The Claim Audit
Every claim your AI product makes, marked and priced
Send us everything your AI product claims — website copy, pitch deck, docs, demo script. We send back a marked table: what's validated, what's measured, what's computed, what's refuted, and what's simply unproven. £1,500, fixed scope, 5 working days.
The offer
One price, one deadline, one deliverable
What you send
Everything your AI product claims, wherever it lives
If a claim isn't in what you send us, it doesn't get audited — so send the whole surface, not just the parts you're proud of.
Website copy
Every page that makes a claim about what the product does or how well it does it.
The pitch deck
Investor decks and sales decks say things the website doesn't. Send both if they differ.
Docs
README, API reference, technical docs — anywhere a spec or a number is written down.
The demo script
What you actually say and show live. Demos make claims too, and they're rarely written down anywhere else.
The marking system
Five markers. Every claim gets exactly one.
No claim is left unmarked, and no marker is a guess — each one says precisely what kind of evidence stands behind the claim, if any.
- V
Validated
Confirmed by an independent party or toolchain — an auditor's report, a published benchmark, a third-party test. Not your own say-so.
- M
Measured
You ran it and got a number. The row names the exact command or artifact that produced it, so it can be rerun.
- C
Computed
Worked out on paper from other numbers, never measured directly. Arithmetic, not evidence.
- R
Refuted
The evidence you gave us contradicts the claim. These rows stay in the report — they are not quietly dropped.
- U
Unproven
Nothing was produced to back it up. Not necessarily false — just not shown to be true yet.
What comes back — a worked example
A marked table, with a caveat on every row
This is a sample only, built from generic claims to show the shape of the deliverable — not a real client's report. Every row in your real audit gets the same treatment: one claim, one marker, one named caveat. The caveat is never blank.
| Claim | Marker | Caveat |
|---|---|---|
| 95% accuracy on our benchmark | M | Measured on an internal 200-example set — command: `pytest tests/eval/accuracy_test.py`. Not yet run against a public or held-out benchmark. |
| SOC 2 Type II compliant | V | Confirmed against the auditor's published report on the client's trust portal. Scope covers the infrastructure listed there; does not automatically extend to features added since. |
| 40% faster than [a named competitor] | C | Derived from published token-pricing arithmetic, not from a timed run on either system. No latency benchmark exists yet. |
| Works fully offline — no data leaves the device | R | The demo script calls an external embeddings API at step 3. The evidence we were given contradicts the claim as written. |
| Certified for HIPAA compliance | U | No certificate, audit report, or named auditor was provided. Nothing behind this claim was produced for review. |
The other half of the deliverable
What you may say, and what you may not
Alongside the table, a plain two-column list: the version of each claim the evidence actually supports, and the version it doesn't. Also a sample — the real list is built from your own claims.
What you may say
- “Measured 95% accuracy on our internal 200-example eval set (script included).”
- “SOC 2 Type II compliant for the infrastructure covered by our current report.”
- “We estimate this could run faster than [competitor] based on token pricing — not yet benchmarked.”
- “We have not completed a HIPAA compliance assessment.”
What you may not say
- “95% accurate” — unqualified, reads as independently verified. It isn't, yet.
- “Fully SOC 2 compliant” — the report has a scope, and the claim shouldn't outrun it.
- “40% faster than [competitor]” — stated as fact when it's an unverified calculation.
- “HIPAA compliant” — with nothing behind it, this is the kind of line that ends a deal in due diligence.
The differentiator
The refuted rows are the valuable part
Most audits quietly drop the claims that don't hold up — an easy way to keep a report looking tidy. We keep them in, marked R, with the specific evidence that contradicts each one. That's not us being difficult. A prospect, a journalist, or a regulator will find those gaps eventually; better it happens on a page only you have seen first.
What this isn't
Said plainly, so there's no confusion later
Legal advice, a certification, or a compliance sign-off. The Claim Audit doesn't make any of your claims true — it tells you, in writing, which of today's claims you can currently support and which you can't. Nothing here implies regulatory approval or accreditation, because none is being given.
The audit is only as good as the evidence you can produce. If you send us claims with no measurements, no reports, and no artifacts behind them, most rows will come back marked U — and that outcome is itself the finding, not a failure of the audit.
Who's doing the marking
The discipline, not the hype
Fifteen years delivering safety-critical systems in global rail and metro — an industry where an unsupported claim about a braking system isn't a marketing problem, it's a legal one, and evidence for every claim is required by law. Consumer AI has never been made to meet that bar. The Claim Audit applies the same discipline — evidence-backed claims, marked and named — to what your AI product says about itself.
Not ready to send everything yet?
Ask a question first
Tell us what you're building and what you're worried a claim of yours won't survive. A real person reads it and replies within one working day.
This form isn't connected yet.
We haven't deployed the service that would receive it, and a form that swallows your details is worse than no form. Write to us directly instead — it reaches the same person, and you get a reply within one working day.
Fixed price, fixed scope
Send us what your product claims
Email the materials, or book a call first if you'd rather talk it through. Either way, the price is £1,500 and the clock is 5 working days from when we have everything.