A gap, not a guess: how to stop AI inventing facts about your company

The proposal said the company serves a certain number of customers. The figure looked right, sat in a sensible sentence and came from nowhere: nobody had given the assistant a customer count, and nothing in the chat said one was missing. The sentence needed a number, so it got one.

Language models write the most plausible next words, and a sentence about how many customers you serve is most plausibly finished with a figure. Unless you have said a blank is acceptable, do not expect one. Both vendors treat that permission as something you have to give. The first technique in Anthropic’s guidance on reducing hallucinations is to let Claude say it does not know [1]. OpenAI’s prompting guide suggests a boundary of exactly this kind: use only the supplied sources, and flag missing information instead of guessing [2].

Company facts are an easy place for a guess to slip through, because they read as the safe part of a draft. Reviewers check the price and the scope and skim the paragraph about who you are. That is where a wrong headcount, an invented client count or last year’s office list reaches a buyer. A blank marked [CHECK] looks unfinished, which is exactly why it is hard to send by accident.

The obvious move is to ask Legal or Marketing for a list of approved facts. That starts a new approval, and new approvals are slow. The quicker question is what they have already signed off. The answers usually exist: the About page, the current sales deck, the boilerplate at the end of the last press release, the annual report if you publish one.

Attach two or three of them in a new chat, in your work account, and type the prompt below. It asks for exact wording because the wording is what was approved. To a lawyer, “more than 200 customers” and “200+ clients” are different claims, even if they are the same fact to everyone else. It asks for the source so every line can be traced back. And it asks for clashes, because the About page and the deck were almost certainly updated at different times.

What comes back is a list of facts, the words used for each and the document each came from, with the conflicting versions marked. Treat it as a draft of the sheet. Take each clash to whoever owns that fact and keep the version they confirm. Then add the two columns the assistant cannot fill in for you: who approved the line, and when. Cut anything that changes weekly. Thirty stable facts with an owner are worth more than three hundred that nobody looks after.

Keep the result to one document. One sheet is easy to review, easy to replace when a fact changes, and easy to point a rule at.

A sheet only helps if the assistant reads it every time and treats it as the whole truth about you. That is the job of a shared project. Claude and ChatGPT both let you keep files and standing instructions in a project, which applies them to every chat inside it. Both let colleagues work in the same project, so the sales team, the bid writers and whoever updates the website draw on one sheet. The steps for each product are further down. The rule itself is two sentences.

In the next draft, the sentence that needed a number reads “We serve [CHECK] customers”. That hole is the point. It tells the writer exactly which fact is missing. The brackets also make every gap findable: search the draft for [CHECK] before it goes anywhere, and each hit is either a question for the owner of that fact or a sentence to cut.

Then read the draft anyway. A project instruction changes how often the assistant guesses. It does not stop the guessing, and a fact can still arrive from somewhere other than the sheet.

Each sentence of the rule does a different job, and both line up with what the vendors themselves advise.

“Use only the sheet for facts about us” narrows the source. Without it, a fact about your company can come from anything the assistant has to hand: what it learned in training, a web search, an old proposal someone attached, a remark in an earlier chat. With it, the sheet is the one place a company fact should come from, and its words are the ones that should reach the draft. That is why the sheet holds the exact wording that was approved rather than a summary of it.

“Anything missing, write [CHECK]” gives the assistant permission to leave a hole and tells it what the hole should look like. Anthropic lists that permission first among its ways to reduce hallucinations [1]. OpenAI’s guidance on checking ChatGPT’s work makes the same argument: a gap stated plainly is easier to deal with than a confident guess. It also tells you to verify the numbers, names, dates and claims in any first draft [3].

The fixed marker matters more than it looks. An instruction such as “be careful with facts” produces careful-sounding prose, and careful-sounding prose is the problem. A token that never appears in real copy can be searched for, counted and, if you want, blocked by a check before anything is sent.

The two-sentence rule covers everyday drafting. These two additions earn their place on bids, tenders and anything a customer will hold you to.

Keep the sheet current, or the rule turns a stale fact into a confident one. When a fact changes, change the line, its owner and its date, and rerun the first prompt on the new About page or deck to catch anything you missed. OpenAI’s checklist for taking a shared skill or plugin company-wide asks for what a fact sheet needs too: a source, an accountable owner, an audience and a review date [16].

Both products have projects, and the set-up is close to identical: the sheet goes in the project’s files, the rule goes in the project’s instructions, and you share the project with the people who write. Use your work account for all of it. The differences are in who can share and who sees what.

The Claude rows are from Anthropic’s help centre pages on projects, sharing and personalisation [4, 5, 6]. The ChatGPT rows are from OpenAI’s Help Center article on projects and its projects documentation [7, 8]. All were read on 11 October 2026, and plans and menu names change, so check the labels in your own account.

Decide on the visibility row before you share. In ChatGPT a proposal drafted inside the shared project is visible to everyone in it, along with its files [7]. That suits a bid team working the same deals and is wrong for anything only some of them should see, so keep those drafts in a project with a smaller membership. In Claude each colleague’s chats stay their own, but anyone with Can view can read the instructions and the knowledge [5], so put nothing in either that the whole audience should not see. A shared ChatGPT project also always uses project-only memory, which keeps its chats from drawing on chats outside it, and ChatGPT Work is not available inside such a project [7].

The rule is guidance the assistant follows most of the time. These are the cases to plan for.

Rolling Claude and ChatGPT out across a company is core 1AYM work. A shared fact sheet is one of the first assets worth setting up in that rollout: which facts go on it, who owns each line, which projects carry it, and how the rule is tested before the wider team relies on it. Where a wrong fact would cost more than an awkward correction, as in a tender or a regulated claim, a rule the model follows most of the time is not enough. The check belongs in the system. That is our AI governance implementation work: controls the system enforces rather than instructions it is asked to follow.

We take small fixed-scope statements of work as well as larger builds, so the sheet, the projects and the checks for one team can be a job of their own. Work can start within a day of the scope being signed. If that would help, book a call below or email enquiries@1aym.com.