Instrument
What an agent is allowed to decide
The question comes up in every first conversation, and rightly so: what does the system do on its own? The answer does not belong to trust but to the contract - as a scope of action defined step by step.
Last updated: July 2026
01
Four levels
- Level 1 - preparing
- The system reads, matches, proposes. Nothing leaves the building, nothing is written. A person sees every case. Suitable for anything new, and the level we usually start on.
- Level 2 - independent within the reversible
- Sorting, matching, creating drafts, proposing appointments. All of it can be undone. A person checks by sample rather than case by case.
- Level 3 - independent with an exception rule
- The unambiguous part runs through, anything with a discrepancy goes to a person - together with the reason why. This is the level most processes settle on for good.
- Level 4 - fully independent
- Only where there is nothing to decide, only to transfer. A system integration reaches this level; producing quotations does not.
02
The rule behind it
The level does not depend on how good the model is but on what an error costs and whether it can be undone. A misfiled message can be found again. A rejection already sent cannot, and neither can a wrong price in a binding quotation.
Irreversible or binding means a person releases it. We do not negotiate that line downwards.
That is not caution on principle but what protects the business: the level is in the specification, it is verifiable, and it can be raised once the process has proved itself over months. The other way round - let it run free first, then row back - costs trust that a business does not get twice.
03
What you keep in every case
- Traceability
- Every case carries what it rests on: which document, which record, which revision. No source, no output.
- The permissions you have today
- A system adopts your permissions, it does not work around them. What someone may not open today, they will not see through the new route either.
- The off switch
- Every process can be halted without the business stopping - the routine then falls back to the way it worked before.
- Uncertainty left as uncertainty
- What the system does not know for certain, it flags. It does not fill a gap with a plausible guess - that is the most dangerous property of these models and the point at which construction decides reliability.
Frequently asked
Does the AI decide on prices or rejections by itself?
No. Anything binding and anything irreversible goes past a person. That sits in the specification as a scope of action, not as an assurance in conversation.
What happens when the system is not sure?
It flags the passage and hands the case over rather than guessing. That property is tested at acceptance: unclear items have to be recognisable as unclear.
Can we change the degree of autonomy later?
Yes, and that is the usual route. Processes start at a low level and are raised when practice supports it. The change is recorded like the original definition.
Does the system see data that individual employees may not see?
No. Permissions stay as they are - the new route gets no more rights than the old one.
Tell us where your time goes.
In the audit we record the process, work out what is worth automating, and you get a plan with fixed prices - credited against the build if you go ahead.