AI agents have left the chatbot tab.

The most useful AI story this week was not a model launch. It was where the agents moved.

Anthropic put Claude into Slack. Not as another private chat window, but as a thing you can tag into a channel, like a colleague. It remembers the relevant context, uses the tools you allow, works while you sleep, and leaves logs an admin can actually read. Anthropic also says its internal version now writes 65% of its product team’s code. Treat that as their claim rather than gospel, but the direction is hard to miss.

OpenAI did the same thing to spreadsheets. ChatGPT now sits inside Excel and Google Sheets, building models, running scenarios, explaining what it changed. The help page says the quiet part out loud: check the formulas and the changed cells before you trust the output.

And Cloudflare has been quietly building the unglamorous end of this. Browser sessions agents can drive, recordings, approval gates before risky actions, durable logs, recovery when something disconnects. The boring plumbing that decides whether an agent is usable or just a demo.

That is the pattern. AI is climbing out of the isolated chatbot tab and walking into the places where work already happens. Slack. Spreadsheets. Browsers. Dashboards. Codebases. Product feeds. Support queues.

The chatbot tab was useful. It got people comfortable. It let your team poke at AI, draft a few things, and realise the ground had shifted. But it was a safe little box, and the box is being opened.

What comes next is messier, and far more valuable. An agent that lives where the work happens needs to see company context. It needs tools. It needs memory. It needs a budget. It needs a line it cannot cross without a human saying yes. And it needs logs, because “Claude did it” is not an incident report.

This is where a lot of companies are about to get caught out.

They think they have an AI strategy because a few people have subscriptions and there’s a Slack channel called ai-experiments. That is not an operating model. That is a stationery cupboard with tokens in it.

The real questions are duller and more important.

What can the agent see? Which channels, which files, which dashboards, which client notes, which analytics account?

What can it do? Draft only, or edit, or publish, or send? Can it change a price? Move a CRM stage? Reply to a customer? Open a pull request? Touch the live site?

And who owns the review? Not in theory. In the actual week, when someone’s busy, the client wants an answer, the spreadsheet looks convincing, and the agent has produced something with a tidy summary at the top.

That last one is the trap. The most dangerous AI output is never obviously mad. It is tidy, plausible, and just close enough that a tired person waves it through.

So prompts are not the unlock. Prompts help, but they do not fix a broken business brain. If your data is scattered, your rules live in three people’s heads, and nobody can say what good looks like, AI mostly makes you faster at being confused.

For an agency, this should change the whole offer.

Selling “AI agents” as a shiny object is already tired. Every vendor has an agent now. Every other LinkedIn post has six of them, twelve automations, and a screenshot of a chat doing something suspiciously neat. The client does not need another mascot. They need someone to map the work.

Where does demand enter the business? Where does research live? Where does positioning get decided? Where do product facts sit? Where does the company actually learn from what happened last month? Map that, and AI finally has somewhere useful to stand.

A marketing team does not need “an AI content agent” floating in the abstract. It needs a bounded loop that pulls from Search Console, analytics, CRM notes and recent calls, then hands a human a draft brief to approve.

An ecommerce business does not need “agentic commerce strategy” on a slide. It needs product data that agrees with itself across the page, the feed, the schema and the reviews, so a buying agent can recommend it without taking a risk.

A service business does not need a chatbot bolted to its website. It needs a way to turn real questions, objections and case evidence into a system that gets a little sharper every week.

This is the post-agency shift, and it is not subtle. The old agency sold output: pages, posts, campaigns, reports. The next one builds the operating layer underneath all of that. The commercial brain, the workflows, the review loops, the measurement, and the human taste that stops the machine flooding the world with beige.

Copy still matters. Design still matters. SEO, ads, strategy, all of it still matters. The difference is they stop being isolated deliverables and become parts of a faster system. And a system has to be designed.

If I were advising an owner this week, I would not start with “which agent platform should we buy.” I would start much smaller. Pick one workflow where the pain is obvious and the risk is contained.

A weekly opportunity brief built from your analytics and sales notes. A product-data consistency check across twenty important SKUs. A support-ticket review that turns repeated objections into sales assets. A follow-up assistant that drafts the next message but cannot send it without a human nod.

Then build the boring scaffolding around it. One owner. One data map. One permission boundary. One approval path. One log. One success measure. One review cadence.

That sounds far less exciting than “we deployed six agents.” Good. Exciting is how you end up with a bot quietly editing the wrong spreadsheet, or inventing a client claim because nobody ever gave it a source of truth.

The better question was never “how many agents do we have.” It is “what part of this business can now run faster without becoming less trustworthy.”

That is where the money is. And that is where the next agency lives.

---

Jason Sibley is the founder of Cleo, a post-agency marketing and AI company. JasonVsTheNoise is where he writes about what is actually happening with AI, marketing, and how businesses should be thinking about both.