Document business rules for AI before automating

To document business rules for AI before automating them, sit with whoever decides today, go through real cases and write each rule with its condition, its action, its reason and its source. Then test it on new cases and have the owner of the process approve it before an AI agent uses it.

This article is part of AI-First hybridisation consultancy.

In short

A company's rules are almost never written down: they sit in people, emails and exceptions.
You draw them out by looking at real cases with whoever does the work, not by asking for principles.
Each rule has four parts: condition, action, reason and source.
A rule is tested on new cases before it is approved.
You do not need to document the whole company: the rules of the first process are enough.

What a decision rule is in a company

A decision rule is a sentence that says what the company does in a specific situation, and why. «If an order goes over the client's credit limit, it does not ship until the accounts team approves it» is a rule. «We look after our clients» is a wish.

An AI agent can only work with rules of the first kind. That is why documenting them is the step that separates a demo from a process that works, and it is where company hybridisation with AI begins.

Where the rules nobody wrote down are

Before you write, you need to know where to look. A company's rules tend to sit in five places:

In the person who has done the job for years and can no longer explain why they do it that way.
In corrections: every time someone says «we don't send it like that», there is a rule behind it.
In exceptions with a name attached: «we always ring this client first».
In old emails where someone decided something once and it stuck.
In what a new joiner gets told during their first weeks.

Manuals and procedures are a starting point, but they describe the ideal process. The rules an agent needs most are the ones that say what happens when a case does not fit the manual.

How to draw them out: cases before principles

Ask someone how they decide and you get a general principle. Show them a case and ask what they did, and you get the rule. The method has seven steps:

Choose one process and the person who does it best.
Sit beside them and go through ten real cases together, including the ones that went wrong.
For each case, ask what they decided and why. Write it down in their words.
Look for what repeats. If they did the same thing for the same reason in three cases, there is a rule there.
Write it with its four parts and read it back to them aloud.
Test it on five cases you have not looked at. If the rule would not have told you what to do, it is missing a condition or an exception.
Ask the owner of the process to review it and approve it.

The second step turns up what nobody would have mentioned in a meeting: that three clients always send their invoices as crooked photos, that one supplier invoices twice, or that the fastest person on the team uses a shortcut that is written down nowhere.

How to write a rule

A well-written rule answers four questions. This example is from an accountancy practice, made up for the article:

The four parts of a decision rule
Part	The question it answers	Example
Condition	When it applies	If an invoice arrives without the issuer's tax number
Action	What is done	It is not sorted: it is set aside and requested from the client with the usual message
Reason	Why it is done that way	Without a tax number it cannot be recorded, and finding out at quarter end means redoing the work
Source	Where it comes from	The head of administration, from the cases returned last quarter

With all four parts, the rule can be checked: find a real case and see whether the condition was met and the action was taken. The same method, written for a professional setting down their own rules, is on the Digital Twin blog: how to write down your decision rules in plain language.

Rules that work and rules that do not

When you review the first rules, these five signs show that one is not ready yet:

It says «it depends» without saying on what. Every «it depends» is a condition that has not been written.
It describes an intention and not an action: «prioritise good clients».
It mixes two decisions. If the action has an «and» in it, it is nearly always two rules.
Two people on the team read it and understand different things.
It has no source, and nobody can say who decided it or when.

The fourth sign often uncovers something more serious: two people who have been deciding differently for years without knowing it. Before the rule is written, the company has to agree. That agreement is one of the most useful results of the exercise, with or without agents.

What does not become a rule for an agent

Some decisions are documented precisely so that it is clear they are not automated:

Decisions about people: selecting, promoting, moving someone to another role or ending a contract. The EU AI Act treats these systems as high risk (Annex III, point 4).
Decisions that would have legal effects on someone if they were taken without human involvement (Article 22 GDPR).
Decisions whose rules change every few weeks. The agent would be working from a rule that has lapsed.
Decisions nobody can explain yet. They have to be understood first.

In these cases, the rule you write is a different one: «a person decides this, and the agent prepares this information for them».

Where they are kept and how they are maintained

Rules written in a loose document age just like the ones that were never written. They need a place with versions, someone responsible and a way to change them. At Legathum that place is the Book of Criteria, which we explain in what a Book of Criteria is and why your AI agents need one.

Two habits keep them current. First, every correction made by whoever oversees an agent is a candidate for a new rule. Second, the whole Book is reviewed on a fixed date, even if nothing seems to have changed.

Having the rules in order also helps when informing staff. If you employ people in Spain, the works council has the right to know the parameters, rules and instructions behind the algorithms or AI systems that affect working conditions (Workers' Statute, Article 64.4.d). Other jurisdictions have their own rules.

How much to document before you start

Only the rules of the first process. Documenting the whole company before automating anything is the surest way never to begin.

In the Legathum AI-First Hybridisation Method, this work has a name. The Decision Loop Map sets out the full process behind each important decision: where it gets stuck, where the judgement lives, what an agent runs and what stays with people. After that, each agent has its specification, with its judgement, its limits and its human oversight, and nothing is built without it.

This is the work we do in the AI-First hybridisation consultancy, and it is the basis of an AI-first company. To choose which process to start with, use the table in which processes to automate with AI first.

Frequently asked questions

What does documenting business rules mean?

Writing down how the company decides in each situation of a process: when each rule applies, what is done, why, and who decided it.

Why do it before automating?

Because an AI agent works with what is written. If the rules are still in someone's head, the agent will answer with whatever seems plausible.

Who has to write them?

Whoever does the work today explains them, someone writes them down in that person's words and the owner of the process approves them.

How many cases do you need to go through?

Ten real cases per process are usually enough to find the first rules, and five more to test them.

What if two people decide differently?

Agree before you write the rule. An agent cannot apply two sets of judgement at once, and the company should not either.

Where do I start?

With a process that repeats and the person who does it best. If you want to do it with your own company's case, the first step is a 15-minute meeting with Deibson Silva.

How to document your company's decision rules before automating them