Our system can refuse to do something in about eleven different ways. Every one of them comes from a real mistake we’ve watched an AI make. The list wasn’t dreamed up in a meeting; it was built one mistake at a time, from things that actually went wrong early on.
The right message sent to the wrong person. A figure quoted correctly — but from last year’s version of the document. A letter to a regulator written in a tone the organisation would never use. The first time each happened it was a small problem. Each one became a named check the system now runs before it acts.
Why it gets skipped
Listing the ways your impressive new AI will misbehave is uncomfortable — it’s much nicer to show it working. So most projects skip it. But safeguards aren’t built against “bad things” in general. You can’t write a rule for “don’t do anything bad.” You can write one for “don’t send anything to someone outside the approved list for this matter.” That rule only exists because someone named the mistake first.
A safeguard is just a mistake someone took the trouble to describe before it happened.
A few real ones
- Wrong recipient. The message is right; the person isn’t. It checks who you’re sending to before anything leaves.
- Out-of-date source. The figure is real, but from a version that’s since been replaced. It flags the mismatch.
- Wrong tone. Accurate, but in a voice your organisation would never sign. This one always goes to a person.
- Sent twice. A retry that looked like a fresh request. It spots the duplicate.
- Shouldn’t have left the building. An internal note headed outside. It refuses, rather than drafts.
From mistake to safeguard
Each named mistake becomes one of three choices: do it, draft it for a person to approve, or refuse. Routine, low-risk things it can just do. Anything that matters, it drafts and a named person approves. A few things it simply won’t do without a separate, explicit say-so. And a new feature doesn’t ship until we can say which of its possible mistakes fall into which bucket.
The list is never finished
We don’t think eleven is the final number — it’s the eleven we’ve found so far. Every new kind of business turns up mistakes the last one didn’t. So the list keeps growing, and growing it is treated as real work. When something goes wrong, the fix isn’t just a patch — it’s a name, a check, and a decision about what the AI should do next time. That’s the half of the job that never shows up in a demo — and it’s the half that decides whether you can rely on the thing.
Why your data should stay in your country.
The cheapest place to run your AI is rarely the right one.
Read N° 04 →The safeguards are the AI’s safety net.
The design agent on the home page sketches a private AI workbench for your business and emails the plan to you.
Talk to the assistant →