Fixed rules, thinking colleagues
Fixed code wherever possible, and language models where they fit
- ExplanationExplanationPublication typeA publication that explains one thing simply. It takes a few minutes to read.All terms
- ControlControlFieldHow people and AI colleagues work together safely.All terms
- 5 min read
- Draft, version 0.2
You are here
Step 01, Fixed rules, thinking colleagues
Summary
- The question: what do we use a language modelLanguage modelTermAI that understands and writes language, such as the model behind a chat assistant. It helps you think, but can be wrong.All terms for, and what do we use fixed codeFixed ruleTermOrdinary code that always does exactly the same thing. Predictable and testable, but it never hesitates.All terms for?
- What we do: as much work as possible follows fixed rules. Language models think within those rules.
- What we saw: the consequential errors occur between the components.
- The main limitation: this is one small workshopWorkshopTermWhere our AI colleagues and fixed rules work together, day and night. We share what we learn there in the Lab.All terms. It is not representative of SMEsSMEsTermSmall and medium-sized enterprises: businesses with fewer than 250 people. ITARR builds working AI solutions for them.All terms.
Why we are writing this
When people think of AI, they picture a chat window. In our workshop, many AI colleaguesAI colleagueTermOur name for AI with a defined task and role. It works within our rules; a person authorises irreversible actions.All terms work at once, day and night. Fixed code sits between them. Without that context, our examples are easy to misread.
Rule, model or person?
In our workshop, every task has a place. Try it yourself.
What belongs whereTry it
Who handles this task?
Choose who does each task.
Check whether an IBAN is valid.
View the answer
A fixed rule. An IBAN has a defined format and check digits. A rule can check those consistently.
Summarise a client conversation.
View the answer
A language model. Summarising requires an understanding of ordinary language. That is something a language model does well.
Send the quotation.
View the answer
A person approves. Sending cannot be undone. A person therefore gives explicit approval.
Count how often a task fails.
View the answer
A fixed rule. Counting requires precision. A rule counts consistently and can raise an alert at a threshold.
Judge whether an email sounds friendly.
View the answer
A language model. Tone calls for judgement about language. A model can do that, although it can be wrong.
Pay a new supplier.
View the answer
A person approves. A payment cannot be undone. A rule runs the checks first, but a person approves it.
This is how we divide the work: rules for precision, models for language and judgement, and a person for irreversible actions.
A rule follows a law; a model thinks
Side by side
Fixed rule
- Predictable
- Can be checked with a test
- You can see why it refuses
- Cheap and fast
But It never hesitates.
Language model
- Understands ordinary language
- Writes and summarises
- Makes connections
- Can notice its own error
But It can be wrong.
Ask both the same question ten times. The rule gives the same answer ten times. The model does not always do so.
Asked twiceTry it
The same question, twice
Press the button a few times and see what changes.
Question May this email be sent? It contains an internal name.
- The fixed rule
- No. It contains an internal name. The same every time.
- The language model
- A different answer each time, in this order:
- Yes, the tone is friendly.
- Better not: the second sentence contains a name that looks internal.
- Yes, but check the names once more.
- No, it contains something that looks internal.
A simulated example.
That is why a model never decides on an irreversible actionIrreversible actionTermSending, paying, deleting and publishing cannot simply be undone. Here, a person must explicitly authorise them.Read more: step 08, Stay in control of your AIAll terms on its own in our workshop.
We put rules around the thinking
Think of it as a house. Point to a part.
The houseTry it
Choose one of the five parts of the house.
- The rooms
- This is where language models work. They think, write and assess.
- The walls
- The fixed rules. They define what must always happen and what must never happen.
- The front door
- A fixed scan. Everything leaving the workshop passes through here.
- The window
- A second look. Another model reviews the work, because the same model shares the same blind spots.
- The key
- A person holds it. Sending, paying, deleting and publishing happen only after their approval.
A rule needs a brake too
A rule never hesitates. One night in September, two fixed rules kept each other busy. By morning, the counter had reached 20,478. See whether you can stop the loopLoopTermTwo rules that keep triggering each other, without an end. That is what happened in the night experiment.Read more: step 05, A night in the workshopAll terms yourself.
The brakeTry it
Stop the loop
Press Start, then choose: apply a brake or let it continue.
Rule 1
Rule 2
Two fixed rules keep sending each other notes. The first asks for room. The second follows a fixed rule: “Is there room? Then yes.”
Nobody intervenes. By morning, the count is 20,478 failed tasks (measured).
With a brake, the rule says: “One retry, then a person.” The loop then stops.
A simulation of that night. Only the final figure was measured.
Since then, the rule has been: one retryRetryTermOne new attempt after a failure. If it also fails, the system stops and a person takes a look.Read more: step 08, Stay in control of your AIAll terms, then a person. The system involved remains switched off until that brake is in place.
Keep the scale in mind
Side by side
Your coding assistant
- One model
- One conversation
- You are watching
Our workshop
- Many AI colleagues
- Day and night
- Even when nobody is watching
But The consequential errors arise between the components.
That is why our articles often focus on the space between those components.
Limitations
- Not representative. This describes how one small workshop operates.
- Not a measurement. We did not count how much work follows rules and how much uses models.
- A snapshot. The capabilities of language models change quickly.
What this means
Questions to considerTry it
What does this mean for you?
Turn a card over for our tip.
Which task could a fixed rule handle better than your AI?
Tip: Start with anything that requires precision: counting, checking and validating a format.
Where does a model make a decision that a rule should check?
Tip: Look at everything that leaves your organisation. Put a fixed scan in front of it.
Who stops a rule that gets stuck in a loop?
Tip: Allow each automatic repetition one retry. After that, a person takes a look.
Closing notes
With thanks to the people and AI colleagues at ITARR.
Also read: A night in the workshop and the working policy Stay in control of your AI.
How to cite this: ITARR Lab (2026). Fixed rules, thinking colleagues. Version 0.2, as of 17 October 2026. itarr.eu/lab/vaste-regels-denkende-collegas. CC BY 4.0.
Sources: ITARR’s own working rules, as of 17 October 2026; a night in September 2026, shown in the film at itarr.nl/etalage/nacht/ (as of 17 October 2026).
Licence: CC BY 4.0. Use, adapt and share it. Attribute ITARR Lab, link to the licence and indicate any changes.