Use case · measured
Check every order. For cents.
Every incoming order checked against your rules before it is picked: customer, address, delivery date, item, quantity, unit, price. Described once in plain language, then run as code.
The AI thinks once. Then your task runs as software.
Example: checking every incoming order. Six steps run as code. One uses a model.
-
Describe it
“Check every new order against our ten rules, and flag anything odd in the customer's remark.”
-
Maurice writes the specification
Six steps become plain code. One step, judging the free-text remark, uses a model.
Readable, versioned, stored with your code.
-
A person approves
Tested on 48 real orders.
- Approve
- Correct
-
It runs, the same way every time
Order #10482: 9 rules pass. Remark flagged for review.
In our benchmark: 1,000 order checks
- 18× cheaper than the cheapest AI agent tested €0.07 vs €1.27, and up to 232× (€16.22)
- 100% success on exact rules no errors over 1,250 runs, vs 3 and 15 errors for the AI agents
- 50 / 50 identical results on a repeated run vs 48 / 50 and 49 / 50 for the AI agents
Synthetic scenario, against Claude Haiku 5.5 and Sonnet 5.5 agents. Execution cost only, excluding infrastructure. Read the methodology and its limits
The full measurement table
| For 1,000 runs | AI agent (Claude Sonnet 5.5) | AI agent (Claude Haiku 5.5) | Maurice |
|---|---|---|---|
| Execution cost | €16.22 | €1.27 | €0.07 |
| Duration of one run | 7.1 s | 6.8 s | 1.1 s |
| Errors on the 9 exact rules | 3 | 15 | 0 |
Do it on your own orders.
The same path works for your rules, your systems and your volumes.
-
Write down your rules
In your own words, the way you would explain them to a new colleague. Ten rules took about 9 minutes in our benchmark.
-
Test on real orders, then approve
Maurice turns the rules into code and keeps a model only for judgement, here the free-text remark. A person approves the specification.
-
Run it, and measure it
Every run is logged with its cost. In a pilot, we measure cost per run, time saved and accuracy on your data.