AI-first company

Data and tools · Explore

Ask AI vendors for proof before you buy, and test on your own material

Do not take an AI vendor’s accuracy figure or “fully compliant” promise on trust. Ask how it was tested and on what, run a small test on your own material, and check the terms on your data and on leaving.

By
Last reviewed
Corrections
Report an error
When Do first. Before you start with AI, or as you start.
Effort Medium (weeks, a few people)
Evidence Emerging

Regulators recommend it, and six regulator cases show vendors’ claims failing or a buyer not asking. No study compares buyers who asked for proof with buyers who did not.

Fits
Industries All industries
In the guide Mid-market companiesTools and systems by size

What it is

Before you sign for an AI tool, get answers to three sets of questions and run one test.

  • Does it work? What was it tested on, by whom, and with what result?
  • What happens to our data? Where it goes, how long it is kept, and whether it is used to train the vendor’s systems.
  • How do we leave? What we can export, how long the contract runs, and what happens to our data afterwards.
  • The test: a small trial on your own material, scored by your own people.

What the evidence shows

Six cases brought by regulators in the United States are the basis for this advice. No study has tested it.

  • A headline figure came from the wrong kind of test. One vendor advertised its AI-text detector as 98% accurate. The regulator said independent testing found 53% on general content, because the system had been tuned on academic writing only.
  • A blanket promise did not hold. Another vendor said its AI plug-in could make any website meet accessibility standards. The regulator said it did not, and that reviews praising it were paid for. The vendor paid $1 million.
  • “As good as a professional” had never been tested. A service sold as a “robot lawyer” had not been compared with a lawyer’s work, and no lawyer checked its output.
  • The buyer was held responsible. Rite Aid bought facial recognition from two vendors. The regulator said it did not ask either vendor how far the system had been tested for accuracy, and did not test it itself. The buyer, not the vendors, was banned from the technology for five years.
  • “AI” that mostly needed a person. The securities regulator found that the “vast majority” of drive-through orders taken through Presto Automation’s own system needed a person to step in, while the company told investors its voice AI removed the need for people to take orders.
  • A small buyer gets little back. The consumer regulator says Air AI sold small firms “conversational AI” it said could replace customer-service staff, with promises that buyers would earn their money back quickly. The regulator alleges some buyers lost up to $250,000. Under a proposed settlement announced in March 2026, the owners are banned from marketing business opportunities. An $18 million judgment is mostly suspended because they cannot pay, and $50,000 goes to the people harmed. These are allegations; the settlement admits nothing.
  • Checking is not a one-off. Australia’s privacy regulator says checking an AI product before and after you buy it “should not amount to a ‘set and forget’ approach”.

A small online shop shows the ordinary version. Its founder compared several vendors before choosing a chatbot, and reports that the vendor helped with set-up and early problems.

How to do it

  1. Write down the task first. One sentence on what the tool must do, and how you will know it does.
  2. Ask for the test behind every number. For any accuracy or time-saving figure: what material was it measured on, how much, and who measured it? If the material is unlike yours, the figure tells you little. For a tool sold as automatic, ask what share of transactions run with no person involved, and how that is measured.
  3. Ask who else has checked. An independent test or audit counts for more than the vendor’s own. Treat online reviews with care. Ask the vendor for two customers of your size you can call. Ask how any usage figure is defined. eBay says more than 10 million sellers have used its AI listing tools; “have used” counts anyone who tried once.
  4. Run your own trial. Take 30 to 50 real examples from your work where you already know the right answer. Run them through the tool. Have someone who knows the work score the results. Do this for two or three tools side by side.
  5. Read the data terms. Find the answers to three questions: is what we enter kept, is it used to train the vendor’s systems, and can it be deleted on request? See decide what data may go into which AI tools.
  6. Check the way out. Can you export your data, prompts and settings in a usable form? Prefer monthly or one-year terms at first. Avoid paying for custom work that only runs on one vendor’s system until the tool has proved itself. Check the terms on price changes at renewal. Software suppliers have raised prices when they added AI to existing plans; see start with AI in software you already pay for.
  7. Ask what support you get. Who helps with set-up, and how fast do they answer when something breaks?
  8. For hiring tools, ask for adverse-impact testing by age and other protected traits. In the United States, a federal court has let a nationwide age-discrimination claim over Workday’s AI screening go ahead against the vendor. The rulings so far are preliminary, with no finding of liability.
  9. Repeat the trial on a schedule. Run the same examples again every six months and after any major update. AI tools change under you.

What to measure

  • The tool’s score on your own test set, against the vendor’s claimed figure.
  • The same score at each repeat. A fall is a reason to call the vendor.
  • Cost per month against the time or errors it saves. See measure against a baseline.

Where it fails

  • Buying on a demo. A demo shows the cases the vendor chose.
  • Buying a promised payback. Be wary of any vendor that sells a payback period or “AI that replaces your staff”. A small firm that is cheated gets little back: in the Air AI settlement, $50,000 for the people harmed against an $18 million judgment.
  • Trusting “compliant”. No tool makes you compliant by being installed. You stay responsible for your website, your decisions and your data.
  • Testing once. The tool you tested in spring may behave differently in autumn.
  • No one who can score the test. If nobody in the company knows the right answers, you cannot tell good output from fluent output.

By company size

  • Micro and small: steps 1, 4 and 5 are enough. An afternoon with 30 examples tells you more than a sales call.
  • Medium: add the exit check and a named person who repeats the trial twice a year.
  • Mid-market and enterprise: a standard set of questions for every AI purchase, with security and legal review, and test sets kept for each tool.

Company cases

Companies that did this

What happened without it

Sources

No published study tests this practice directly, so it is rated Emerging. The rating rests on the company cases above and on the sources below.

  1. Office of the Australian Information Commissioner (October 2024, updated January 2025). Guidance on privacy and the use of commercially available AI products. oaic.gov.au
  2. US Federal Trade Commission (25 September 2024). FTC announces crackdown on deceptive AI claims and schemes. Press release. ftc.gov
  3. US Securities and Exchange Commission. In the Matter of Presto Automation Inc., File No. 3-22413, Release 33-11352 (January 2025). sec.gov
  4. US Federal Trade Commission (24 March 2026). Press release on the proposed settlement with Air AI and its owners. ftc.gov
  5. Value Added Resource. Report on eBay’s AI listing tools for small sellers, with the publication’s own tests. eBay’s usage figure is a company statement. valueaddedresource.net