Automation that someone is answerable for
This policy applies to every program we design, build or operate, and it is written into each statement of work. Reviewed 11 September 2026.
Most of what goes wrong with AI in operations is not a model error. It is that nobody decided in advance what the system was allowed to do, who would notice when it was wrong, and how to undo it. This policy is our answer to those three things.
Six rules every program is built under
A named person owns every automated decision
Each program names the person on your side who is accountable for what the system decides, and the system's permissions are written down before it runs. "The model did it" is not an answer we accept, so we do not build systems where it would be the only one.
People stay where the consequence is high
Decisions that affect a person's money, health, employment or legal standing are proposed by the system and confirmed by a person, unless you decide otherwise in writing and the reversal path is in place.
Models see the minimum
A model receives the data it needs for the task and nothing else. Access is scoped per workflow, logged, and reviewable by you.
Your data is not training material
We do not use client data to train or fine-tune models for anyone else, and we select model providers and contract terms so that they do not either.
Every decision leaves a record
What the system saw, what it decided, and why, is logged in a form your auditors can read. Reversal is a designed path, not a support ticket.
Measured before and after
Accuracy, cost and time are measured against the manual baseline during the parallel run, and the numbers are yours whether or not the program continues.
What we will not build
- Systems whose purpose is to make a decision about a person without any route to a human review.
- Automation that is designed to conceal from customers, staff or regulators that they are dealing with a system.
- Surveillance of employees beyond what a process genuinely requires and what those employees have been told.
- Any program where the client cannot name who is accountable for the outcomes.
How this is enforced
The permissions, the human review points and the logging described above are deliverables of the design stage, reviewed by you before anything is built, and re-checked at the gate into managed operations. The security and data governance page describes the controls that make them hold in production.