Insights · August 2026 · 4 minute read
The most dangerous AI tool is the one that is never unsure.
Every vendor says their tool is accurate. Almost none will tell you what it is not sure about. That gap is the whole problem with buying AI right now.
A confident answer and a correct answer look the same
On a screen they are identical. You learn which one you bought when it is wrong, in front of a client, a regulator, or a court. By then the cost is not the licence fee.
This is why a tool that always sounds certain is a liability, not a strength. Certainty is easy to render. Being right is not.
The flag is the feature
Every tool we build flags its own uncertainty. When a draft contains something the model cannot stand behind, it says so, and it routes that line to a competent person before anything leaves the building.
In regulated work this is not a nice to have. A risk assessment that reads “this control may conflict, review before signing” is safer than a polished one that is quietly wrong. It follows Toyota Principle 8: pull technology to solve a real problem. The real problem is not producing more text. It is knowing which text to trust.
Three questions to ask any AI vendor
- What does it do when it is not sure? If the answer is “it is always sure”, walk away.
- Who signs off before the output is used? The competent person stays in charge. The tool gives them a stronger starting point, not a replacement.
- Can you show me a case where it declined to answer? A tool that never abstains was not built to.
Why we build this way
Half credit for “I think I made an error” beats zero credit for confidently wrong. We would rather ship a tool that looks less impressive in a demo and holds up on a site, in a clinic, or in front of a regulator.
The flag is not the limitation. It is the point.