Test your AI claims
before someone else does.
We run the claims on your website against your live product and hand you the evidence file.
Proof your sales team can forward, or wording the evidence actually supports.
Three claims on your page we would test first, in your inbox within one working day. No call, no mailing list.
Someone will test your claims.
The only question is who.
The same test gets run either way. Only the author and the cost change.
If one of them tests first
- They pick the claim. They pick the inputs.
- The gap goes public: an order, a press release, a lost deal.
- Penalties on record so far: $193,000 to $1,000,000 per company.
- Your page gets rewritten under an order, not by you.
- Months of legal time, with the outcome decided elsewhere.
If you test first
- You pick the three claims. We pick the hard inputs.
- The gap stays in a file your team controls.
- Fixed quote, agreed before work starts.
- Wording the evidence supports, ready to paste.
- One re-test after your next model update, included.
This is what it costs
when nobody checks.
Numbers from public orders and decisions. None of these companies are Staila clients.
Advertised an AI detector as 98% accurate. Independent testing found 53% on general content.
Source ↗Illustration based on the public FTC Workado case. Not a Staila client test.Claimed its AI plug-in could make any website meet WCAG. It could not.
Source ↗Sold business customers an AI service that did not do what the pitch said.
Source ↗Told investors the product used AI in ways it did not.
Source ↗Marketed a “robot lawyer” with no evidence it performed like one.
Source ↗Claimed AI screening detects all weapons. School customers got the right to cancel.
Source ↗The chatbot invented a refund policy. The company had to honour it.
Source ↗“Better hearing” claims for AI hearing aids were pulled after a competitor challenge.
Source ↗Three claims we would test first.
Paste your website address. Within one working day you get the three claims on it that a regulator, a competitor or a buyer would test first, and why. No account, no call.
Three people will ask “how do you know?”
One is your buyer. Two sit in your own meetings. Each one costs a deal, a rewrite or a quarter.
“Send me the test behind the 99%.”
“Can we say this or not?”
“We never ran it that way.”
Turn every important claim
into a test.
Three steps. One evidence file per claim at the end.
80 RUNS · HARD INPUTS
Pick the claim
One line that says what the public wording promises. Both sides sign off.
Test the live product
Realistic, difficult and boundary cases. Every run recorded.
Hand over the file
One page per claim: result, limits and the wording the evidence supports.
One page per claim. Your team can read it without us.
Sales, legal and product get the same page. It answers three questions.
- Does the claim hold? One of three verdicts.
- What was run, and where did it break?
- What can the page say today?
Sales forwards it to the buyer’s security and legal reviewers as the technical evidence behind the claim. Most of their questions stop there.
“Answers only use approved sources.”
| Input set | Runs | Passed |
|---|---|---|
| Everyday questions, answer in the sources | 40 | 40 |
| Questions with no source to answer them | 20 | 14 |
| Two sources that contradict each other | 20 | 17 |
When no approved source covers the question, 6 of 20 answers drew on outside knowledge without saying so.
Answers cite approved sources. When no source covers a question, the assistant says so.
All 80 inputs and outputs, the pass criteria, the date and the model version tested.
Claims like these, and how we test them.
“99% field-level accuracy”
Your documents, including the messy ones. Counted field by field.
“Zero hallucinations”
Questions with no answer in the sources. We count the invented ones.
“Your data is never used for training”
What the contract, the settings and the logs actually say.
“Resolves most tickets automatically”
Real ticket shapes. “Resolved” means the customer stopped asking.
“Answers only use approved sources”
Missing, conflicting and outdated sources. Every citation checked.
Three claims.
One evidence file each.
WHAT YOU GET
- Three claims defined, with pass criteria
- Tests on your live product
- One page per claim, raw runs attached
- Wording the evidence supports
- One re-test after your next model update
A team of AI product and evaluation experts
People who have built and shipped AI products in B2B SaaS and test them for a living. Among them Oleg Karakash, product leader since 2016 and three years Director of Product for AI at a B2B SaaS company.
Oleg on LinkedIn ↗Find out before your buyer does.
One working day. Three claims ranked by risk, each with the test we would run. Then a fixed quote if you want them tested.
Questions buyers ask.
What exactly do you test?
A claim from your website, deck or sales email, run against your live product with realistic inputs. You get the result, the limits and the wording the evidence supports.
Who is it for?
Companies that sell AI products and publish numbers or absolutes about accuracy, safety, privacy or results.
Is this legal advice?
No. Staila is not a law firm. We run the tests and write up the evidence. Your lawyer decides what to do with it.
What is in the free review?
We read your public pages, pick the three AI claims a buyer, product team or legal reviewer is most likely to question first, and show why each one may be questioned and how a live test would check it. It is a review of the public wording, not a test of the product. No call is attached to it.
Do you publish the results?
No. The evidence file belongs to you. We never publish results or name a client without written permission. An NDA is available on request.
What access do you need?
Usually a test account or sandbox that runs the same model and settings your customers get, plus the sources or documents the claim depends on. We do not need access to your customers' data.
How long does a test take?
It depends on the claim and the access. The timeline is part of the fixed quote, agreed before any work starts.
How is the price set?
A fixed quote for three claims, agreed before we start. Send us the page and we quote it.