Codify: turning what your people know into instructions an assistant can follow

When a team first connects an AI assistant to their work, the failure is rarely the model. It is that nobody can state, in writing, how the work is actually done. The rules live in one person's head, in an email thread from 2023, and in a spreadsheet nobody owns. An assistant asked to help will confidently fill those gaps itself — and that is the moment governance is lost.

Codify is the first stage of our method for exactly that reason. Before anything is personalised and before anything is governed, the knowledge has to exist outside a person's memory in a form a machine can follow and a human can audit.

What codifying actually means

Codifying is not writing a policy document. Policy documents describe intent; codified knowledge describes behaviour. The test is simple: could a competent newcomer — or an assistant — follow this and reach the same outcome as your best person?

• The decision, named: what is being decided, by whom, and on what trigger.
• The inputs: which facts, documents or data the decision depends on, and where they live.
• The rules: the conditions that change the answer, including the exceptions people apply without noticing.
• The boundary: what must never be decided by the assistant, and who it escalates to instead.
• The evidence: what should be recorded so the decision can be reviewed later.

How to run a codify pass

1. Pick one recurring, bounded task — not a department. Something done weekly, with a clear start and finish.
2. Watch it done once and write down every question the person asks themselves. Those questions are your rules.
3. Separate fact lookups from judgement calls. Fact lookups can be automated first; judgement calls need oversight design.
4. Write the boundary before the capability. Decide what the assistant must refuse or escalate, and to whom, by name or role.
5. Record the source for each rule — a policy, a regulation, a contract, or a named person's professional judgement. Unsourced rules are the ones that quietly go stale.
6. Review with the person who does the work. If they say 'well, except when…', you have found the exception that would have caused your first incident.

Why codifying pays off even without AI

A codify pass produces artefacts you need anyway: an accurate process description, a data-source inventory, a list of decisions with human owners, and an explicit escalation route. Those are the same artefacts a Data Protection Impact Assessment, an audit, an Ofsted conversation or a due-diligence questionnaire will ask you for. The AI benefit arrives second.

It is also what makes the next stage possible. You cannot personalise knowledge that does not exist in a shared form, and you cannot govern behaviour you never wrote down.