Five places I still won't let AI near client work

Five places: any fact that leaves the building unchecked, a scare statistic borrowed from somebody else's job, a deliverable you've promised the client exclusive copyright in, work that carries a disclosure duty nobody's looked at, and confidential material going into a tool whose terms nobody has read.

Each of them costs real money, and each has evidence behind it rather than a feeling.

Everywhere else, use it, and use it hard. The gap between agencies isn't whether they use AI. Everybody does. It's how — and these five are the how.

You've heard about the lawyers. The detail that matters gets lost in the retelling.

In Mata v. Avianca, a filing went into a federal court containing case citations that did not exist. The judge's June 2023 order is a public document and the language is short: "A penalty of $5,000 is jointly and severally imposed on Respondents."

You don't file court documents. But swap the nouns and you've got a familiar week: a market-sizing number in a new business deck, a competitor claim in a strategy doc, a statistic in a client's blog post. The failure in that case wasn't that a machine invented something. Machines do that. The failure was that it went out of the building without anyone checking, and the person who signed it owned the consequence.

That's the rule, and it's not a new one. It's the same rule you already have about a junior's copy deck. AI didn't create the need to check work — it just made unchecked work much faster to produce.

You'll get quoted a hallucination rate. The one that gets quoted at agencies is Stanford's finding of 17% to 33%, and it is real research. It is also specifically about AI legal-research tools answering legal questions. That is not your job. It is not a measure of what happens when a model writes a product description with your client's own spec sheet in front of it.

A benchmark overview reports rates from 22% to 94% across twenty-six models on one task, while a different long-context benchmark puts leading results in the low single digits. Same technology. Wildly different answers, because the task, the model, the retrieval setup and the scoring method all changed.

Which means a single "AI hallucination rate" isn't a meaningful thing to hold an opinion about. If a vendor quotes you one, ask which model, which task, and measured how. And if you catch yourself repeating one in a client meeting, do the same to yourself.

The honest position is narrower and more useful: error risk scales with how far the model is working from source material you supplied. Summarising a document you handed it is a different act from asking it what's true about the world.

This is the one I see agencies walk into blind, and it's contractual, so it bites later.

The U.S. Copyright Office's position, published in January 2025, is that generative AI output can be protected by copyright only where a human author has determined sufficient expressive elements — and it says plainly that prompting alone does not provide that control.

The panic version of this is wrong. AI assistance does not automatically strip copyright from a larger work. Human-authored material that's perceptible in the output, creative selection and arrangement, real modifications — those carry protection. The line isn't "did a machine touch this." The line is whether a person made the expressive decisions.

Practically, for an agency: your standard contract may promise the client full ownership and exclusivity in every deliverable. If some of those deliverables are wholly generated with no meaningful human authorship, you may be promising something that isn't yours to give. Have that conversation with your lawyer before it becomes a dispute rather than after.

I'm not a lawyer, and nothing in this article is legal advice.

If any of your work touches the EU, the AI Act's transparency obligations under Article 50 have applied since 2 August 2026. Generative and interactive systems and deepfakes are among the named categories. Whether that reaches a given piece of your work is exactly the question, and it isn't one you can settle off the category list.

Two things decide what you owe. Whether you count as a provider or a deployer, because the duties differ — that's a question for a lawyer, not a box you tick. And the territory: this is the EU's rule. The platforms run their own disclosure policies separately from it, and when I went looking for the current wording from Meta, Google and TikTok I couldn't pin it down to something I'd repeat — which is itself the point. Check it per platform, per campaign.

What you can do is stop treating it as a legal problem to be solved once and start treating it as a line on the brief: does this piece need a disclosure, and who says so. Ask the client what they need disclosed before the work starts rather than after it ships. Most of them haven't thought about it, and being the agency that raised it first is a good look — it also means it's their answer on the record, not your assumption.

People will tell you confidently that pasting client material into a chatbot leaks it. I went looking for a named, documented case of that happening to an agency, with the contract term that would have prevented it. I couldn't find one to a standard I'd repeat.

OpenAI's business services agreement — which covers the API and the Enterprise and Business plans, and explicitly does not cover consumer ChatGPT — states that, except as expressly set out, it grants them no IP rights in your content and you no IP rights in the services. Which sounds reassuring and is — but the real protection comes from which agreement you're on, which plan, what retention is configured, and how the product is set up.

So the habit is simple: before an unreleased product, a set of financials, or an unannounced campaign goes into any tool, somebody at your shop should have read that tool's actual terms. Not the marketing page. The agreement. It takes an afternoon once, and then you know.

Those five are all work where the model is asserting something about the world, or where ownership and disclosure attach. That's a real slice of agency work, and it's a small one.

The large slice is production. Resizing, versioning, reformatting, drafting from a brief you wrote, working against a brand's own guidelines, generating options for a human to choose from. In that work the model isn't inventing facts, it's executing against material you gave it — and the risk profile is completely different. That's the same distinction I'd draw between a machine assembling variants and a person deciding what the ad should say.

Use it there without apology. Just know which side of the line you're standing on.

Open your standard client contract, find the clause where you promise ownership and exclusivity in the deliverables, and read it against what you actually shipped last month.

Ten minutes, and it's the one thing here you've already put your name to.

When should an agency not use AI?

Where the cost is documented. A fact that leaves the building unchecked — in Mata v. Avianca a US federal court imposed a $5,000 sanction over case citations that did not exist. A deliverable you've promised exclusive copyright in, when the US Copyright Office said in January 2025 that prompting alone doesn't meet the human-authorship bar. Work carrying a disclosure duty, with EU AI Act Article 50 in force since 2 August 2026. And client material going into a tool whose terms nobody has read.

What is the actual AI hallucination rate?

There isn't one. Stanford researchers measured factual hallucination rates of 17% to 33% in leading AI legal-research tools — that figure is about legal research and does not transfer to other work. A separate benchmark overview reports 22% to 94% across twenty-six models on one task, while a long-context benchmark puts leading results in the low single digits. The rate moves with the model, the task, the retrieval setup and the scoring method, so a single "AI hallucination rate" is not a meaningful number to hold an opinion about.

Can a client own the copyright in AI-generated work?

In the US, only where a human author determined sufficient expressive elements — the Copyright Office said in January 2025 that prompting alone doesn't meet that bar. AI assistance does not automatically defeat copyright in a larger human-authored work: perceptible human material, creative selection and arrangement, and real modification all still carry protection. Positions outside the US differ and I haven't verified them. Check what your own contracts currently promise.

Do I have to disclose that an ad was made with AI?

It depends on territory, role and format. Article 50 of the EU AI Act has imposed transparency obligations since 2 August 2026, per European Commission guidance, and whether you are a provider or a deployer changes the duty. Platform policies run separately from that rule. Ask the client what they need disclosed, and get the answer before the work starts.

Is it safe to put client material into a chatbot?

It depends on which agreement, plan and configuration you're on. OpenAI's business services agreement covers its API and the Enterprise and Business plans and explicitly does not cover consumer ChatGPT — so the plan an employee is pasting into decides the answer, not the marketing page. Read the terms once, decide a policy, and tell your team what it is.

• The ad you approved isn't the ad that ran — what the ad networks change without asking.
• GEO vs SEO: what the hell is the difference? — what the engines actually read before they answer.
• How to make money with AI visibility — what to sell, and who'll pay for it.