What the number does and doesn't tell you
Accuracy is one number across many kinds of decision. It hides which cases the model gets wrong, and those may be the ones that cost you most. A missed defect and a harmless reject count the same in it.
Ask how the figure was measured: on whose data, from which period, who chose the examples and who labelled the right answers. Results from other scanners, cameras or customers may not transfer to yours. A test set the team wrote itself shows only that the product passes its own exam.
How I would test it
Take a sample of your own past cases where the right outcome is already known, such as claims, scans, footage or customer conversations. Run every vendor on the same set, and have your staff score the results without seeing which vendor produced them. If the claim is to beat your current system, tune that system first so the baseline is fair.
Agree the pass criteria before the test starts. Count misses, false alarms and harmful answers separately, by case type and customer segment. Include cases that look different without differing in quality, and questions your sources can't answer.
Depending on the use and your seat
If the AI only sorts work for a person who still decides, a lower accuracy can still be useful. If it acts on its own, the bar is much higher, and you should check what your regulator expects. Settle before the test where your data goes and whether the vendor trains on it.
If you're on the board or an investment committee, ask whether the evaluation changed since it was approved. Easier tasks or vendor-selected examples can make progress look better than it is. Require results on the original basis as well as the new one.
What to check before you decide
- Ask on which data the figure was measured, from which period, who chose the examples and who labelled the answers.
- Test on a sample of your own past cases with known outcomes, with pass criteria agreed before the test.
- Count misses, false alarms and harmful errors separately by case type, and look hardest at the expensive ones.
- Give competing vendors the same set, and have your staff score the results without seeing which vendor answered.
- Ask what the model does when it is unsure, and how a person sees and overrides its decision.
- Check what the contract says if accuracy on your cases drops after go-live.
- Confirm where your data goes during the test, who may use it, and whether the vendor trains on it.
Questions people ask
How should a health board assess competing claims of responsible AI?
Translate each claim into evidence for the intended use. Selection depends on limitations, escalation, affected users, and appropriate clinical, privacy, and legal review rather than policy statements alone.
Can an investment committee rely on a model validated using vendor-selected examples?
Require evaluation relevant to the intended decision and independent of supplier selection. Reliance depends on representative cases, documented exclusions, scoring ownership, and treatment of adverse outcomes.
Should investors trust AI progress reports when evaluation criteria keep changing?
Preserve comparability while allowing justified evaluation changes. Reliance depends on recorded changes, parallel results where feasible, and a clear explanation of how the business claim changed.
Three AI customer service vendors make similar claims, how do I run a fair comparison quickly?
Give all three the same sample of your real customer conversations and judge the answers with your own staff, blind to the vendor. A few weeks of that beats any demo. Then compare the data terms and the cost at your volumes. It depends on your languages and channels, and on how much of your service needs access to customer records.
How should we compare inspection vendors when rejected stock has different costs?
Compare the business consequences of each error type on your products. Selection depends on quality decisions, reinspection capacity, and the cost of unnecessary rejection, with safety questions owned appropriately.
Radiology AI vendor shows good results from abroad, do we need to validate on our own scans before buying?
Yes, because results depend on scanner models, protocols and patient mix, and yours differ from the study's. Run the tool on a set of your own past scans with known outcomes and compare. It depends on what the tool is cleared to do in your jurisdiction and whether the vendor allows a retrospective test.
How do we know our AI agent's accuracy numbers are real before an enterprise pilot?
Accuracy on a test set the team wrote tells you the product passes the team's own exam. Build an evaluation from real customer requests, scored by someone outside the team, and track failures by type. It depends on whether you can get representative requests from the pilot customer before you start.
AI vendor claims it can beat our advanced process control on yield, how do we design a pilot that actually proves it?
The pilot needs a baseline that already includes your control system at its best, a long enough run to cover feed and season changes, and a measure of yield the plant already trusts. Use alternating periods or a comparable unit as control. It depends on how stable your feedstock is and on whether the optimiser only advises or writes setpoints.
Choosing AI video analytics for transport safety, how do I test vendor accuracy claims and handle privacy?
Test on your own footage, your camera angles and your lighting, with incidents you already know about, and count misses and false alerts separately. Decide before the test what the system is allowed to record, retain and identify. It depends on your privacy obligations for passengers and staff and on which incidents matter most.
Is an AI forecast reliable enough to choose new charging locations?
Treat unobserved sites as an extrapolation problem. Useful evidence depends on comparable locations, available local inputs, and whether simpler estimates support the same decision.
Is our manufacturing AI inspection pilot representative enough to justify expansion?
Test whether defect evidence survives changes unrelated to quality. A purchase decision depends on performance across packaging variation rather than a mixed headline score.
Is a hospital knowledge assistant ready when its citations cannot be verified?
Evaluate answer support separately from retrieval relevance. Suitability depends on whether each material statement is supported by the cited passage and whether unsupported questions are handled safely.
Is Hebrew clinical transcription accurate enough to justify a hospital pilot?
Assess meaning preservation rather than transcript fluency alone. Suitability depends on negation, uncertainty, mixed-language speech, and an appropriate review process for the intended use.
How should we compare fraud platforms that block different legitimate customers?
Evaluate prevented loss and legitimate activity disruption separately across relevant segments. Suitability depends on error costs, review capacity, and whether differences reflect evidence or evaluation bias.
How I can help with this decision
- Ask or talk (Free)
- I give my view on what the vendor's evidence proves for your data, and the one test that would settle the claim.
- Review (Pay if it was worth it)
- I write an independent assessment of the vendor's claim, the test results on your own cases and the data terms. I recommend to sign, test further, narrow the use or decline.
- Retain (When it makes sense)
- I stay close through rollout to review results against the agreed test, and flag when the use expands past what was tested.