"What is an AI disclosure?" Owners ask it the week it first turns up in front of them: a line in a customer's new contract, a question on a buyer's due-diligence list, a box on the insurance renewal, a form from a platform the company sells through. Until then nobody asked, and the honest first answer is that the words mean less than they sound. An AI disclosure is a short, true statement of where your company uses AI in work a customer receives, who checks that work before it counts, and what the tools are never allowed to do on their own. That is all. Writing one is easy if the company has already made three decisions, and impossible to do truthfully if it has not.
I was COO and then President of a public company that closed more than forty acquisitions, and I ran the integration on most of them, which meant answering many questionnaires from buyers, lenders and insurers about things nobody had written down. I am a lawyer by training, and nothing here is legal advice; the wording of any contract, questionnaire or form you were given is for you and your lawyer, and it comes first. Now, at 013 Labs, I write fixed-fee AI plans for CEOs of private-equity-backed and owner-led mid-market companies, and the disclosure is usually the last page of the plan, because by then it writes itself.
The short answer
A disclosure is a statement of fact about your own company, not a promise and not a policy. It answers four plain questions in a few sentences. Where do tools do part of the work a customer receives? What information do they touch? Who, by role, checks the result before it goes out? And what do the tools never do alone? If you can answer those four from memory, you can write the disclosure in an afternoon. If you cannot, the gap is not in the writing, it is in the company, and the questions on the page about who owns AI in a company are the place to start.
The statement is true or it is not; that is the whole test. It is not a marketing claim, not a reassurance, and not a place to describe what you intend to do next quarter. A buyer, a lender or an insurer who asks is asking one thing underneath: does a person stand behind the work, and would the company notice if a tool got it wrong? A few true sentences answer that better than a page of confident ones.
Who asks, and where it lives
The request arrives in a handful of shapes, each with its own wording, and that wording governs what you write. Some customer contracts, especially from larger companies and public bodies, now ask whether any part of the deliverable is produced with these tools and how it is reviewed. Some buyers put the question on the due-diligence list, next to data and insurance, because they are pricing the risk of a quiet mistake. Some lenders and insurers ask on a renewal form, in the same spirit; the pages on what to tell your bank about AI and whether your insurance will ask about AI cover those conversations. Some platforms a company sells or publishes through ask sellers to mark work made with these tools. And sometimes a customer simply asks in an email after noticing a change in tone.
So the disclosure lives in several places at once, and they must agree with each other. The clause in a proposal or a contract, which your lawyer reviews. A short line on the website where a careful buyer will look for it. The answer typed into a questionnaire. The two sentences you send a customer who asks. The content is the same four answers each time; only the length changes. When the website says one thing and the questionnaire another, the reader notices, and the mismatch costs more than either answer would have.
Nothing here says what any law requires of you. Some contracts, questionnaires and platforms ask; many do not. Start from the wording you were given, and when it is a legal term rather than a plain question, that is for your lawyer, not this page.
What a good one says, and never says
A good disclosure uses plain words and names the specific work. "We use software tools to prepare first drafts of routine correspondence and to summarise meeting notes. A named member of staff reviews every item before it is sent. Customer records, pricing and anything you have given us in confidence are not entered into these tools." Three sentences, every one checkable by walking the office.
It names the check by role, not by brand: a customer does not care which tool drafted the letter, only that a person they could name has gone through it. It states the data rule, the sentence a buyer or an insurer is really looking for. And it stops there; a disclosure that runs to a page has started to argue.
It never promises an outcome. "Our tools eliminate errors" is a claim you cannot keep and will be held to. It never offers a tool's brand as a guarantee, because the brand is not the one signing your contract. It never says more than is true to sound modern, and it never says less than is true to sound careful; both are the same mistake, a statement written for effect rather than for accuracy. And it never describes a control the company does not actually run. If the review step exists on paper and not in the office, the disclosure is the document that will prove it.
The data rule
Every one of the questions behind a disclosure comes back to information: whose it is, where it went, and under what terms. So before anything is written, the company needs its one-sentence data rule, the one everyone can repeat. Not the customer list, not pricing, not payroll, not drawings, specifications or recipes, nothing about an employee, nothing under a confidentiality agreement, and nothing a customer would be surprised to find in a tool they have never heard of, into any tool the company has not approved on a company account under terms somebody has read.
Then the disclosure says so, in the customer's terms rather than yours. "Your records and anything you give us in confidence are not entered into these tools" is a sentence a customer can hold you to, which is exactly why it is worth writing. The page on whether a well-known chat tool is safe for business use explains why the terms and the account matter more than the tool, and the page on shadow AI explains why the rule has to be written before the disclosure: the common failure is a careful process on paper while a member of staff pastes customer details into a free personal account twenty feet away.
Five checks before you write it
First, walk the work, not the org chart. For each thing a customer receives, ask the people who make it whether a tool touched it this month. The answers will surprise you, and the disclosure has to match them, not the version in your head.
Second, name the check by role. For each item a tool touches, who reads it before it leaves, and what do they look at? If the answer is "whoever is free", there is no check, and the disclosure cannot say there is. The page on human in the loop is about placing that person deliberately.
Third, write the data rule first and the disclosure second. The sentence about what never goes into a tool is the sentence the reader wants; everything else is context.
Fourth, compare every place the statement will live. Proposal, website, questionnaire, email: same four answers, different lengths, no contradictions.
Fifth, decide who owns the statement. One person updates it when the company changes a tool or a process, and the date on it is real. A disclosure nobody owns is true on the day it is written and false within a quarter.
Check yourself
Seven yes or no questions. Each no is a decision still to make.
- Could you list, from memory, the customer-facing work where tools do part of the job today?
- For each of those, could you name the role that checks the work before it leaves?
- Is there a written one-sentence rule about what never goes into an outside tool?
- Do the tools the company uses run on company accounts under terms somebody has read?
- If a buyer's questionnaire arrived tomorrow, could you answer the AI question without a meeting?
- Would the answer match what the website and your proposals already say?
- Does one named person own the statement and its date?
Mostly no? That is the normal starting point, and it is the company, not the paperwork, that needs the work first. The AI readiness scorecard asks ten questions like these across the whole business and shows which of the three decisions is missing.
Writing it in an afternoon
Three conversations and one page. Ask the people who produce what customers receive what the tools do for them this month, and write down the jobs in their words. Ask who reads each item before it goes out, and write down the role. Find the data rule, or write it if it does not exist (the one-page shape on whether you need an AI policy is enough), and confirm with whoever pays the bills which accounts the tools run on and whether anyone has gone through the terms.
Then write the long form once, four short paragraphs: the work the tools touch, the information they may and may not be given, the person who checks, and what the tools never do alone. From that page, cut the website line (two sentences), the proposal clause (a paragraph, then to your lawyer), the questionnaire answer (the same words, sized to the box) and the customer email (two sentences plus a name to ask). Date it, give it an owner, and put a quarterly reminder in the owner's calendar to walk the work again.
Picture it going wrong
A made-up example, to show the shape. A ninety-person specialty food manufacturer wins a place on a large retailer's supplier list. The retailer's contract asks the supplier to disclose any use of these tools in documents the retailer receives: specifications, certificates, correspondence. The owner, in a hurry, writes that the company does not use them. Two months later the retailer's quality team notices an odd shared phrasing in three allergen statements, asks, and learns that a junior technologist has been drafting them with a free tool on a personal account since the spring, pasting each product's recipe in to do it. Nothing was wrong with the statements. Everything was wrong with the disclosure, and the conversation that follows is about trust.
The same company, a different start. Before signing, the owner walks the quality office and finds the technologist, who is proud of saving an hour a sheet. The owner writes the data rule (no recipe, no supplier pricing, no customer document into any tool not on a company account), moves the drafting onto an approved account, names the quality manager as the person who reads every statement before it is issued, and writes three true sentences into the contract response. The retailer's team nods and moves on; the technologist keeps the hour. The tool is the same in both stories. The disclosure is the difference, and the walk made it possible.
Mistakes that repeat
- Writing the disclosure before walking the work, so it describes an intention rather than the company.
- Answering "we do not use AI" because the owner does not, without asking the people who do the work.
- Promising an outcome, or naming a tool's brand as if it were a guarantee.
- Describing a review step that exists in a policy document and nowhere else.
- Different wording in the proposal, the website and the questionnaire.
- Treating a legal clause as a plain question, or a plain question as a legal clause; the first needs your lawyer, the second needs an honest sentence.
- No owner and no date, so the statement is stale by the second request.
- Saying nothing at all while staff paste customer details into free tools, which is the only version of this that ends badly for the customer.
Frequently asked questions
What is an AI disclosure?
A short, true statement of where a company uses AI tools in work a customer receives, what information the tools touch, who checks the result before it counts, and what the tools never do alone. It is a statement of fact, not a policy and not a promise.
Do I have to tell customers that my company uses AI?
That depends on the contracts, questionnaires and platform rules you have agreed to, and on what any law applying to you says, which is a question for your lawyer. Some ask; many do not yet. The practical answer is to be able to tell them truthfully the moment one asks, which means walking the work and writing the page first.
What does an AI disclosure look like?
Three or four plain sentences. The specific work the tools touch, the information they are never given, the role that reviews the result before it goes out, and what the tools never do on their own. Questionnaires get the long form; the website line and the customer email are two sentences each, in the same words.
Should the disclosure name the tools we use?
Usually not. Customers, buyers and insurers want to know that a person checks the work and that their information is protected, not which product drafted a letter. Naming a brand can read as a guarantee you cannot give, and it goes out of date the day you switch tools. If a questionnaire asks for the list, answer in the box that asks.
What if we use AI in the back office but not in anything customers receive?
Then the honest disclosure says so, and says what information those tools touch, because a buyer or an insurer is asking about the data as much as the deliverable. Payroll, pricing and customer records going into an outside tool matter even when no customer sees the output.
Who should write and own the disclosure?
The person who owns the AI question in the company, with the owner's sign-off and a lawyer's review of any contract language. Not the vendor, not the IT provider, and not the marketing team alone, because the statement has to match the office rather than the brand.
How often should it be updated?
Whenever a tool or a process changes, and at least once a quarter by walking the work again. The date on the page should be a date someone can defend. A disclosure that was true a year ago is a liability today if the company has moved on.
Not sure where you stand? Take the 3-minute AI readiness scorecard.