What you can honestly promise a client about AI work

Somewhere in your material there is a sentence about AI. On the services page, in the proposal template, in the thing you say twenty minutes into a first call. It promises a speed, an accuracy or a capability, and it was written in a good week, right after an afternoon where the model did something that genuinely surprised you. It has not been checked against a bad week since.

This guide is about the distance between that sentence and what you actually hand over, and about closing it yourself before a client, a regulator or a bad month closes it for you. It is written for a freelancer, solo operator or small team that sells work AI now helps produce: copy, bookkeeping, research, design, code, support, admin. It is not for building an AI product other people rely on, where the obligations are heavier than anything here. It is also not an argument about whether to use AI. Assume you do, and assume you say so.

The criticism that lands is the delivery gap

In August 2026 Dario Amodei, the CEO of Anthropic, said the public backlash against AI is “fundamentally a crisis of trust”, and offered a diagnosis that is more useful than it first looks: “by far the most accurate criticism of AI companies including Anthropic is that we haven’t yet delivered on our big promises to benefit the world” [1]. He is describing his own industry, at the scale of what he offered as the test of real delivery, “actually curing cancer” [1]. The structure of the complaint is the same one that operates at the scale of a 4,000-word content retainer.

Notice what the criticism is not. It is not that the technology fails, or that using it is somehow disreputable. It is that the promise was larger than the delivery, and that the gap was left for other people to find. That is the failure mode that survives every model release, every price change and every argument about regulation, because it is a failure of arithmetic between two sentences you control: the one you said, and the one describing what arrived. Everything else in this guide is a way of keeping those two sentences close together.

A capability claim is one you have to be able to prove

The rule that applies to your website is older than the technology, and enforcers have said so directly. When the US Federal Trade Commission announced Operation AI Comply on 25 September 2024, a sweep of five cases against companies using AI claims deceptively, the FTC’s then-chair Lina Khan put it as “Using AI tools to trick, mislead, or defraud people is illegal. The FTC’s enforcement actions make clear that there is no AI exemption from the laws on the books” [2].

The DoNotPay case in that sweep is the one worth reading if you sell professional services. The company marketed its product as “the world’s first robot lawyer” and, according to the complaint, told consumers they could “sue for assault without a lawyer” and “generate perfectly valid legal documents in no time” [2]. The allegation that should make you uncomfortable is procedural rather than moral: the FTC said the company had not tested whether its chatbot’s output matched the level of a human lawyer, and employed no attorneys [2]. Under the settlement DoNotPay agreed to pay $193,000 and to notify subscribers from 2021 to 2023 about the service’s limitations, with the order barring claims that its product substitutes for a professional service without evidence to back them [2]. The problem was not that a machine wrote the documents. It was that nobody had checked the claim before it went on the page.

The same rule runs in the other direction, which catches people out. On 18 March 2024 the US Securities and Exchange Commission settled charges against two investment advisers, Delphia and Global Predictions, for false and misleading statements about their use of AI, with civil penalties of $225,000 and $175,000 respectively [3]. Then-chair Gary Gensler’s line was “Investment advisers should not mislead the public by saying they are using an AI model when they are not. Such AI washing hurts investors” [3]. Overstating what your AI does and overstating that you have any are the same offence. If your site says “AI-powered” and what you have is a ChatGPT tab open in the next window, you have made a claim you would struggle to describe under questioning.

Your supplier’s terms are the ceiling on what you can promise

You cannot sell a guarantee upstream of you that your own supplier explicitly refuses to give. This is the fastest sanity check available, it takes about ten minutes, and almost nobody does it.

OpenAI’s terms of use, effective 1 January 2026, state that “Output may not always be accurate” and that “You should not rely on Output from our Services as a sole source of truth or factual information, or as a substitute for professional advice” [4]. They go further and put a duty on you: “You must evaluate Output for accuracy and appropriateness for your use case, including using human review as appropriate, before using or sharing Output from the Services” [4]. They also prohibit using output relating to a person “for any purpose that could have a legal or material impact on that person, such as making credit, educational, employment, housing, insurance, legal, medical, or other important decisions about them” [4], and separately bar you from representing “that Output was human-generated when it was not” [4]. OpenAI’s usage policies, effective 29 October 2025, add the practice rules: no “provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional”, and no “automation of high-stakes decisions in sensitive areas without human review” [5].

Anthropic’s usage policy, effective 15 September 2025, is more specific about what you have to tell people. A consumer-facing chatbot “must disclose to users that they are interacting with AI rather than a human”, and that disclosure has to appear “at a minimum at the beginning of each chat session” [6]. For high-risk uses, a list that includes legal interpretation, healthcare and mental health, insurance underwriting and claims, financial and lending decisions, hiring and housing eligibility, academic assessment, and journalistic content, two conditions attach: “a qualified professional in that field must review the content or decision prior to dissemination or finalization”, and “If model outputs are presented directly to individuals or consumers, you must disclose to them that you are using AI to help produce your advice, decisions, or recommendations” [6].

Read those as pricing information rather than legal reading. Human review is not a nice-to-have you can drop when a deadline compresses; it is a condition of the licence you are working under, and it has an hourly cost that belongs in your fee. If a client asks for a turnaround that leaves no room for it, the honest answer is that the turnaround is not available, not that the review quietly stops happening.

Promise the process, because that is the part you control

The rewrite is mechanical once you see it. Take each claim and ask what a sceptical client could verify without taking your word for anything. “AI-powered accuracy” fails, because there is no observation that would settle it. “Every deliverable is read line by line by me before it goes out, and I carry the errors” passes, because the client can test it by finding an error and watching what you do. “Cuts your turnaround in half” fails unless you can show the two measurements. “Your monthly report lands by the fourth working day” passes, because a calendar settles it.

The pattern is that outcome claims about a model’s correctness are unbackable, and claims about your own process, accountability and turnaround are checkable. So make the second kind. This also happens to be the honest description of what changed when you started using AI: the work is faster to draft, and the checking is now the job. Say that. A client who understands that they are paying for judgement over a fast draft is a client who is not surprised later, and being unsurprised is the whole of trust as a commercial matter.

Two specifics are worth writing down rather than leaving implied. The first is who is accountable, in one sentence, with a name in it. The second is what you disclose. There is no general obligation to itemise every tool you touch, but there is a floor: do not present model output as human-written when someone asks directly [4], and follow the disclosure rules that attach to consumer-facing chatbots and high-risk advice [6]. Between the floor and full itemisation, pick a position and apply it consistently, because clients notice inconsistency long before they notice policy.

Nobody gets to promise the model will be right

Anthropic’s own documentation on reducing hallucinations, after listing the techniques that work best, says: “Remember, while these techniques significantly reduce hallucinations, they don’t eliminate them entirely. Always validate critical information, especially for high-stakes decisions” [7]. That sentence is written by the people with the most access, the most context and the most incentive to claim otherwise. Whatever your prompt setup, your accuracy ceiling is not above theirs.

The techniques are still worth using, because they change what an error looks like. Anthropic recommends explicitly giving the model permission to say it does not know, extracting word-for-word quotes from long documents before working on them, and having it cite quotes and sources for each claim so the response is auditable [7]. The value for your client work is that a quoted passage from their own document is checkable in seconds, while a confident paraphrase is checkable only by redoing the research. Grounding does not make the output correct; it makes wrongness cheap to find, which is what your review time is actually buying.

Promises also expire because the tool underneath them turns over. The Claude lineup available today is Opus 5 at $5 per million input tokens and $25 per million output, Sonnet 5 at $2 and $10, Haiku 4.5 at $1 and $5, and Fable 5.1 at $10 and $50 [8]. Those names and figures will move again. A promise anchored to what a particular model did in a demo in March has a shelf life measured in months, so anchor the promise to your process instead, and put a date in your calendar to re-run your standard task on whatever you are actually running now.

An hour to find the claims you cannot back

Start by collecting them, which is the step people skip because they assume they remember. Your services page, your pricing page, the proposal template, the onboarding email, the pitch deck, your profile on whatever marketplace sends you work, the LinkedIn post from a good week, and the sentence you say out loud on calls. Write each AI claim on one line.

Then sort into two piles. Pile one is claims a client could check: dates, who reviews, what you fix and at whose cost, what you disclose. Pile two is everything else: percentages with no measurement behind them, accuracy adjectives, comparisons to firms you have never benchmarked against, and the word “AI-powered” attached to nothing in particular. Pile two gets rewritten into pile one or deleted. If a claim is true and you can show the working, keep the working somewhere you could find it in five minutes, because that file is the entire difference between a defensible claim and an allegation you did not test [2].

Finish by pricing the review. If your material says a person checks the work, that person has hours, and those hours are either in your fee or in your evenings.

What still goes wrong

The rewrite makes your material duller, and duller loses some pitches. A competitor promising an accuracy figure they have not measured will sometimes win the job. That is a real cost and pretending otherwise is its own kind of over-promising. The compensation arrives late and quietly, in the clients who stay because nothing you said turned out to be untrue, and it is genuinely hard to feel that in the month you lose the work.

Process promises fail the same way any process fails, which is under load. A review step that survives a normal week gets skipped in the week three deadlines collide, and the claim on your site does not skip with it. The only durable fix is to keep the promise small enough that it holds on your worst week rather than your best one, and to say no to the turnaround that requires the exception. Enforcement bodies are not the main risk to a two-person firm here; the main risk is a client discovering, at the worst possible moment, that the thing you told them happens to every deliverable does not.

There is also a limit to how much of this you can settle in advance. The rules move: the OpenAI and Anthropic policies quoted above carry 2025 and 2026 effective dates [4][5][6], and both vendors update them without asking you. Diary a check twice a year against the terms and usage policies of whatever you actually use, and treat any promise you made that depends on a supplier’s behaviour as provisional until you have re-read the page it rests on.