Two independent studies published in September 2026 give that question some evidence. A randomised experiment embedded in the ICML 2026 peer-review process found that allowing reviewers limited AI assistance made reviews 5.5 to 7% longer, with no evidence of an effect on paper scores or final decisions. A separate study of AI coding agent pull requests found that 61.4% of agent pull requests received no recorded human review, and that when a merged agent pull request needed a follow-up fix, the agent itself usually opened it. Neither study is about marketing. Both are about what happens when output gets cheap and checking does not.

Apply that warning to marketing; a CMO who can generate an animated infographic from a dull PDF in minutes still has to check every number, every label and every claim inside it before it reaches a stakeholder, a client or a prospect. That verification step does not get faster because the file got easier to build. The bottleneck has not disappeared at all. In fact, it has grown to be increasingly complex due to the nature of the work packaged for a human review. Also, it has shifted from the concept and build steps to a human reviewer, and the last person who signs off the work or changes on it.

The Build Step Just Got Shorter

Claude Motion turns reports, charts, and product walkthroughs into short animated explainers. Anthropic says Claude writes each animation as code rather than using a video generation model, so there is no generated footage and no AI-generated people, and every word, number, and timing stays editable before you export into an MP4 file. For a marketing team used to briefing a designer, waiting on concept and mood boards, then to production days for a first draft, and iterating twice more before an animated asset is applicable and usable. A quarterly results walkthrough or a pricing comparison can now take an afternoon rather than a week. One practical limit matters before you plan around it. Motion is currently in beta on Claude Team and Enterprise plans only, so individual Pro and Max subscribers do not have it yet.

Claude Motion works directly from a source document. Feed it a quarterly report and it produces a moving chart walkthrough rather than a static slide. Feed it a product spec and it produces a stepped tour rather than a bullet-point PDF. For a marketing team used to briefing a designer, waiting days for a first draft, and iterating twice more before an animated asset is usable, this collapses a week of production into an afternoon.

This is a genuine capability. I have used enough of these build-speed tools, from drafting to layout to now animation, to know the pattern by now: whatever used to take days inside the production pipeline eventually takes minutes, and the team that adopts early gets a real head start on output volume. A stack that cannot talk to these tools yet is the first problem that’s time consuming and worth fixing.

From a Singapore practitioner’s view, and this is my own observation rather than measured data, solo and fractional marketing roles in the region were already stretching thin headcount across a lot of output. Agentic AI marketing content verification was already a stretched discipline before Claude Motion arrived. Adding a faster production layer on top of a verification process that was already thin is exactly the moment the bottleneck becomes visible.

But a faster build step changes what gets produced, not what gets checked frequently. An animated infographic carries the same numbers, the same comparisons and the same claims that a static one did. Nothing about the animation verifies that a chart’s axis is scaled correctly, that a comparison table is current, or that a claim about a competitor’s pricing is still true. The file moved, and the facts inside it still need a person to confirm them. Because Motion writes each animation as code, a wrong number can be corrected in seconds once someone spots it, which makes correction cheap, and spotting it is easy, however, anyone can miss it too as we are bombarded by too much information all the time.

Two Studies, One Warning: Output Outpaces Review

The ICML study is instructive precisely because it was not about marketing at all, and because its result is easy to read with various interpretations. The experiment ran inside the real ICML 2026 peer-review process, covering 24,661 papers and 17,886 reviewers. Reviewers were assigned to either a conservative policy that prohibited LLM use or a permissive policy that allowed limited assistance, and the researchers compared outcomes. They found no evidence that the policy changed paper scores, reviewer confidence, or final paper decisions, although reviews written under the permissive policy ran 5.5 to 7% longer. The study did not measure how well anyone verified anything, so it would be wrong to claim it proves a verification gap. AI assistance changed how much text was produced more visibly than it changed what was decided, and a longer review is not evidence of a more thoroughly checked one.

The pull request study is the sharper warning for anyone shipping AI-built assets. Across 891 repositories and five coding agents, 4.5% of merged agent pull requests drew a verified follow-up fix within 30 days, and agent merges carried 1.62 times the odds of a follow-up fix compared with human merges in the same repositories. Of 263 verified fixes on agent merges, 69.6% were opened by the same agent, 27.4% by a human, and 3.0% by a different agent. The authors also cite a figure that should make any marketing leader pause, which is that 61.4% of agent pull requests received no recorded human review. The honest picture is therefore not a human reviewer buried under an agent’s output. It is a review step that often does not happen at all, and a correction loop that the same tool is trusted to close.

The same relocation is already visible in B2B vendor evaluation, where AI agents now do comparison work marketing teams used to own, and the human sign-off step simply moves later in the process rather than disappearing. Marketing content built with tools like Claude Motion sits on the same curve. The build step compresses. The review step, done properly, does not.

What makes this an AI generated marketing content review bottleneck rather than a passing inconvenience is the compounding effect. One animated infographic a week was manageable under a loose review habit. Five a week, each produced in minutes rather than days, is a different workload entirely, and most marketing teams have not rebuilt their review process to match the new volume. The gap between what a team can now build and what it can now verify widens every time the build step gets another speed upgrade, and the pull request data suggests that, left unmanaged, the gap is filled by skipped review rather than faster review.

My Personal Anecdote

For me, this is not a loss as I have gotten around to building my own engines and adding layers of security from the ground up; from AI SEO Agent, an AI Market Desk Intelligence platform, and to video generation engines (internal use) for my platforms. And like what I meant from the above, the bottleneck is super real. I used to have very long review loop queue of work, and it kinda burn me out. That is no longer the case now, since I have adapted and consistently building, while updating my own rendering engines. All these started from a super boring workflow that I did not even want to start with, while pushing back on adopting AI work because of its incapable and unreliable outputs, as well as deliveries. It was through automation that I got the hang of managing agentic AI, its foundations, production layers and engines, and the tenacity to keep pushing the frontiers and capabilities of what my very own rendering engines could do. Well, of course it took me some weeks to experiment, review, change, and to spot the inconsistencies of its outputs across my social accounts.

Building from scratch is not an enjoyable job for some, but for me, it has been fruitful in shaping the outputs I want from my AI models, but not in a great imaginable way. It still requires a longer runway, but I am quite satisfied with what it has accomplished thus far. Knowing that this is the quality we may have to succumb to in the future, do not expect authentic originality, creativity and superb quality output from any AI model. I have given that up and because of that, I am learning to embrace its “unexpected quality outputs” like any humans with deep expertise.

What This Costs a Marketing Leader Who Skips the Check

Here is the practical risk. A team adopts Claude Motion, starts producing animated infographics at a volume that would have been impossible a month earlier, and the review discipline that worked for one static asset a week does not scale to five animated ones a week. Something gets missed, a stale statistic, a mislabelled axis, a comparison that quietly favours the wrong product, and it goes out under the brand’s name.

Readers are increasingly meeting marketing content through AI summaries before they ever reach the original asset, which means an error in an animated infographic can be restated, flattened and redistributed by a chatbot before anyone on the marketing team notices the original mistake. The review step was never just about catching an error before a human reader sees it. It is about catching it before an AI system amplifies it somewhere the brand cannot identify or correct.

For a fractional CMO or a solo marketer, this is not an argument against the tool. Claude Motion is worth adopting where your plan includes it, and resisting a genuine production speed-up on principle is its own kind of mistake. If your plan does not include it yet, the same verification habit applies to any tool that builds assets faster than you can read them. The argument is for building a verification step that scales with output volume rather than assuming the old and slower review habit will somehow catch up because the content now looks slightly polished. Three things make this workable in practice.

1. A named reviewer per asset, not a shared assumption that someone will catch it.
2. A short written checklist covering every number, label and comparison the animation contains, rather than a general read-through.
3. A rule that no animated asset leaves the building without a dated sign-off against that checklist, so the verification step leaves the same kind of evidence trail the build step does.

None of this is expensive. It costs less than the tool did, and it is the only part of the pipeline that genuinely protects the brand once the build step is no longer the constraint.

Final Thoughts: The Bottom Line

Claude Motion and tools like this will keep compressing the concept and build steps. That is good news for marketing and content production teams that are under pressure to produce more with less. But the queue I have watched move through almost every efficiency tool, got deployed inside organisations does not really disappear when the concept and build steps shrink. It changes shape easily due to how its being produced, and it checks in with whoever signs off before publishing. Build the review discipline before the volume arrives, not after the first mistake goes out under the brand’s name.

Part of my experimental short animated videos were posted, where AI build speed meets marketing judgement, on TikTok at @marcomfintech, and on Instagram at @fsmarcomtech with upcoming short videos releasing through this month, November and till almost end of December, hopefully.

Ready to put a verification layer around your own AI-generated content? Start with LITV AI SEO Agent 2.0 and build the audit habit, whereby context tree evidence reveals your brand’s secret weapon.

Frequently Asked Questions (FAQs)

Does Claude Motion make marketing review faster?

Not on its own. Claude Motion speeds up building an animated explainer from a report or chart, but the numbers, labels, and claims inside it still need a person to check them. The build step shrinks, the verification step does not, so the queue moves to whoever signs off before publish. Because Motion writes each animation as code, corrections are quick once an error is found, but finding the error is requires manual work.

What is Claude Motion?

Claude Motion is an Anthropic feature released in beta on 8 October 2026 for Claude Team and Enterprise plans. It turns reports, charts, and product walkthroughs into short animated explainers that Claude writes as editable code, with no video generation model involved, and you can export the result as an MP4.

Which Claude plans include Claude Motion?

At launch, Claude Motion is in beta on Claude Team and Enterprise plans. Individual Pro and Max subscribers do not have it yet. Plan availability may change while the feature is in beta, so check Anthropic’s current announcement before you build a workflow around it.

Why does AI content review not speed up when generation does?

Generating content and verifying it are different kinds of work. In the ICML 2026 peer-review experiment, allowing limited AI assistance made reviews 5.5 to 7% longer with no evidence of a change in paper scores or final decisions. In a study of AI coding agent pull requests, 61.4% of agent pull requests received no recorded human review. Both point to the same risk. Output gets cheaper, and the checking either stays manual or quietly gets skipped.

Who should check an AI generated animated infographic before it publishes?

Whoever owns the final claim, usually the marketing leader or the fractional CMO on the account, not the tool. Every number, label and comparison in the animation needs a named human sign-off before the asset leaves the building.

Does faster content production reduce marketing headcount need?

Not for verification work. Faster production tools compress the build step, which can reduce drafting hours, but the judgement step, checking accuracy and framing, still needs a person with context on the account. That role does not shrink with the tool.

How should a marketing team adopt Claude Motion safely?

Treat the animated file as a first draft, not a final asset. Build a short verification checklist, covering each number, label and claim, and assign one named reviewer per asset before it goes to a client or publishes externally.

Here is the number that prompted this piece. Alli AI analysed 24,411,048 proxy requests across 69 customer websites between 14 January and 9 March 2026, and reported that OpenAI’s ChatGPT-User made 133,361 requests against Googlebot’s 37,426. That is 3.6 times Googlebot’s volume, and more than Googlebot, Amazonbot and Bingbot combined. The analysis ran on Search Engine Journal as sponsored content written by Alli AI’s own chief executive, so read it as a vendor dataset rather than independent research. It is still a useful prompt, because many technical SEO audits size infrastructure thresholds around Googlebot’s volume, and a fetcher running at a multiple of that deserves a hard look at the rules sitting in front of your origin server.

Where the Policy File Stops and the Server Starts

Robots.txt is a request, not an enforcement mechanism. It tells a well-behaved crawler what it may fetch, and it says nothing to the web application firewall, the rate limiter in front of PHP-FPM, or the CDN’s bot-management ruleset, none of which read it. In many stacks those three layers were configured with Googlebot’s behaviour in mind. Googlebot adjusts its own pace in response to server health, which is part of why thresholds tuned around it have felt safe for so long. Google retired the crawl rate limiter tool in Search Console on 8 January 2024, citing improvements to its crawling logic and noting that Googlebot already slows itself down when a server returns errors or responds slowly. A newer class of AI fetchers, with its own user agents and request patterns, was not part of that tuning in many configurations.

OpenAI’s own crawler documentation splits its traffic across several user agents. GPTBot crawls content that may be used to train its foundation models. OAI-SearchBot surfaces websites in ChatGPT’s search features, and OpenAI recommends allowing it in robots.txt and also allowing requests from its published IP ranges. ChatGPT-User visits a page when a user’s question or a Custom GPT action calls for it, and OpenAI states that robots.txt rules may not apply to it because a person initiated the action. The same page says ChatGPT-User is not used to decide whether content may appear in Search, and that OAI-SearchBot is the one to manage in robots.txt for that purpose.

Read that carefully, because it changes the audit question. GPTBot, the crawler most people check in robots.txt, is the least relevant of the three to live retrieval. OAI-SearchBot, which decides whether you appear in ChatGPT search answers, is easy to overlook. And ChatGPT-User, the fetcher behind the Alli figures, may not consult robots.txt at all, which makes your firewall and CDN rules the main lever you hold over that traffic.

The wider picture is moving quickly and does not point in one direction. Botify and Nectiv analysed roughly seven billion OpenAI bot log events from enterprise sites between November 2024 and March 2026, and found that OpenAI’s automated crawling roughly tripled after GPT-5 launched in August 2025. OAI-SearchBot activity rose about 3.5x and GPTBot about 2.9x. ChatGPT-User moved the other way, falling 28% between December 2025 and March 2026 in that dataset, and even after the tripling, OpenAI’s crawlers came to about 4% of Googlebot’s volume. Two datasets paint two very different pictures, which is exactly the point. Alli’s sample is small, WordPress-heavy and drawn from customers who opted into a crawler enablement platform, while Botify’s skews towards large enterprise sites. Neither is a portrait of your site.

My Personal Anecdote

I run LadyinTechverse on managed WordPress hosting at Verpex, after starting out self-hosting on DigitalOcean, and robots.txt was always the last line of my crawler checks. Nothing in it ever looked wrong, which is exactly the problem. After an audit, I realised a clean file says nothing about what reaches the origin. I honestly did not know whether my own site was turning ChatGPT away at the door. My robots.txt says it is welcome, but that file is only a sign on the gate. Until I checked the records of who actually got through, all I have is a hunch.

Why the Audit Checks One File and Stops

Many technical SEO audits, including ones I have run on WordPress, first self-hosted on DigitalOcean and now on managed hosting at Verpex, treat robots.txt as the finish line for crawler access control. Confirm nothing important is disallowed, tick the box, move to the next line item. That habit comes from a decade in which Googlebot was the only crawler that mattered at meaningful volume, and a clean policy file was a reasonable proxy for access.

AI fetchers break that proxy, because every user agent has its own rules and the infrastructure in front of your origin classifies bots on its own terms. Take a rate limiter set at 60 requests per minute per IP range for bot traffic. That figure is purely illustrative, and whether it would ever catch an OpenAI fetcher depends on how those requests arrive, in what bursts, and from which published ranges. Public data cannot answer that for your site. Averaged across Alli’s sample, ChatGPT-User made roughly 35 requests per site per day, which is my own arithmetic from the published totals and which hides whatever spread exists underneath. In the same dataset ChatGPT-User received a successful response 99.99% of the time, but Alli serves pre-rendered pages from the CDN edge, a very different set-up from a self-hosted WordPress stack. The sample shows what is possible, not what your origin and firewall would do.

When a block does happen, it is usually recorded in the firewall or CDN’s own logs, and it may never appear in the origin server log at all. That is why the log you check matters as much as the rule you read. I covered the allow-or-block decision itself, the GEO policy layer, in an earlier piece on AI crawler access control. This piece is about the layer underneath that decision, the one that could quietly reverse it.

Cloudflare’s Bot Preference Sync, announced on 21 August 2026 and covered by Slobodan Manic in a column for Search Engine Journal on 18 September, generates a robots.txt file from three dashboard settings for Search, Agent and Training bots. Cloudflare’s stated rationale, as Manic reports it, is that when a site’s declared preferences and its enforced rules disagree, some crawlers treat that gap as grounds to disregard the preferences. I read that as a tacit acknowledgement from the infrastructure side that policy and enforcement drift apart. Manic’s own objection is that three categories cannot express a per-crawler policy, and he notes that at the time of writing the feature had not yet appeared in Cloudflare’s bots changelog, so treat it as announced rather than universally live.

The Four-Layer Audit That Confirms Access

A technical SEO audit that wants to know whether OpenAI’s fetchers can reach a site, rather than whether the policy file welcomes them, needs to inspect four layers in order. First, robots.txt itself. Confirm GPTBot and OAI-SearchBot are treated the way you intend, remember that ChatGPT-User may not follow the file at all, and check the directive syntax against OpenAI’s documentation, since a malformed user-agent match is its own silent block. Second, the WAF ruleset. Pull the rate-limit and bot-management rules by name rather than by assumption, and identify which request-per-minute or request-per-hour ceiling applies to the IP ranges and user-agent strings OpenAI publishes for each of its bots. Third, the CDN layer, which on many self-hosted WordPress stacks sits in front of the WAF and makes its own bot-classification decision before traffic ever reaches the origin. Fourth, the logs. Start with the WAF and CDN logs, then check the origin server logs, and look for a run of 403 or 429 responses against OpenAI’s published ranges, because that is the tell that policy said yes while infrastructure said no.

This fourth layer matters because it cannot be faked by good intentions. A marketing leader can confirm robots.txt is correct and still have little ChatGPT traffic reaching the origin, because the WAF may have returned a 403 before the origin log line was ever written. Auditing only the policy file can produce false confidence. The audit passes, the citation eligibility it was meant to protect does not materialise, and nobody connects the two because the failure happened one layer below where anyone looked.

What This Costs a B2B Marketing Leader Who Never Audits It

GEO and AEO visibility depend on AI systems being able to retrieve current content, not on a policy document saying they are welcome to. A B2B site that has invested in structured data, clean chunking and citation-worthy answer formatting, the work I have written about in the context of retrieval failure modes behind wrong AI agent citations, gets non of that benefit if the fetcher carrying the content back to the model is throttled or blocked at the edge. I cannot show from public sources that this is happening on any particular site, and that is exactly why the audit matters. From the infrastructure side, blocking an aggressive, high-volume bot is simply a rate limiter doing its job, so the content team never hears about it and the infrastructure team never thinks to flag it.

This also compounds the measurement blind spot covered in why AI agent traffic hides inside Chrome analytics. A marketing leader who cannot see AI agent traffic in GA4, and whose infrastructure team may be rate-limiting that same traffic is working from two independently incomplete pictures, each one making the other look less urgent than it is.

Final Thoughts: The Bottom Line

Robots.txt is policy. WAF rules, rate limiters and CDN bot management are enforcement. For ChatGPT-User, OpenAI’s own documentation says the policy file may not even apply, which makes enforcement the layer that matters most. The two sets of controls are configured by different teams, reviewed on different schedules and audited by different checklists, which is how a site can say yes at the policy layer and no at the infrastructure layer without anyone noticing the contradiction.

The fix is not a bigger robots.txt file. It is a technical SEO audit that treats crawler access control as a four-layer question, with logs as the layer that tells the truth about what the other three did. The datasets cited here cannot tell you whether your stack is blocking OpenAI’s fetchers. Only your own logs can. If AI citation eligibility matters to your marketing outcomes in 2026, this is an audit line item worth adding. Read how I approach that work.

Frequently Asked Questions (FAQs)

Does allowing GPTBot in robots.txt guarantee ChatGPT can reach my site?

No. GPTBot is OpenAI’s training crawler. Surfacing sites in ChatGPT search is handled by OAI-SearchBot, and user-initiated page visits are handled by ChatGPT-User. Each has its own rules, and a firewall, rate limiter or CDN can block any of them regardless of what robots.txt says.

Does robots.txt control ChatGPT-User?

OpenAI’s documentation says that because ChatGPT-User actions are initiated by a user, robots.txt rules may not apply. It also says ChatGPT-User does not decide whether content appears in Search, and that OAI-SearchBot is the agent to manage in robots.txt for that. For ChatGPT-User, server-level rules are the control you have.

Why might a WAF or rate limiter affect OpenAI crawlers when robots.txt allows them?

WAF, rate-limiter and CDN rules do not read robots.txt. Alli AI’s sponsored analysis of 69 sites found ChatGPT-User making 3.6 times Googlebot’s requests, so thresholds tuned around Googlebot may behave differently with that traffic. No public dataset shows how often blocking actually happens, which is why you should check your own rules and logs.

How do I check whether my site is blocking OpenAI’s crawlers?

Pull your WAF and CDN bot-management rules by name, compare their rate-limit ceilings with the IP ranges OpenAI publishes for each bot, then check WAF, CDN and origin logs for 403 or 429 responses against those ranges. A cluster of them means enforcement is overriding policy.

Is this the same issue as deciding whether to allow or block AI crawlers?

No. The allow-or-block decision is a GEO policy choice. This is the separate infrastructure layer, made up of rate limiters, WAF rules and CDN thresholds, that can quietly reverse an allow decision after it has been made.

Why does this matter for AI citation eligibility?

GEO and AEO visibility depend on AI systems being able to retrieve your content. OAI-SearchBot is the agent OpenAI says surfaces sites in ChatGPT search, so it is the first one to verify. Content investment in structured data and clean formatting delivers no citation benefit if the fetcher is blocked before it reads the page.

At the same time, AI is lowering the barrier to building software itself. The emerging ArtCraft project is a vivid example. Software engineer, Brandon Thomas says he used AI-assisted development with Anthropic’s Claude Opus 5.5 to create a free, open-source suite of applications intended to replicate workflows found in seven Adobe products (Adobe Creative Cloud).

The last time I looked at the ArtCraft page, it read “Seven apps. One craft.”. As of 19 October 2026, it reads “Thirteen apps. One craft.” The pace alone is mind-blowing. The size additions cover interface design, word processing, audio, presentations, CAD, and spreadsheets, including early stage equivalents of Word, Excel, and PowerPoint. This article stays with the seven Adobe equivalents because they are what the first reports covered.

These two stories are uncannily connected. As software becomes easier to reproduce, the tools people use may change faster. But when an AI agent acts inside those tools, responsibility does not become any clearer. The engagement contract, insurance policy and liability clause may still be written around a human doing the work.

It seemed like Brandon has fully open sourced MS Office. The last time I saw his page, it wrote “Seven Apps. One Craft”. Its now showing “Thirteen Apps. One Craft”. mind-blowing epic software reverse engineering

The Application Layer is Starting to Flatten

For a long time, building a serious creative application required substantial engineering time, expertise and resources. That made established software suites difficult to challenge — not only because they were familiar, but because reproducing their capabilities was expensive and slow.

ArtCraft points toward a different possibility. In early October coverage, TechSpot and PetaPixel reported that Thomas used AI tools to help build open-source alternatives to seven Adobe applications, including tools for photo editing, vector graphics, video, page layout, and PDFs. The project includes PhotoCraft, VectorCraft, FilmCraft, and LightCraft. It is built in Rust, and is intended to run on macOS, Windows and Linux.

That is not the same as proving Adobe has been replaced. ArtCraft is an early stage application. PetaPixel reports that Thomas describes the suite as alpha, and Techspot reports that commenters have encountered problems with basic functionality. Thomas’ stated ambition to reach feature parity is a target, not evidence that the applications already match Adobe for reliability, performance, or professional use. TechSpot also reports that Thomas first promised 100 percent parity within a month, then said on Hacker News that reaching 99 percent would take months rather than years. But the signal is significant: the work of creating an application may be getting less exclusive.

If AI-assisted development makes it easier to reproduce familiar workflows, more developers may be able to build alternatives to established products. The near-future software market could become flatter: more tools, more competition, and fewer applications that are difficult to challenge simply because building a competing product once required a large team, and a long timeline.

When the Tools Multiply, So Do the Access Questions

That shift matters to any professionals using AI tools. A CMO may work across client-owned CRMs, ad platforms, analytics tools and automation systems. Those applications may increasingly include AI features, third-party integrations, and purpose-built alternatives. The stack can change quickly, while the permissions and accountability arrangements around it remain harder to update.

A client’s security team may not view an agent as simply “the CMO working faster”. If the agent has a scoped API key or service account, it may be managed as a separate non-human identity, with its own permissions and activity logs.

The practitioner relationship remains human, but the access pattern does not.

That distinction becomes more important as AI agents take on work across a growing number of applications. A client may approve the human engagement, grant the agent access and permit specific actions — yet the contract may never name the agent, or explain how responsibility applies when work is performed through it.

Why The Paperwork Can Fall Behind

Software products can evolve quickly. An agent can be connected to a CRM, an ad platform or a newly adopted application without the engagement agreement being revisited. The technology changes; the written assumptions may not.

Many service agreements are framed around a named professional delivering work personally, or through staff under their supervision. Their confidentiality, liability and performance terms may not clearly address an AI agent holding its own credential, and taking actions inside client infrastructure.

This is not only a question of which application the agent uses. It is also a question of what the agent can do, which identity it uses, who approved that access, and how its actions are attributed.

A governance framework may explain who approves an agent’s actions. The contract still needs to make clear how agent-mediated work fits within the engagement, and what happens when an agent acts within the permissions it was given but produces an unexpected result.

The Insurance Question is Still Open

Professional indemnity policies, and client contracts may not expressly address every AI agent scenario. Coverage depends on the wording of the specific policy, the work performed, and the circumstances of a claim. It should not be assumed that a policy automatically covers, or automatically excludes an agent’s actions simply because the professional approved its use.

The practical issue is disclosure and clarity. If an agent will access client systems, the professional should understand whether the client and insurer know about that access, how it is controlled and how responsibility is treated in the relevant documents.

That question becomes more pressing as software alternatives multiply. If teams can adopt new applications, or connect agents to new tools more easily, each change may introduce a new access path. A tool being cheap, or free doesn’t decide who’s responsible for what its AI agent does.

What Fractional Professionals Should Do Before The Next Engagement

Three steps can help make the arrangement clearer:

1. Ask how the client classifies, and manages the agent’s access. Find out whether it will use a dedicated service account, API key, or other credential, what permissions it will receive, and whether its activity will be logged separately from your own.
2. Check the agreement against the actual workflow. If an agent will perform work in your own or a client’s systems, make sure the contract addresses agent-mediated work and clarifies how that work is attributed to the signing professional. Do not assume a clause written for human performance automatically answers every question.
3. Ask the insurer directly in writing. Describe the agent, the systems it will access, and the tasks it will perform. Ask how the policy applies to that arrangement before renewal, or before the work begins. As applications become easier to build and replace, the number of tools in a professional’s stack may grow. The accountability arrangements should not be left behind.

My Personal Anecdote: When the System is Your Own, the Questions Do Not Go Away

Few CMOs give an agent a scoped API key for their own company’s CRM, ad accounts, or automation platform, where no client or engagement contract is involved. That can feel like internal housekeeping, yet nothing about the risk changes. A credential exists, it carries permissions, and a named person is accountable for what the agent does with it. It is also where sandboxes live, thrive and fail at the same time.

With no client to ask the awkward questions, the company has to ask them of itself. Which identity does the agent use, who approved its access, and who is the named human accountable for it? What does it log, and when did anyone last review it? If software keeps getting cheaper to build, these questions become more urgent, not less. A team that can build or adopt an internal tool quickly may connect an agent to it just as quickly, and every connection is a new access path that nobody scheduled time to review.

I put the same questions to myself. For my own platform, there is no client and no procurement team to set the rules, so I set them. When I work with Claude and ChatGPT, I ask for judgement calls to be put in front of me as options before anything is executed, and I ask for proposed changes as a list I can review, approve, or reject. I also have an AI agent acting as my QA and QC gatekeeper. It checks the quality before anything reaches my review, and if I do not have the time to review the document, it sends me a reminder before the deadline. Those habits are small, but a security team would recognise them: a defined role, a named human who is accountable, and a record of what was proposed. IYKYK, right — the agent acting as my QA and QC gatekeeper can once-in-a-blue-moon sleep on its job, breaks off from the workflow due to insufficient tokens, missed a few audited lines, and did not check on the latest scopes. That’s where I have to come back to my computer to fine tune the entire process, after each model or patch updates. I have to be very restrictive for certain type of workflows. Therefore, the agent may have decided to bail on me without any conscience, when I told the agent that it is having a ‘dementia’ moment.

Final Thoughts: The Bottom Line

The application may be easier to replace, however, accountability isn’t. ArtCraft is too early to prove that established creative software is about to disappear. It does offer a timely illustration of how AI-assisted development may lower the barrier to creating competing applications.

That could make software more abundant, and the application layer flatter, but it does not flatten responsibility. When an AI agent uses a scoped identity inside a client’s systems, the human engagement, access permissions, and insurance terms still need to line up.

The risk is not simply that an agent makes a mistake. It is that the systems allowed the agent to act, the contract never clearly described that arrangement, and everyone assumed someone else had resolved the gap.

Frequently Asked Questions (FAQs)

What does it mean for the software application market to become “flatter”?

It means the barriers to building competing applications may fall, making it possible for more developers to attempt products that reproduce familiar workflows. It does not mean every new application will match established products in quality, reliability, or support.

Does ArtCraft already replace Adobe Creative Cloud?

No. PetaPixel describes the suite as alpha, and TechSpot reports commenters encountering problems with basic functionality. Its stated feature parity goals should be treated as ambitions, not as proof that it is already a professional substitute.

What is an AI Agent’s machine identity?

It is a non-human credential, such as a service account, or an API key, that lets an agent access systems without using a person’s login. The credential can have its own permissions, and activity records.

Who is liable if a CMO’s AI agent makes a mistake?

That depends on the contract, the facts, the permissions granted, and the applicable insurance policy. The important practical step is to make agent-mediated work explicit in the engagement, and confirm the insurance position directly with the insurer.

What should I clarify before letting an agent use a client’s systems?

Clarify the identity, and permissions it will use, how its actions will be logged, what is is authorised to do, how agent-mediated work is treated in the contract, and whether the insurer has been informed.

Do the same rules apply when an agent works inside my own company’s systems?

Yes. With no client, nobody outside the company asks the questions, so the company has to ask them itself: which identity the agent uses, who approved its access, what is logged, and who reviews it.

Is agentic AI governance the same as contractual accountability?

No. Governance can define who approves, and monitors an agent’s actions. Contractual and insurance terms address how responsibility is allocated if those actions cause harm, or produce an unexpected outcome.

What is PHP2Shell?

On 21 October 2026, the PHP project released fixes for two critical flaws, CVE-2026-109618 and CVE-2026-109619. Both score 9.2 on CVSS v4.0, and allow an attacker to run code on a web server without logging in. The fixed versions are PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35. The problem lives in PHP itself rather than in WordPress, which is why WordPress.org confirmed that no WordPress update is required or planned. Wiz researchers showed unauthenticated code execution working against the latest official WordPress, Drupal, and Joomla container images at the time of testing.

The disclosure itself went sideways. Details leaked on 20 October right before the embargo. PHP brought its release forward by a day, and the UK National Cyber Security Centre advised organisations that cannot patch immediately to consider taking systems offline. Amazon’s honeypot network recorded an exploitation attempt on 17 October, four days before public disclosure, and watchTowr reports observing exploitation attempts. Huntress had not see in-the-wild exploitation across its partner base as of 21 October.

Where AI is the actor in this story

This is where the actual incident emphasised the headline version of this story. The researchers at Wiz, working with the PHP team, Calif, and OpenAI, say they used Wiz Atlas, Gemini 4 Argon, and GPT-5.6 Sol to find the flaws and build working exploits. Their claim is that one researcher completed in about a week with what would once have taken a world-class team months, and that the models let them develop, adapt exploits across several applications within hours. That is somewhat defenders using AI. On the attacker side, John Hammond of Huntress said the leak gave criminals an early warning window, and that they were undoubtedly using AI to find targets, and adapt exploits. That is an informed expectation, which no one wants to end up with finding out after its done.

So no, I cannot tell you that AI has been hacking the Internet since yesterday. What I can tell you is that AI has collapsed the effort between a hard bug, and a working exploit, and that is the shift every founder, business owner, and site owner needs to think through, upgrade reliable cybersecurity premiums, and much more — if it’s a very mature business that has been lagging behind digital, security and marketing transformations.

My Personal Anecdote

I have supported WordPress sites for years, and I have watched plenty of patch cycles come and go. Most felt like routine housekeeping. This is not the first big patch I have seen land, but it is the first where frontier AI sits at the centre of the story, at least from where I stand. That is why I think enterprises should treat it as a rehearsal to practice anticipations, and reactions when the time comes. Extra vigilance is now an operating cadence, and not to settle with some monthly reviews.

Here is where I am not pretending as a theorist. Between 22 September and 11 October 2026, my firewall logged automated probing from eight locations: Sydney, Seoul, Dhaka, Tokyo, Zurich, Groningen, The Dalles in Oregon, and Pakistan. Whoever these idiots are, they clearly have nothing better to do. One Sydney address tried /wp-admin/install.php under at least 12 different paths in about six seconds, hunting for a forgotten WordPress setup screen it could complete on its own terms. Others asked for files name /db.php and cache.php, the kind of names backdoors like to hide behind, or tried to climb the directory tree to read a configuration file. One went looking for a known flaw in an SMTP plugin. Every request in these screenshots are blocked by my firewall. None of it is PHP2Shell because it all predates the disclosure. It is the Internet’s current affairs’ noise, and that noise is why I do not leave PHP patching to chance. A fair caveat: an IP location shows where a connection from, not where a person sits, and three of these resolve to Google Cloud servers. If its a bot, it is pretty obvious, but if its a human, it could be an AI agent at the behest of its owner.

What Claude Fable 5.1 and Opus 5.5 Can and Cannot Do

You may be wondering what the latest Claude models can hack, so here’s the short report. Claude was not among the models Wiz credited in the PHP2Shell research, so I will not attach this story to Anthropic. What Anthropic has published is that Claude Fable 5.1 and Mythos 5.1 are the same model with different safeguards, and its system card describes the strongest cyber capabilities of any model it has released. On general access, safeguards block or reroute offensive work such as reviewing source code for bugs remains permitted. Claude Opus 5.5 was released on 22 September 2026. It was described as comparable to Mythos 5.1 in cybersecurity and biology, and delivers with similar safeguards. Vetted defenders can get fewer restrictions through Anthropic’s Cyber Verification Program, which expanded into three tiers in October.

In layman’s terms, the capability exists and allows better access control rather than inability. Anthropic itself notes its safeguards have limits, and Wiz used other frontier models. The plan for a capable adversary, whichever the logo is found on the AI model.

Where AI Finds The Easy Wins

The PHP flaw is a deep engineering bug, but most real-world compromises still start with something boring. Wiz also showed that a related memory disclosure path could expose passwords, token, API keys, and certificates from a PHP worker’s memory, which is a reminder that your secrets are only as safe as the server they pass through. In my view, the weaknesses AI will exploit fastest are the ones we keep tolerating: unpatched software, passwords that are easy to guess or reused, encryption keys shorter than current guidance, and passkeys, or recovery methods synced across more devices, and accounts than anyone has audited. Passkeys are a genuine upgrade, but a passkey is only as safe as the weakest place it is stored. I covered how convincing AI-enabled fraud has become in my earlier piece on AI scams, and the lesson shows here. AI lowers the cost of attacking you, so the margin for carelessness and sloppy basic security practices are key fundamentals that facilitate these attacks.

What To Do Today

Start by finding your PHP version. In WordPress, go to Tools, then Site Health, then Info, then Server. Your host may already have patched, and some hosts backport the fix so the version number does not change. Verpex, for example, tells customers to check that the phpinfo() build date shows 21 October. If you manage your own server, update to a fixed version, and restart every PHP worker, as well as treat anything on PHP 8.1 or older as unsupported by the PHP project.

After patching, look for new web shells, unexpected scheduled tasks, and unfamiliar admin accounts, and consider rotating passwords and API keys your site has handled. Enterprises should add one more step, which is an inventory of every PHP install, including the ones hiding inside vendor software and appliances. Palo Alto Networks, for example, confirmed that all PAN-OS and Panorama versions are affected, and had not delivered fixed releases at the time of reporting. If your leadership team has not rehearsed this kind of crisis event, start with my guide to AI-assisted crisis communications for CEOs.

Check your host before attackers do

I host LadyinTechverse on Verpex, and during this incident the thing I valued most was transparency. Its status page logged the PHP update from 21 October. As of 22 October, it said patches were being rolled out for cPanel and Plesk shared hosting across PHP versions from 5.2 and above with no customer action required, Enhance panel sites were patched from PHP 7.4 and newer, and self-managed VPS owners were expected to apply vendor updates themselves. It also recommends moving to PHP 8.4 or newer if you are on 8.3 or older. The incident was still marked as recovering at the time of writing, so check the live page at status.verpex.com rather than trusting a blog post, including this one.

Disclosure: the link below is an affiliate link, and I may earn a commission at no extra cost to you.

Frequently Asked Questions (FAQs)

Is WordPress itself vulnerable to PHP2Shell?

The flaws are in PHP, not in WordPress, so WordPress.org says no WordPress update is required or planned. However, a WordPress site running on an unpatched PHP version is exposed, and Wiz demonstrated unauthenticated code execution against the latest official WordPress container image at the time of testing. The fix is a patched PHP build from your host, or server vendor.

Which PHP versions contain the fix?

PHP 8.5.12, 8.4.27, 8.3.36, and 8.2.35. The PHP project no longer issues security updates for 8.1 and earlier, although it has published minimal source backports for 7.4, 8.0, and 8.1 for teams that build from source. Some distributions, and long term support vendors also backport fixes, so ask your provider rather that assuming.

Is AI actively hacking websites through this flaw right now?

Exploitation attempts have been reported. Amazon’s honeypot network logged one on 17 October, before public disclosure, and watchTowr says it is observing attempts, while Huntress had not seen exploitation across its partner base as of 21 October. Researchers used AI to find and weaponise the flaws, and Huntress expects attackers to do the same, but attacker use of AI in this campaign has not been confirmed.

Can Claude Fable 5.1 or Opus 5.5 be used to hack a website?

Not on a general request. Anthropic says both deliver with safeguards that block or reroute most offensive cyber work, such as penetration testing and exploit generation, while ordinary secure-development work such as reviewing source code for bugs remains permitted. Vetted security teams can apply for reduced restrictions through the Cyber Verification Program. Safeguards have limits, so plan for capable adversaries using any model.

What should a small business do today?

Confirm your PHP version and patch status with your host, move to a supported PHP branch, and verify the patch rather than assuming it. Then check for unfamiliar files, scheduled tasks, and admin accounts, rotate credentials your site has handled, and use long unique passwords with two-factor authentication or passkeys you have audited.

New AI models keep arriving, and the best one this month may not be the best one next month, so building your business around any single model is a risky bet. A safer plan is to own the computer, and the rules that decide which AI does which job. That way you can swap in a better or cheaper AI when one appears, and keep your private work away from other companies’ systems.

The Knife Problem: Blades Wear Out Faster than Handles

Generated by me on Flow.

Picture a Swiss Army knife like above AI-generated video. One handle holds many tools, and you can swap the blades. Now imagine the blades changing very fast, and you have today’s AI market. Think of the handle as your computer, your files and your rules. The blades are the AI models, and they come and go.

Mozilla’s State of Open Source AI report (September 2026), looked at free-to-download models and found the best one is only about three points behind the best paid model on a combined test score. My reading is simple. When the gap is that small, the model you pick matters less than the setup around it.

Models can also disappear without warning. On 12 June 2026, Anthropic switched off its newest model, Claude Fable 5, for every customer, to follow a US government order that limited who could use it. Fable 5 came back worldwide on 1 July. A free model already saved on your own machine would have kept working through those weeks because nobody can switch off a copy you hold from the outside.

What You Should Own

If the models are the blades, here is what the handle is made of.

• First, your files. That means your documents, notes, client folders and the instructions you give to AI, saved in formats you can open anywhere.
• Second, your rules. This is a short written list that says which AI does which job, and which jobs must never leave your own machine.
• Third, your connections or connectors, like self-hosting MCP server, a 24/7 device and so forth. These are the small links between your tools, kept simple enough that you can change the AI at one point without rebuilding everything.
• Fourth, your record. This is a log of what the AI did, so when something goes wrong you can find out why.

None of this is exciting for non-technical and not-IT individuals, but all of it lasts longer than any one model. A startup team that writes and commits to these four things down can change AI providers without starting from scratch.

What Running AI at Home Can Do Today

On 29 September 2026, Artificial Analysis, a company that tests AI, published a speed test for AI helpers that run on home-sized computers. It tried four machines and four free-to-download models. The fastest machine had a high-end graphics card (also called a GPU), the RTX 5090. For every model that fit in its memory, it finished the work more than 3.5 times faster than the other machines.

The cheapest machine tested was a MacBook Pro laptop with 64 GB of memory, at about USD 3,700. The biggest gap between these machines was waiting time. The laptop could go quiet for up to 39 seconds at a stretch, while the GPU machine paused for about two seconds. The test also says most of these machines now sell well above their launch prices.

So running AI at home is real, and highly depends on your objective and weekly / monthly usage. The catch is speed, and how willing you will invest in it than going for a mainstream service provider.

Compute is also moving towards homes from the supply side, and the detail matters. On 13 April 2026, Span (a smart electrical panel maker), announced XFRA, a network of small data centre nodes installed at homes and small commercial sites, with NVIDIA as a launch partner supplying liquid-cooled RTX PRO 6000 Blackwell Server Edition GPUs. The nodes are built to serve hyperscalers and AI cloud providers, while the homeowner’s side of the deal is a smart panel, battery backup and fixed, discounted electricity and Internet rates. Span said initial deployments would begin later in 2026, and with gigawatt scale targeted for 2027. My view is that hosting an outdoor unit that runs other companies’ AI workloads is not the same as owning the machine.

What It Cannot Do Yet

Home AI is not yet as good as the best paid AI for every job. Mozilla’s report says one of the top free models, Kimi K3, needs about 1.56 terabytes of memory spread across 64 or more specialised AI chips. That is a small data centre, far beyond any home computer. The best free model that fits on one server scores about 10 points below the top and best paid AI. The best one that fits on a single graphics card scores about 23 points below it.

Running free models is also harder than it looks. A SlashData developer survey, run in May 2026 and reported in Mozilla’s study, found that 51% of people using free models had them in realtime, everyday use, compared with 63% for paid models. Mozilla says the gap comes from tooling, meaning the effort of setting things up and trusting the result, not from how clever the AI is.

The Simple Rule: Private Work Stays Home, Public Work Goes to the Cloud

Here is a simple rule. Ask three questions about any task. Would a leak hurt me or my clients? Do customers touch it? Does it need to be online all day, every day?

If a leak would hurt and customers never touch it, run it at home. Contracts, client files and pricing are good examples. If customers touch it, or it must never go down, use the cloud, which means AI you rent over the internet. A home computer cannot promise it will always be on. Put your answers in a simple table and share it with your team, or if you’re operating solo – with your AI agent.

Many AI tools talk to home models and cloud models in the same way, so you can switch between them without rebuilding your work.

My view is that graphics cards will become the standard home machine for private AI. In the speed test, the GPU machine was fastest by a wide margin. But a card only holds models that fit in its memory, which was 32 GB in the test, so bigger models need a different machine. Memory is the limit. The four machines in the speed test held between 32 GB, and 128 GB, so bigger models obviously need a higher-spec machine.

A real case shows why holding a model yourself matters. In July 2026, Hugging Face, a large AI website, found that someone had broken into its systems. OpenAI later said during a test, its own AI models were behind it. To read the attack records, Hugging Face first tried paid AI services. It says their safety filters blocked the work because the filters cannot tell a defender from an attacker. So it ran a free model on its own computers, and says no attacker data left its systems. It also says this is no argument against safety filters on paid AI. Hugging Face has big computers, so a home laptop cannot copy this. The lesson is the plan, not the size.

One more point: a single machine can fail, so back it up and control who can use it. While you can have a 24/7 single machine running on the clock, you would definitely need to invest in a solid backup drive. For Mac users, it would be Time Machine Backups with Western Digital for Mac (highly recommended by me for anything above 1 TB). For Windows users, its either Western Digital, Seagate or Kingston SSD.

My Personal Anecdote: My Few Days of Running AI on My Own Laptop

It started with the craze on using Ollama when ChatGPT, Claude, Gemini, Perplexity were pushing frontiers for their models. Ollama wasn’t just friendly to home-based users, it was kinda refreshing knowing that a person can own his/her stack and train the AI models in the comfort of their own home. However, this comes with a cost too deep for heavy usage users – the enormous effort drawing the compute power, and if you’re using an entry to mid level Mac machines or PCs, good luck to your chips and electric bills. It was because I wasn’t using a super powerful Mac machine, it started blowing fans louder than any other days, so I had to stop it. It was fun while it lasted but heck, available time is not exactly on everyone’s schedule. If you’re a student or unemployed, this is something you could start dabbling with and learn it while you’re hands-on, rather than relying on free basic virtual GenAI courses out there that don’t really benefit as much as when you learn it yourself, practice it yourself, while being practical on the AI model usage tokens.

A Few Tests Before You Buy the Real Hardware

Before you spend thousands, try these five tests on the computer you already own.

• Run a free model for three to seven days on your real work (start with a personal project first), using a free app such as Ollama or LM Studio.
• Measure your waiting time, using your biggest real files and not using short test questions.
• Check the memory and monitor spikes in your computer activity. A model that does not fit in your computer’s memory will not run well.
• Check the licence. Free to download does not always mean free to use in a business, and the terms differ from model to model.
• List which tasks truly must stay private, then check today’s price for the machine.

Final Thoughts: The Bottom Line

AI models will keep changing, so do not tie your personal projects or business projects to one of them. Keep your computer, your files, and your rules because those are the parts that last. Rent the best AI for public and hard jobs, and keep private work on a machine you own. Start small, test for a week, and only then spend.

Access full suite of channels and platforms by LadyinTechverse: bio.ladyintechverse.com, and my B2B personal brand: marcomfin.tech for more AI insights.

Frequently Asked Questions (FAQs)

Can I run AI on my own computer at home?

Yes. Free apps such as Ollama and LM Studio let you run free-to-download AI models on a Mac or a PC. A September 2026 speed test found that home-sized machines can do the work, but speed and price vary a lot.

Is home AI as good as paid AI?

Not yet for every job. Mozilla’s September 2026 report says the best free model is about three points behind the best paid one, but that model needs a small data centre. The best free model that fits on a single graphics card is about 23 points behind.

What computer do I need?

It depends on the model because the model must fit in the computer’s memory. In the September 2026 speed test, a computer with a high-end graphics card was fastest, and a 64 GB MacBook Pro was the cheapest of the four machines tested at about USD 3,700.

Should private work stay at home and public work go to the cloud?

That is a sensible starting rule. Keep private work on a machine you own, and use the cloud for anything customers touch or anything that must always be online. Back up your home machine because one machine can fail on anyday.

What should I own first?

Start with your files and your rules. Save your documents in formats you can open anywhere, and write down which AI does which job. These last longer than any single model.

The Protocol Anthropic Built, and Then Let Go

MCP is Anthropic’s, in the way that matters for a founding story. The company released it in November 2024 as an open specification for connecting AI models to external tools, data sources and systems without a bespoke integration for every pairing. Agent Skills followed, a related format for packaging reusable agent capabilities so they can be invoked consistently across different runtimes. Both solved a real problem: before MCP, every model-to-tool connection was its own one-off wiring job, and the industry was heading towards an integration mesh nobody could maintain.

Here is the part most people miss — Anthropic stopped being the only caretaker of MCP well before Agent Plugins 1.0 appeared. MCP, short for Model Context Protocol is the standard that lets AI tools connect to other software. In December 2025, Anthropic co-founded that foundation with Block and OpenAI, so MCP now has a neutral home instead of a single owner.

That is a generous move by open-source standards. It also means that calling MCP “Anthropic’s protocol” is only about half right. Although Anthropic invented it, they no longer own it. Inventing a standard and deciding how it gets packaged, and shared are two different jobs. This article is about who ended up with the second job.

What Agent Plugins 1.0 Does

Agent Plugins 1.0 is not a rival to MCP. Think of it as the box everything packaged in left to deliver. MCP lets an AI tool talk to other software. Agent Skills are reusable instructions that teach an AI how to do a task. Agent Plugins 1.0 is a standard folder layout that bundles both, so one package can be installed in many different tools. Before this, developers often had to build a separate package for each tool, even when the contents were identical. Upstash, the company behind the Context7 service, described needing four separate plugins just to offer one MCP server.

According to the official list of compatible clients, the standard now works with ChatGPT, Codex, Cursor, GitHub Copilot, VS Code, Amazon’s Kiro, and several other tools. Skills and MCP settings carry over between them. Tool-specific extras, such as custom commands still have to be set up for each tool. Vercel started the proposal. Amazon Web Services, Anysphere (the company behind Cursor), GitHub, Microsoft, and OpenAI helped refine it. The steering committee that runs the standard has five seats, held by Amazon, Anysphere, Microsoft, OpenAI, and Vercel. Google announced on the same day that it was joining as a core maintainer, which brings the group to six companies, and Anthropic is not one of them. It is worth noticing who is at that table — Microsoft owns both VS Code and GitHub Copilot. Amazon has Kiro. Vercel is a platform for deploying AI applications. These companies compete with each other in many areas, yet they agreed on a folder layout for building blocks that Anthropic created.

I have watched this pattern before in enterprise software. The company that invents a standard does not always end up in charge of who benefits from it. It is often the power sits with whoever controls distribution, meaning the tools people open everyday.

Why “Built on Anthropic’s Tech” is No Longer a Safe Shortcut

Until recently, marketing leaders comparing AI agent vendors could use a simple rule of thumb. If a tool was built on Anthropic’s MCP, it probably worked well with other MCP tools. That rule is now too thin to rely on. MCP only covers how an AI tool talks to other software. How a tool is packaged and installed across different products is a separate question, and a group that does not include Anthropic now decides it.

In practice, when a vendor says its tools are “MCP-native”, that tells you how the tool connects. It does not tell you whether the same tool will install cleanly in ChatGPT, Copilot, or Cursor. A better question is whether the vendor’s packages follow the Agent Plugins standard, and whether the vendor has any say in how that standard changes. A vendor with no plan to follow it, and no voice in its future is the one most exposed to lock in.

Agent Plugins does not remove the risk of tools failing to work together. It moves the risk. The old question was whether systems could connect at all. The new question is who decides how they connect. I have written before about AI assistants starting to compare software vendors on a buyer’s behalf, and about crawler access controls. Both follow the same logic. Whoever controls the access layer influences which tools get considered at all.

My Personal Anecdote

MCP joined my knowledge base when I started dabbling with vibe coding more than a year ago (how time flies ). I started with Cursor, the AI-native application for B2B work. And then tried VS Code, and got stuck to it ever since. It was until Claude released MCP and a few months later, I looked it up to my astonishment — “what the hell was I reading into???”. As a non-backend-engineer/developer, I thought to myself, this MCP seemed like its going to be here for a long time. Heck, it eventually is and how convenient it was for having MCP to plug into pipelines and workflows ( what joy!). Thereafter, I decided to keep using MCP throughout to test drive some of my automations, and it brought me immense joy because my tools and plugins can finally talk to each other. But not all though due to incompatibility versions at that period of time. It was after several months that MCP servers started processing heavily because of the growing complexities surrounding workflows and a growing team of agents . And that is when I cut down and improvise in areas where MCP servers are not necessary to exist in every workflow. Here’s some reasons why not having many MCP servers are good. And another one here, “Do You Really Need MCP Servers? A Senior Developer’s Reality Check“.

This is just a gist of how MCP is designed as a flexible system that connects AI models with external tools, and data sources through three main components: servers, clients and hosts. It enables smooth and efficient interaction, as well as provide secure, scalable and real time access without custom integrations.

Final Thoughts: The Bottom Line

Anthropic created MCP and Agent Skills. In December 2025, it gave MCP to the Agentic AI Foundation under the Linux Foundation so that no single company would own it. Little did we know about eight months later — on 6 August 2026 — a different group of companies announced the packaging format that decides whether an AI skill installs across the tools your team already uses. Anthropic was not part of that group, and Google joined it on launch day.

None of this makes Anthropic’s work less important. It does mean “built on Anthropic’s tech” no longer tells you much about how easily a tool works with others, or how locked in you might be. The next time an AI vendor pitches you on its protocol pedigree, ask which standard its packaging follows and whether it has a voice in how that standard changes. That is the first thing I check, and it is worth adding to your own buying checklist. If you want a working example of what that checklist looks like applied to a live MarTech stack, the LITV AI SEO Agent is built to make exactly that kind of governance visibility a standard feature, not as an afterthought.

Frequently Asked Questions (FAQs)

What is Agent Plugins 1.0?

Agent Plugins 1.0 is a vendor-neutral packaging format (a shared way to package AI add-ons), delivered on 6 August 2026 by OpenAI, Amazon, Microsoft, Anysphere, and Vercel that bundles Model Context Protocol and Agent Skills into one folder that works across ChatGPT, Codex, Cursor, GitHub Copilot, VS Code, and other compatible tools.

Did Anthropic build Model Context Protocol?

Yes. Anthropic released MCP in November 2024 and Agent Skills later. In December 2025, Anthropic donated MCP to the Agentic AI Foundation, a group under the Linux Foundation that it co-founded with Block and OpenAI.

Was Anthropic a founding member of Agent Plugins 1.0?

No. Anthropic is not on the steering committee, and is not among the companies named as helping refine the standard. Google joined as a core maintainer on the same day as the announcement.

Does using Claude lock a company into Anthropic’s tooling ecosystem?

Not necessarily. MCP and Agent Skills are open, and Agent Plugins lets the same package install in several other tools. One thing to check is that, at the time of writing in October 2026, Claude Code is not on the official list of compatible clients. Confirm support for the exact tools your team uses.

Why does this matter for B2B marketing leaders shortlisting AI agent tools?

“Built on Anthropic’s tech” no longer tells you who controls how easily a tool works with others. The packaging standard is now decided by a different group, so that is where your questions should go.

How should a marketing leader evaluate an AI agent vendor’s interoperability claims now?

Ask four things:

• Which packaging standard does the vendor follow, and which version?
• Which of your team’s existing tools has the vendor actually tested it in, and can it show you?
• Does the product rely on tool-specific extras that will not carry over?
• Does the vendor take part in the standards group or follow its roadmap?

Validity’s 2026 Email Deliverability Benchmark Report found APAC senders reaching just 78.4 percent inbox placement, against 84.3 percent in North America. That six-point gap comes down to things a Singapore-based B2B team can fix themselves: whether their email domain is trusted, the technical settings prove their emails are genuine, and their mailing lists are clean and up to date. I have audited Singapore B2B senders whose content quality was years ahead of their list hygiene. The inbox placement gap was never an AI problem.

That Six-Point Gap

Getting emails into inboxes is not just a successful number that looks good in a report. It depends on various surrounding factors that shape the email delivery rate. It also determines whether people see your campaign content in their main inbox, find it in the promotions folder, or miss it entirely. APAC’s inbox placement rate of 78.4 percent, compared to North America’s 84.3 percent, means roughly one in five APAC B2B emails fails to reach an inbox where someone can open, read, or respond to it. For teams working with smaller mailing lists, and fewer sales opportunities than their North American counterparts, that is a real loss. It also reduces the value of everything else they spend on marketing.

This gap is not caused by AI-powered spam filtering. Litmus’ 2026 report on email production speed supports that view. Across the market, teams have cut the time it takes to produce an email from roughly two weeks to three days, trying to keep up with the pace of content driven by AI search. But sending emails faster without keeping mailing lists clean and the sender’s reputation healthy only makes the existing problem worse. A broken process does not improve when you speed it up, since it just adds on new problems on top of the old problems like a vicious cycle that never ends.

Gmail and Outlook judge whether to trust a sender by checking things such as whether emails pass identity checks, how often people mark them as spam, how many fail to reach valid addresses, and how recipients have responded to previous emails. These checks are tied to the sender’s email domain and email sending server, not based on how good or rich the content is. These email recipients do not usually have the time to read outside-of-work emails that are not expected during their work schedule. Email clients do not need AI to work, and have been using these checks since long before Generative AI existed. For the first time, a public comparison puts a clear number on how far APAC senders are falling behind on these basics.

Authentication Discipline: SPF, DKIM and DMARC are Not Optional

Three records decide whether a mailbox provider trusts that an email actually came from the domain it claims to. Sender Policy Framework (SPF) lists which servers are authorised to send on a domain’s behalf. DomainKeys Identified Mail (DKIM) cryptographically signs each message so the receiving server can confirm it was not altered in transit. Domain-based Message Authentication, Reporting and Conformance (DMARC) tells receiving servers what to do when SPF or DKIM fails, and reports back on failures so a sender can see abuse or misconfiguration.

A Singapore B2B team running multiple sending tools, a CRM, a marketing automation platform, and a transactional email service, each need to be declared in that domain’s SPF record and each need their own DKIM signature. Miss one, and every message from that tool sends with a lower trust signal than the rest of the domain’s traffic, even though the domain itself is properly configured elsewhere. This is the single most common gap I find in Singapore and APAC’s B2B stacks: authentication configured once at setup, then never revisited as tools were added or swapped.

DMARC deserves particular attention because it is the record most APAC senders either skip entirely or set to a policy of “none,” which reports on failures without acting on them. A policy of “none” tells a mailbox provider the domain owner is watching but not enforcing, which is a weaker trust signal than “quarantine” or “reject.” Moving from monitoring to enforcement, once the SPF and DKIM records are confirmed correct across every sending tool — is one of the highest-value half-day projects a marketing operations lead can run this quarter.

List Hygiene is a Practitioner Habit, Not a Software Feature

Authentication earns a domain the right to be evaluated fairly. Engagement earns its right to stay in the inbox tab. Mailbox providers track how recipients interact with a sender over time: opens, clicks, replies, deletions without opening, and moves to spam. A list padded with contacts who have not engaged in a year drags that engagement signal down for every subscriber on it, including the genuinely active ones.

The practitioner habit that most Singapore and APAC B2B teams skip is a scheduled sunset policy: contacts who have not opened or clicked in 90 to 120 days move to a re-engagement sequence, and contacts who do not respond to that sequence are suppressed from regular sends entirely. This is pretty common, and definitely not uncomfortable because it shrinks the list on a dashboard marketing team sees. It is also the fastest lever available for improving inbox placement because it removes the exact signal mailbox providers weight most heavily.

As discussed in the The First-Party Data Imperative, teams need to fix their email delivery problems before making plans to protect themselves from losing search traffic to AI. If their own email channel is already falling six percentage points behind in reaching inboxes — even before they measure AI’s impact — they are losing on two fronts where AI platforms are taking away search traffic, while neglected email systems are stopping them from reaching people directly.

My Personal Anecdote

When I audit a B2B sender’s deliverability, the order matters critically than the checklist itself. First, pull the domain’s current SPF, DKIM and DMARC records and confirm every sending tool is declared and correctly signing. Second, review the DMARC policy setting and its aggregate reports for the last 30 days, looking specifically for authentication failures tied to tools nobody remembers adding. Third, segment the list by engagement recency and calculate what percentage has not opened anything in 90 days. Fourth, confirm whether that unengaged segment is still receiving full-frequency sends, which is the single most common finding in this order.

These steps need to happen in this order because trust problems build on each other. If email identity checks fail, providers lose trust in your sending domain, even if your content is good. If fewer people then open or respond to your emails, that trust drops further. Cleaning up your mailing list while leaving a broken DMARC setting, helps providers check whether your emails are genuine could mean things are being fixed in the wrong order of steps. This is the same layered-fix discipline behind MarTech stack rationalisation for APAC B2B: fix the foundation before adding another tool on top of it.

From a Singapore practitioner’s view, this is also where a CMO earns their fee differently; a Southeast Asia engagement typically starts with exactly this audit because it is fast to run, cheap to fix, and produces a measurable inbox placement improvement inside one sending cycle, long before any broad strategy conversation is useful. If the domain cannot reliably reach the inbox, no amount of positioning work matters, a discipline that sits alongside the governance thinking in Agentic AI Governance: What Singapore CMOs Must Build First.

Final Thoughts: The Bottom Line

The 2026 benchmark numbers settle an argument that should not have needed settling. APAC B2B senders are not losing inbox placement to AI filtering they cannot control. They are losing it to authentication gaps and neglected list hygiene, which they could have fixed at any point in the last several years, and the six-point regional gap against North America is the proof.

1. Authentication — this sets the trust limit everything else works within.
2. List hygiene — once that trust limit is raised, cleaning your mailing list gives the quickest improvement.
3. Speed of production — as the industry’s rush towards faster email production shows, making a broken process faster doesn’t fix it. It just makes you hit the same problems sooner than later.

Ready to see where your own sending domain stands before the next campaign goes out? Reach out to me here.

Frequently Asked Questions (FAQs)

Why is B2B email deliverability lower in APAC than North America?

APAC B2B senders reach 78.4% inbox placement against 84.3% in North America (Validity, 2026), a gap driven by domain reputation and list hygiene, not AI-specific filtering. Singapore-based teams close it through authentication discipline and engagement-based list maintenance, not by changing content for AI search.

What are SPF, DKIM and DMARC, and why do they affect inbox placement?

SPF authorises which servers can send for a domain, DKIM cryptographically signs each message, and DMARC tells mailbox providers what to do when either check fails. Misconfigured or missing records across sending tools are the most common cause of poor inbox placement in Singapore and Hong Kong B2B stacks.

How often should a B2B list be checked for engagement decay?

Segment monthly, and move contacts unopened for 90 to 120 days into a re-engagement sequence before suppressing them from regular sends. This engagement-based sunset policy is the fastest lever for improving inbox placement once authentication records are confirmed correct.

Is the APAC deliverability gap caused by AI email filtering?

No. Validity’s 2026 benchmark and Litmus’s 2026 reporting both point to domain reputation and list hygiene predating any AI-specific filtering change. The gap exists because of unmaintained authentication and engagement signals, not because AI models are evaluating content differently.

What should a fractional CMO check first in a deliverability audit?

SPF, DKIM and DMARC records across every sending tool first, then DMARC aggregate reports for authentication failures, then list engagement recency. Fixing authentication before list hygiene matters because reputation signals compound in that order.

Does faster email production make the APAC deliverability gap worse?

It can. Litmus’s 2026 reporting shows production cycles compressing from roughly two weeks to three days industry-wide. Speed without authentication and list hygiene discipline widens the exact reputation gap APAC senders already carry, rather than closing it.

Agentic AI browsers such as Perplexity Comet and OpenAI’s browser-based ChatGPT agent run on Chromium, so they present standard Chrome user-agent strings. Google Analytics 4 has no agent-detection layer, so these sessions record as ordinary human Chrome visits, inflating direct traffic and understating true agent-driven activity across B2B analytics dashboards.

The Browser Everyone Trusts is No Longer Only Human

Chromium Became the Default Skin for AI Agents

When OpenAI retired its standalone Atlas browser in July 2026, it did not retreat from agentic browsing. It folded the capability into ChatGPT itself, delivering a browser-based agent that opens tabs, clicks through checkout flows, and fills forms on a user’s behalf. Perplexity’s Comet does the same, and both sit on Chromium, the open-source engine behind Chrome (by Google) itself. Chromium is free, well documented, and already compatible with every website’s rendering assumptions, so it was the obvious choice for agent builders who needed to run a browser with speed. The ultimate consequence that nobody has anticipated: an agent running on Chromium requests a page exactly the way a human on Chrome does, headers and all. As a result, some platforms are quick to detect bot traffic (non-human behavioural pattern navigating the login webpage), and would instantly block AI agents from accessing.

OpenAI’s own help centre now frames this as a deliberate product evolution rather than a stopgap, describing the shift as evolving a standalone browser into agentic work performed inside ChatGPT directly. That framing matters for B2B marketers because it signals that the capability is going deeper into mainstream products, not fading out as a niche or home based experiment. The agentic browser is part of a fundamental learning-and-growing phase. It is the default interface where a growing share of AI users will route routine web tasks through, and every one of those tasks touches a website somewhere in a B2B marketing funnel.

HumanSecurity’s Number is the One CMOs Should Sit With

HumanSecurity’s State of Agentic Traffic report, published on April 2026, puts browser-based agents at roughly 71% of all observed agentic activity online. That is not a typical behaviour confined to developer sandboxes. It is the majority mode by which AI agents now move through the open web, and it arrives at your site carrying a user-agent string your analytics stack was never built to question in the first place. The remaining agentic activity runs through APIs and headless automation, which at least leave a distinct technical fingerprint. Browser-based agents leave none because they were never designed to be caught. Instead, they were only designed to work. However, this will change in due time as agentic infrastructure evolves around how businesses function after one to two years’ time, when AI adoption grows and matures.

Why GA4 Cannot Tell the Difference

User-Agent Strings Were Never a Trust Signal

GA4 classifies a session by parsing the browser and operating system out of the user-agent string the browser volunteers. That string is self-reported, and it always has been. Traditional bot detection works by spotting obvious red flags, like suspicious web traffic sources, blacklisted IP addresses, or clicks that happen faster than any human could manage. A well-build agentic browser hardly sets off any of those warning alerts. Comet and ChatGPT’s browser agents do not rush. Instead, these AI agents browser in a way a careful human would, like waiting for pages to fully load, allowing content to appear, and filling in forms slowly and deliberately, not all at once like a prompt-scripted machine.

Direct Traffic Was Already a Dumping Ground

Marketing teams have spent years treating an inflated direct channel as a known-unknown, a bucket for dark social shares, copy-pasted URLs, and app-to-browser handoffs GA4 cannot trace a referrer for. Agentic browser sessions now sit inside that same bucket, indistinguishable by any label GA4 exposes. A CMO reviewing quarter-on-quarter direct growth has no native way to ask what fraction of that lift is a Chromium-based agent completing a task, not a person reading a page. The bucket that used to absorb noise now absorbs a specific, growing, and economically meaningful traffic class, and it does so silently.

What This Breaks in a B2B Reporting Stack

Lead Quality Reports Inherit the Data Contamination

If an agentic browser fills a demo-request form while executing a task on a buyer’s behalf, that submission lands in the CRM looking like a qualified lead from a human visit. Sales development representatives then work a lead record with no signal that the browsing behaviour behind it was automated. Search Engine Journal’s 2026 coverage of Google’s own guidance on AI agents notes that publishers and marketers are only beginning to build the infrastructure to treat agent visits as their own traffic class, distinct from both human sessions and known crawler bots. Most B2B MarTech stacks are behind that curve, not ahead of it, and a sales team chasing a “phantom lead” spends real hours on a session that was never a buying signal.

Attribution Models Assume a Human Journey That May Not Have Happened

Multi-touch attribution models are built on an assumption: a session reflects a person moving through a consideration journey, one touchpoint building on the last. An agentic browser executing a single delegated task, checking a price, comparing two vendor pages, filling one form, does not have a consideration journey in the way the model expects. This is a distinct failure from the zero-click problem this site has already covered (see LadyinTechverse’s earlier insights on why B2B marketing attribution is broken in the AI search era). That piece addressed AI answers that replace a web visit. But, the problem flips: the visit does happen and gets counted, just under the wrong label — human instead of AI. That is harder to spot because the dashboard shows the wrong answer, which is just a number that looks quite right but it actually isn’t. A missing page session raises questions, a misclassified one raises none, until someone looks deeper into it.

What a Singapore Practitioner Checks First

From a Singapore practitioner’s view, the fix does not start with a new analytics platform. It starts with an honest audit of what “Direct” and “Chrome” have been quietly absorbing for the past year. Three practical checks belong on that audit: flag “Direct Sessions” with unnaturally short duration paired with high page depth, watch for spikes in form-fill volume that break away from historical patterns, and isolate sessions moving faster than a careful human, yet smoother than a crude bot. This is the signature of a well-built agentic browser. None of this requires new tooling, and it certainly doesn’t require tracking on every AI firm’s release schedule; it simple means reading the existing GA4 export as behavioural evidence rather than a finalised static number.

Related governance discipline around AI crawler access, covered in LadyinTechverse’s GEO framework for CMOs, applies here too: what you allow to read your site and what you allow to act on it are two separate decisions, and most stacks have only made the first one deliberately. Crawler access control governs whether an AI system can index your content at all. Agentic browser traffic governs something else entirely, whether an AI system can act on your site once it has decided to visit, and not everyone has built a policy for that second question yet. Across SEA-6 markets, the compliance conversation has focused on data protection, and AI-generated content disclosure. Traffic misclassification has not yet reached that agenda, which makes it exactly the kind of gap a CMO gets paid to find before C-level team asks about it.

My Personal Anecdote: When the Dashboard Stopped Telling the Whole Story

As a CMO practitioner and digital marketer, I used to treat Chrome traffic, direct visits and form submissions as reasonably reliable signals of human intent. Those numbers shaped my input: where I invested budget, which content I prioritised, and which campaigns I believed were influencing the pipeline.

Then the rise of agentic AI “headless or chromeless mode” browsers made me question what I was really measuring. A visit that appeared to be an ordinary Chrome session could have been a GPT bot or an AI agent comparing vendors, reading our content or completing a form on someone’s behalf. The output still looked familiar, whereby, sessions, leads, conversions and attributed revenue showed performance. However, the meaning behind those numerical data had become ambiguous and increased uncertainty.

That outlook affects ROI in subtle but significant ways. We may overestimate the influence of a campaign, misread direct traffic as brand demand, or send sales teams after leads that do not represent genuine ‘human buying intent’. As a content and AI strategist, I now see analytics less as a final answer and more as one layer of evidence. The question here seeks difficult answers, “How many people visited?”, but “Who or what created this signal, and did it contribute to revenue?”.

This is why some startups, mid-age and established organisations are moving closer to the market presence again to reconnect globally: investing in field marketing managers who can run end-to-end in-person meetings, conferences, virtual events, roundtables and brand partnerships. When web and AI-search measurement becomes increasingly difficult to interpret, trusted human interactions create the context that dashboards cannot. They help connect awareness to conversations, conversations to relationships, and relationships to revenue.

For me, the lesson is pretty clear: digital measurement still matters, but it can no longer stand alone on its legs like it used to. The traditional digital marketing legwork is sidetracked. The future of marketing ROI will belong to teams that combine cleaner analytics (MarTech) with relevant human touchpoints, and tell the story between the data points.

Retrieval Failures and Attribution Failures Share a Root Cause

LadyinTechverse’s earlier breakdown of the three retrieval failure modes behind wrong AI agent citations made a related point about a different layer of the stack: systems built before agentic AI became mainstream were not designed to be interrogated by it, and the gaps show up as misclassification rather than optical errors. Analytics is the same story told in traffic, and not in citations. A dashboard that has never been wrong can still have been wrong for several months, and the fix in both cases is the same discipline: treat every automated system that touches your stack as a class of visitor you have to name AI as, not one you can leave as ‘others’ or ‘miscellaneous’.

Final Thoughts: The Bottom Line

A 71% figure from one vendor’s report will not settle every argument about how much of your “Direct” traffic is agentic. What it should settle is whether the question gets asked at all. B2B marketing leaders who wait for GA4 or Google to deliver finished native agent detection are choosing to keep reporting a number they already have reason to doubt. The practitioner move is to audit now, using session-level signals already sitting in the export, and to treat every future MarTech vendor conversation as an opportunity to ask directly whether agent traffic is separated from human traffic in the product roadmap, not based on assumptions. Agentic AI B2B activity in markets from Singapore to U.S., Western Europe, and the UK is not slowing down for anyone’s reporting cycle to catch up.

Build strategy that survives contact with agentic AI, not just the buzzwords about it. See how your full audit shows, and the availability of tooling dashboard to help you overcome the technical SEO many are not prepared for.

Frequently Asked Questions (FAQs)

Why does my Google Analytics show Chrome traffic that isn’t human?

Agentic AI browsers such as Perplexity Comet and OpenAI’s browser-based ChatGPT agent run on Chromium, so they present standard Chrome user-agent strings. GA4 has no agent-detection layer, so these sessions record as ordinary human Chrome visits, inflating direct traffic and understating true agent-driven activity.

What percentage of AI agent traffic runs through a browser?

HumanSecurity’s State of Agentic Traffic report, published April 2026, found roughly 71% of observed agentic activity runs through browser-based agents rather than APIs or headless automation, meaning most agent traffic arrives looking exactly like a human Chrome session.

Can GA4 detect AI agent sessions on its own?

Not currently. GA4 classifies sessions purely from self-reported user-agent strings and has no built-in layer to flag agentic browsers separately from human ones, so the distinction has to be reconstructed manually from session-level behavioural signals.

How is this different from AI search reducing website visits?

Zero-click AI answers replace a visit entirely, so the session never happens. Agentic browser traffic is the opposite problem: the visit does happen, gets recorded, and is misclassified as a human session, which is harder to catch because the dashboard shows no missing data at all.

What should a B2B marketing team check first?

Segment Direct traffic by session duration and page depth to isolate unnaturally efficient patterns, cross-reference form-fill timestamps against known agentic AI product launches, and flag sessions with behavioural velocity between a careful human and a scripted bot.

Does this affect lead quality reporting too?

Yes. If an agentic browser completes a demo-request form while executing a delegated task, that submission enters the CRM looking like a qualified human lead, and sales development reps then work a record with no signal the underlying session was automated.

The Real Number: What Four AI Subscriptions Actually Cost

I run LadyinTechverse’s content pipeline on neither this kind of stacked, multi-model budget, and I track it to the Singapore dollar every month. This piece is a reconciliation, not a hypothetical excel spreadsheet exercise.

Standard-tier pricing across the four leading AI model assistants sits close together. OpenAI prices ChatGPT Plus at USD 20 per month. Anthropic prices Claude Pro at USD 20 per month, Perplexity prices its Pro tier at USD 20 per month, Google prices AI Pro at USD 20 per month. Together these four AI model assistants cost a solo operator roughly USD 80 a month, or nearly USD 960 a year. Anyone layering in a workflow tool such as Make.com, n8n, Gumloop, an image model, a video model, or a dedicated agent framework on top of these four is looking at a higher annual cost again. I know a few who used to subscribe to these 4 big AI models. These days, all I hear is ChatGPT and Claude. For Google ecosystem users (those who hold android phones, uses Google workspace and emails, Google drive, Google Maps, etc.), Google AI Pro and above plans are pretty common to hear.

Why the Total Sum Stays Invisible

The reason this total sum stays hidden is inevitable, and this has nothing to do with too much moolah to spend. Each subscription appears on a different card statement, arrives on a different billing date, and is listed under a vendor name that rarely mentions “AI subscription”. Its because these are viewed separately, the four charges seem small and easy to overlook. Few solo operators compare multiple statements unless the are intentionally auditing expenses, and some solopreneurs run their own agencies have no finance team to do it for them. The fragmented billing simply makes it easy to miss.

This matters more for a solo operator than it would inside a larger team, where a finance or procurement function typically owns software spend review on a quarterly cadence. A solopreneur carries that function personally, on top of delivery, sales and everything else, which is exactly why the audit gets skipped. I have written before about how AI-native CRM platforms quietly duplicate capabilities a business already pays for elsewhere in the martech stack, and the same blind spot applies to the AI assistant layer sitting underneath it.

Why Solo Operators Only Need a Minimum of 2 AI Assistants

A solo operator does not need four AI subscriptions to build an effective content workflow. Two well-chosen assistants can cover the essentials: one for producing and refining work, and another for research, verification, or handling longer briefs.

ChatGPT works well as a rapid-response partner for brainstorming, quick drafts, and fast back-and-forth iteration. Claude is useful for longer, structured assignments that require a large brief to stay in context, which is why some work outputs are produced within a Claude-run agent pipeline. Together, they can support most of the core writing process.

Additional tools become worthwhile only when they solve a specific operational gap. Perplexity can add real-time source verification for agencies publishing factual claims (research-focused), while Google AI Pro can be valuable for marketers who spend much of their day working in Docs, Sheets, Slides, and Forms. The goal is not to collect every capable model, but to start with one assistant first, then two assistants once you get the hang of it. That is enough to cover the work, and only add another assistant when its specialised function justifies the cost.

You would agree where this logic takes you. It is also exactly what makes the total sum easy to justify one subscription at a time and hard to question as a stack. Each individual USD 20 charge looks small enough to wave through without scrutiny. The stack-level number, close to USD 80 a month is the one that actually deserves a decision, and it sits alongside every other pricing decision a solo operator makes, including how they price their own fractional delivery work now that AI compresses the time a project takes.

My Personal Anecdote

When I first started out with Generative AI, nobody warned me that learning the tools would be the tricky part, and that filtering through all the noise would be the real struggle. AI-native tools, AI-assisted tools, and established brands bolting a chatbot on as a shiny new plugin all arrived at once. It felt like a mammoth maze made of unsolvable puzzles. It took me a few months of trials and errors to filter through and work out what works for me, what I like best, and what suits me. If you have doom scrolled past them or heard the names before, chances are I tried it.

My journey was bumpy and 100% non-linear, partly because OpenAI, Gemini, and Claude were still in the early stages of iterating their own GenAI models. I dabbled with Midjourney, pushed out AI-animated series through Leonardo AI, and tested Canva Pro, Jasper AI, and HyperWrite AI. On the automation side, I tried n8n first, then Gumloop, then Make.com, and realised that n8n was not for me.

I did not jump into subscribing to everything at once. My rule is simple. A trial of two weeks to one month has to begin the moment I sign up, otherwise I will not bother. As for tracking, I did not use a spreadsheet. After many years of work, I am tired of looking at cells and rows, and I did not need that type of approach here. Instead, I drafted my expenditure list with its invoice dates, line by line, and I go through it every month in Singapore dollars.

As each month passed, my trials and subscriptions leaned more towards the big four AI firms, and not Copilot for sure. Gemini led on image and video generation out of the three, excluding Perplexity. I let Google AI Pro run on my subscription money for a couple of months, given its decent and quite generous three-month introductory offer at SGD 7.28 per month. Then OpenAI caught up and improved tremendously in its image models. I dropped Gemini, and since then I have never looked back.

Ever since Claude Design (formerly called Claude Design System) came out, I found it tardy, a confused model that cannot get granular edits right for the many times it was expected to, and it wasted the free credits that Claude had distributed upon its launch. It was intended to replace my Canva, but after several attempts to train it to respond accordingly, it got better but I left it to code thereafter. Canva is great at producing consistent work because its focus is on templates and canvases of specific dimension sizes, anytime ready for a fast export. That holds even though it hosts poor, average, and slightly good quality images, videos, and design elements such as UI icons and graphics. It was incredibly hard for me to accept that, given how deeply trained and experienced I am in this field.

This is not about whether Canva can replace my skillsets. If I let it generate creative quality outputs, can it still replace my skillsets? Definitely not for me, and not for the other creative comrades who share the same sentiment. If I want a tool to replace my creative skillsets, it should deliver the optimal quality that humans naturally expect. Sadly, it cannot, and it never will because creativity has its “unique signature ID” in every human’s DNA. So, I had to drop my expectations much lower to accommodate mediocre outputs . To be fair, the quality of viral and high-engagement posts, filled with generic and comic-like visuals on Instagram and TikTok, contains uglified meme graphics and visual elements, an aesthetic that started on GIF making platforms, and then Tumblr .

Soundtrack, and sound effects generation was another rabbit hole I went down. I tested Suno AI, ElevenLabs, and a few others, and in my opinion, Suno is an interesting platform. What I like about it is its unique way of reading a prompt description and generating something as close as possible to a full music orchestration, from the instruments to the beats, rhythm, bass, tinkles, synthesisers, amplifications, pitch, tone, and vocals. ElevenLabs, on the other hand, was more focused on AI voice clones and radio-style podcasts, and it grew to run media generation features within the app.

Video generation is the one place I did not expect AI to catch up so fast. A few plugged-in AI model platforms have emerged as quite close to good AI animation, and Higgsfield has come out victorious. Tech Funding News reported in August 2026 that it raised USD 400 million at a USD 5.4 billion valuation. One detail worth separating out for developers is that Higgsfield’s GitHub repository is an open source, Apache 2.0 licensed GPU orchestration and machine learning framework for training large models. It is not the code behind its video platform.

Beyond subscriptions, I tried self-hosting LLMs with Ollama and Docker, among other tools I can no longer remember, and I built customised widgets and apps for mobility. Then came inventing Skills and Agents, which meant defending the work, integrity, and dedication I have put in with brain juice, energy, hours of late nights, and early mornings to deliver quite accurate outputs. No subscription does that part for you.

Whether you pay for one AI subscription or all four, the charges are real, and so is the total once they sit side by side on one invoice. Every tool I kept had to name a job that nothing else in my stack did as well, and every tool I dropped could not. That is the audit I am about to walk you through because the real 2026 cost of four AI subscriptions, only makes sense once each seat has earned its place.

The Stack Audit: From “Which Tool” to “Does Every Seat Earn Its Keep”

A stack-audit question looks different from a single-tool question. “Which AI tool should I use” wants one answer. “What does my current stack actually cost, and does every seat earn its keep” wants you to consolidate. To run this simple checklist audit, list every AI subscription you pay for, its monthly or annual charge, and the one job it does that nothing else in your stack does as well. If a subscription cannot name that job, it is not a permanent piece for your entire stack. You would either have to cut it or merge it into something you already pay for.

Once the audit sits on paper, most solo operators find at least one seat duplicating a capability already covered elsewhere in their stack. This is the same rationalisation logic I have set out in the four-layer AI operating system framework for solopreneurs, applied here to the specific case of overlapping AI subscriptions rather than the broader tool stack. The AI subscription layer is not exempt from that logic just because each individual charge looks trivial next to a full software licence.

From a Singapore practitioner’s view, the currency conversion matters more than it looks on paper. At roughly SGD 105 a month, four AI subscriptions add up to over SGD 1,200 a year, a real line item against a solo operator’s annual revenue, not a rounding error absorbed somewhere in overhead cost. Framed as an annual SGD figure rather than four scattered monthly USD charges, the decision to keep, cut or consolidate gets far easier to make a decision on since “the AI tools are getting expensive as workflows start to get more complicated”.

Final Thoughts: The Bottom Line

The solo operators who get real value from a multi-model stack are the ones who treat it like a stack, with each seat justified against a specific job and reviewed on a schedule, not the ones who let four separate billing emails quietly become a permanent fixture of overhead. The total for standard-tier ChatGPT Plus, Claude Pro, Perplexity Pro and Google AI Pro comes close to USD 80 a month, about SGD 105, as of September 2026. Reconciling that number is a 20-minute exercise.

Add every AI subscription your agency currently pays for, the monthly charge and the specific job each one does that nothing else in your stack does as well. If any seat cannot name that job, it is a consolidation candidate. Run the audit this month, before the next renewal date passes it by unexamined. For a fuller framework on scaling capability without adding headcount, the builder story behind LadyinTechverse’s own AI stack is worth the read.

Frequently Asked Questions (FAQs)

How much do AI subscriptions cost a solo marketer running an agency alone?

Running ChatGPT Plus, Claude Pro, Perplexity Pro and Google AI Pro together costs about USD 80, roughly SGD 105, a month at standard 2026 pricing, before any add-on automation, image, or agent tool. Most solo operators never total this because each subscription bills separately, on a different card.

Why do solo marketers not notice how much their AI subscriptions cost in total?

Each subscription renders on a different card statement, billing date and vendor name that rarely says “AI subscription” on the line item. Nobody scans four separate statements side by side unless deliberately reconciling a budget line, so the combined total stays invisible.

Do I actually need four separate AI subscriptions as a solo operator?

Only if each one does a job the others cannot cover as well. Fast drafting, long structured writing, real-time source verification and Workspace integration are genuinely different jobs. If a seat cannot name a distinct job, it is a consolidation candidate, not a permanent fixture.

How do I audit my current AI subscription stack?

List every AI subscription by name, its monthly charge, and the specific job it does that no other tool in your stack covers as well. Any subscription without a distinct job becomes a candidate for cutting or consolidating into a tool you already pay for elsewhere.

What does the AI subscription total look like annually for a Singapore-based solo operator?

At roughly SGD 105 a month, four standard-tier AI subscriptions add up to over SGD 1,200 a year. Framed as an annual SGD figure rather than four scattered monthly USD charges, the keep-cut-consolidate decision becomes far easier to make.

Does this AI subscription cost include automation tools or image generation add-ons?

No. The roughly USD 80 a month figure covers only the four standard-tier chat assistant subscriptions. Automation platforms, image generation tools and agent-orchestration add-ons sit on top of this total and raise the annual figure further.

Google’s Universal Cart, unveiled at I/O 2026 and rolling out through summer 2026, lets Search, the Gemini app, YouTube and Gmail complete a purchase inside a single AI-managed cart. OpenAI’s narrower Agentic Commerce Protocol (ACP) survives in app-based form after Instant Checkout was scaled back in March 2026 for weak conversion. For B2B marketing leaders, the mechanism that matters is not the cart. It is that vendor comparison and shortlisting are moving inside a surface marketing does not control, the same disintermediation AI search already caused for informational queries, now reaching the evaluation and transaction layer.

Like many marketers, I have built demand generation motions around a controlled funnel from lead to close. Agentic commerce protocols remove marketing’s visibility at exactly the stage that used to justify the budget line: the evaluation moment where a shortlist forms and a vendor gets chosen or dropped.

What Agentic Commerce Protocols Do

Google’s Universal Cart works by letting an AI agent hold context across Search, Gemini, YouTube and Gmail, then execute a transaction without the user leaving the conversation. According to Google’s own product announcement, the cart persists across these surfaces so a comparison started in one place can complete in another. OpenAI’s Agentic Commerce Protocol took a narrower path. Instant Checkout, its consumer-facing transaction feature, was scaled back in March 2026 after CNBC reported weak conversion, and the protocol now survives mainly in app-based integrations rather than a universal cart.

Universal Cart is a horizontal layer sitting across Google’s entire surface area. ACP is a narrower, app-specific mechanism. Both point at the same structural shift: an AI agent acting as a buyer who is clicking through a comparison page, now holds the shortlist.

Why B2B Vendor Selection is the Next Layer to Move

AI search already removed marketing’s visibility into the discovery stage. A buyer asks an AI system a question, gets a synthesised answer, and never lands on a vendor’s comparison page. Marketing teams adapted by building for AI citation, not just search ranking, a shift covered in a piece on AI crawler access control.

Agentic commerce protocols extend that same mechanism, one stage further into evaluation and transaction. When Universal Cart persists context across Search, Gemini and Gmail, the agent is not just answering a question. It is holding the comparison set, weighing options, and in some flows executing the purchase. For a B2B software or services buyer, that comparison set is the vendor shortlist. The marketing team that built content, case studies and comparison pages to win that stage now has no visibility into whether the agent even surfaced them.

TechCrunch’s coverage of the Universal Cart launch described it as designed to follow a buyer’s entire shopping journey across the Internet. For consumer retail, that is a convenience feature. For B2B vendor selection, it is a structural threat to the funnel marketing has spent a decade building.

What Marketing Teams Lose When Evaluation Moves Inside Chat

Three specific things disappear when an AI agent handles shortlisting.

The first is comparison page attribution. Marketing teams invest heavily in comparison content, pricing pages, testimonials, and case studies designed to win the moment a buyer evaluates two or three vendors side by side. If an agent performs that comparison inside a chat surface, the team has no way to know whether its content shaped the outcome, echoing the attribution blind spot documented in this piece on why B2B marketing attribution is broken in the AI search era.

The second is sales handoff timing. A traditional funnel assumes a buyer engages sales after building intent through content. If an agent completes both discovery and shortlisting before a human buyer speaks to anyone, sales inherits a decision already narrowed, sometimes already made with no record of what content or messaging influenced it.

The third is control over the comparison narrative itself. A vendor comparison page lets marketing frame the criteria that matter. An agent building its own comparison set from indexed content, structured data and prior context is not obligated to use those criteria at all. This is why owned, first-party relationships with buyers matter more, a point covered at length on the first-party data imperative.

None of this means the funnel disappears. It means the part marketing could see and shape is compressing into a layer marketing cannot yet read or comprehend. Ultimately, the buyer journey erodes the surrounding functions that once made Top of Funnel, Middle of Funnel and Bottom of Funnel (ToFu, MoFu and BoFu) a highly effective instrument.

What to Do Before the Evaluation Layer Fully Closes

The practical response starts with structured, machine-readable comparison content. If an agent is going to build its own shortlist from indexed material, that material needs to be legible to the agent, not just persuasive to a human reader. Pricing, feature comparisons and differentiators need to exist in a form an AI system can parse cleanly, not buried in marketing copy designed for human scanning.

The second response is instrumentation. Marketing teams should start tracking referral patterns from AI-mediated sessions the same way they built dark traffic baselines for AI search, treating any spike in direct or branded traffic following an agent-mediated evaluation as a signal worth investigating, even without full attribution.

The third is relationship depth before the evaluation stage. If an agent is going to shortlist vendors based on indexed authority and structured data, the vendors with the deepest first-party relationships, the strongest existing pipeline of warm buyers, and the clearest machine-readable proof of expertise have the advantage before the agent even starts comparing.

Final Thoughts: The Bottom Line

Agentic commerce protocols are not a shopping cart feature. They are the evaluation and transaction layer of B2B buying moving inside an interface marketing teams do not control. The funnel built around visible, owned evaluation stages is compressing, and the response is not to wait for full visibility to return. It is to make comparison content machine-legible now, build first-party relationships that survive disintermediation, and treat this as the same structural shift AI search already forced at the discovery stage, arriving one layer deeper. However, we are still at an early stage, and I do not see this window of opportunity expanding so far that agentic buying outgrows human buyers (tell me if I am wrong), unless one- to three-person startups start multiplying with full agentic AI operating at scale. This is where they may run into a lack of resources and require AI agents to help them decide and shortlist vendors.

Frequently Asked Questions

What is an agentic commerce protocol?

An agentic commerce protocol lets an AI agent hold shopping or evaluation context across multiple surfaces and complete a transaction without the user navigating away. Google’s Universal Cart and OpenAI’s Agentic Commerce Protocol are the two current implementations, differing mainly in scope: horizontal across Google’s ecosystem versus narrower app-based integrations.

How does agentic commerce change B2B vendor selection?

Agentic commerce protocols such as Google’s Universal Cart and OpenAI’s Agentic Commerce Protocol let AI agents complete transactions inside chat and search surfaces. For B2B buyers, the same mechanism extends to vendor comparison, meaning marketing teams lose visibility into evaluation the moment an agent handles discovery and shortlisting.

Does this mean B2B buyers will purchase software through a chat window?

Not necessarily as a completed transaction today. Most B2B purchases still require procurement, contracts and human sign-off. What changes first is the comparison and shortlisting stage, where an agent narrows options before a human buyer engages sales, not the final signature.

What happened to OpenAI’s Instant Checkout?

OpenAI scaled back Instant Checkout in March 2026 after weak conversion, reported by CNBC. The Agentic Commerce Protocol continues in a narrower, app-based form rather than a universal, cross-surface cart, which is the model Google pursued with its May 2026 launch.

Can marketing teams track traffic from agent-mediated evaluation sessions?

Not directly yet. There is no standard referral tag for agent-mediated sessions comparable to UTM parameters. Teams can approximate visibility by tracking spikes in branded or direct traffic following periods of AI search growth, similar to the dark traffic baseline method used for AI search attribution.

What should a comparison page look like for an AI agent to parse it correctly?

Pricing, feature sets and differentiators should exist as clearly labelled, structured content rather than narrative copy alone. Consistent terminology across pages, explicit criteria statements and machine-readable formatting give an agent a cleaner basis for comparison than persuasive prose written for a human scanning the page.