Approved Claims: How to Stop AI Inventing What Your Product Does

AI invents product claims when it does not know the facts, because it fills gaps with whatever sounds plausible. The fix is an approved claims list: a product fact sheet where every statement has a source and a flag saying whether it may be published. AI writes only from facts marked publishable, and anything new goes to a person.

Why does AI invent product claims?

Because language models predict plausible text. When asked to describe your product without the facts in front of it, a model draws on what similar products usually do and writes that confidently. The result reads well and may describe features you do not have, integrations that do not exist or promises you would never make.

The problem is well documented in high-stakes fields. According to Stanford HAI's 2024 study of general-purpose language models, hallucination rates on specific legal queries ranged from 69% to 88%. Business writing is less demanding than legal research. Nevertheless, the mechanism is the same: without grounded facts, the model guesses.

The danger is not obviously wrong output, which gets caught. It is plausible output. A follow-up email promising "seamless integration with your existing CRM" sounds exactly right, and nobody checks until the customer asks how.

What is an approved claims list?

An approved claims list is a product fact sheet built for AI. Each line is one specific statement about the product or company, with the source it comes from, an owner, and a flag saying whether it may be published externally, used internally only, or still needs checking. AI writes only from publishable lines.

The flag is the important part. Many true statements should not be published:

• In testing. A feature that works but has not launched.
• Unconfirmed. A claim someone in sales makes that the product team has not verified.
• Sensitive. A customer name you have no permission to use.
• Volatile. A price or limit that changes and must always be dated.

A simple three-way flag covers most cases: yes, publish; no, internal only; check, confirm before use.

What goes in an approved claims list?

Capabilities, limits, integrations, pricing rules, security and compliance facts, named customers you may mention, and the phrases you use to describe the product. Each entry needs the claim itself, the source, the owner, the publishable flag and, for anything that changes, the date it was checked.

Limits deserve as much space as capabilities. Telling AI what the product does not do is often what stops the worst errors, such as promising an integration on the roadmap.

How do you connect approved claims to AI work?

Put the list where every person's AI and every agent can read it, and instruct them to state product facts only from entries marked publishable. When a draft needs a claim that is not on the list, the AI should flag it for a person rather than write it. Review drafts against the list before anything goes out.

In practice that means three habits. First, the claims list is one of the shared files every AI reads before writing about the product. Second, workflows tell agents to cite which entry a claim came from, which makes review fast. Third, anything new that comes up, such as a capability a salesperson mentions on a call, goes to the owner to confirm rather than straight into copy.

Privacy rules apply to what AI says about people, too. Australia's privacy regulator notes that personal information AI generates or infers, including inaccurate content about an identifiable person, is covered by the Australian Privacy Principles. Keep customer names and details on the approved list only with permission.

Who should approve product claims?

The person accountable for each area: the product lead for capabilities and limits, finance for pricing, legal or compliance for regulatory statements, and marketing for how claims are phrased. Each should own their lines, settle contradictions and approve anything new before it is used externally.

Approval should be a recorded act, not a feeling that something is probably fine. A named approver and a date beside each entry make it clear what is current. That also makes the list trustworthy enough for people to rely on.

People feel the absence of this. According to Atlassian's 2026 Pulse survey, only 12% of workers say their AI understands their organisation like a tenured senior colleague. A senior colleague knows what you can and cannot promise; the claims list is how AI learns it.

The claims list sits inside the product facts set of a company brain.

Skein builds approved claims into the shared knowledge layer of every Company OS, so agents cite their source for each product statement. That suits a business where several people write about the product with AI. A business with a very simple offer may manage with a short list of dos and don'ts.

Frequently asked questions

How do I stop AI making up product features?

Give it an approved claims list: product facts with sources, owners and a flag showing which may be published. Instruct AI to state product facts only from publishable entries and to flag anything it cannot source. Then review drafts against the list before anything is sent.

What is an approved claims list?

It is a product fact sheet built for AI use. Each line is one specific claim with its source, its owner and a flag saying whether it can be published, is internal only or needs checking. AI writes only from publishable lines, so invented or unapproved claims stay out of external content.

Why do AI hallucinations happen in marketing content?

Because the model fills gaps with plausible guesses when it lacks the facts. Asked to describe a product it knows little about, it writes what similar products typically offer. Supplying approved, sourced facts and telling it to flag gaps prevents most invented claims.

Who should approve claims AI uses?

Each area's accountable owner: the product lead for capabilities and limits, finance for pricing, legal or compliance for regulated statements, and marketing for phrasing. Owners settle contradictions and approve new claims, ideally with their name and a date recorded beside each entry.

Should pricing be in an approved claims list?

Yes, if you publish prices, but always with the date it was checked, because prices change. If pricing is not public, mark it internal only so AI never quotes it externally. A pricing rule, such as "from" figures only, can also be recorded as a claim.

Related reading