Fig. 01Four levels

Put human intelligence where it creates the greatest value.

Level 1 · thresholds met. Only when every check meets its threshold. For back-office teams → An invoice is booked automatically only when every match meets its threshold and the amount is small. Everything else goes to a person. Every confidence above its threshold; AND amount under the limit; → done, with the reason logged

Level 2 · the pre-filled answer. The answer is ready. The person confirms, corrects or skips. For back-office teams → The person reviews the answer instead of writing it. Confidence 70–89 → pre-filled, waits for a person; Below 70 → not filled, reason given

Level 3 · explanation at three depths. The AI explains. The person judges. For senior experts → Every figure is labeled measured, estimated or unknown. Click it to see it at three depths: glance, plain meaning, how it's made. See one AI answer, shown two ways →

Level 4 · where AI is not allowed. Some steps never use AI. Exact figures come from fixed formulas, not from a language model. Signed approvals stay with a person. We write down which steps are level 4 before we build anything.

Fig. 02Explore, verify, freeze

How a rule is made.

Explore · AI. AI goes through the past cases and proposes patterns. It is fast, but wrong often enough that nobody should rely on its first answer. Each proposed pattern comes with the cases it was drawn from.

Verify · a person. An expert checks each pattern against real cases. Failed patterns are dropped. The ones that hold are written down, exceptions included. The expert's know-how goes into the rule, not into a prompt.

Freeze · a written rule. The checked pattern becomes a written rule. Anyone can read it, test it and change it. 450 rules were written this way from one accountant's know-how, in about a month.

Fig. 03Research that tries to prove itself wrong

Every claim carries its label. Including ours.

Reading backwards. Workarounds, complaints, what people use instead. The need comes from that evidence.

Red team. Before we recommend anything, we try to prove ourselves wrong. The final report shows what held up and what didn't.

Evidence labels. Every number shows how we know it. Measured, estimated from a stated basis, or not known. Unknowns stay visible. The same labels are used in the systems we build.

Bring one process you can't hand to a machine. We'll show you where each level goes.