How to teach an AI agent how your business works

Most business procedures leave out the parts that matter. The document shows the standard steps. An experienced employee also knows which supplier needs extra checking, which missing field is harmless and when a manager must step in.

Teaching an AI agent how your business works means capturing both the written procedure and this practical knowledge. In agent software, that package of instructions is often called a skill.

Begin with work that has already happened

Choose one contained process and collect five recent examples from beginning to end. Include an ordinary case, an incomplete one and at least one that took an unexpected turn.

Supplier onboarding is a useful example. The written checklist may ask for a registration record, bank details and a signed contract. Recent cases might show a bank change sent from a personal email address, two legal entities using the same trading name or an expired document that a manager accepted.

Review the cases with the person who does the work. Ask them to show what they opened, checked and changed. Pay attention when they use words such as “normally”, “unless” and “it depends”. Those phrases often point to a missing rule or a decision that needs an owner.

Do not assume every past exception was correct. One may reveal a sensible policy. Another may show a control that people have been skipping.

Define the agent's job in one paragraph

Write a short boundary before writing detailed instructions. For example:

Prepare a supplier-onboarding review from submitted documents and approved company records. Flag missing or conflicting information and produce a review pack. Do not approve the supplier or change bank details.

This says where the agent's work ends. It prepares and checks. A person makes the approval decision, and a controlled system action updates the supplier record.

A useful job boundary names:

• what starts the work
• which sources the agent may use
• what it should produce
• what it may do without approval
• when it must stop and pass the case to someone

If the team cannot agree on these points, pause the AI work and settle the process. An agent cannot repair unclear ownership on its own.

Separate company facts from the steps

Some information changes often. Supplier categories, policy documents and email templates are examples. Keep these as separate reference files so the owner can replace them without rewriting the whole skill.

The main instructions should describe the usual flow of work. For supplier onboarding, that may mean checking the legal entity, comparing bank details and preparing the review pack.

Judgement needs special care. Spell out decisions that depend on context and name the person who handles the unclear cases. An urgent request may move up the queue, for instance, while the bank verification still has to happen.

The Agent Skills specification supports this structure by allowing instructions, reference files, templates and scripts to sit together. The file format matters less than keeping the material understandable for the process owner.

Write down what happens when things go wrong

The normal case is usually easy to describe. Most of the useful work lies in the exceptions.

For each step, cover missing information, conflicting records and failed checks. State when repeated attempts should stop. If the agent receives a request outside its job, tell it where the request should go.

“Send it to a person” is too vague. Name the role and specify what the handover should contain. A finance reviewer may need the two conflicting bank records, their sources and a list of checks already completed. A good handover saves the reviewer from starting again.

This exercise often improves the human process too. If every incomplete form causes the same exchange of emails, create a standard request for the missing information.

Use examples that contain real friction

Give the skill a few cleaned examples from past work. Remove personal or commercially sensitive details unless the system is approved to hold them.

Include an everyday case and one that needs a good handover. Add a case that looks acceptable at first but fails a later check. Show an out-of-scope request that the agent should decline or redirect.

Examples should support the written rules. If every example rejects a certain document but the instructions never explain why, the policy is hidden. Add the rule or link to the source so the next person can understand it.

Test the instructions before live use

Run the skill against past cases and compare its work with the final human decision. Check whether it used the right sources, found the important conflicts and produced a useful handover.

Then try it with a small number of live cases where every output is reviewed. Record failures in plain language. “Used the trading name instead of the legal entity” gives you something to fix. “Bad result” does not.

Failures can lead to a clearer instruction, a new test, cleaner data or a stronger system control. Permissions and mandatory approvals should sit in the agent's operating environment, as explained in our plain-English guide to agent harnesses and skills.

Give the process an owner

Company procedures keep changing after the first version goes live. Someone must own the skill, review errors and update its references when the business changes.

The best result is a written method that an experienced employee recognises and a new colleague can follow. Once that exists, an AI agent has a much better chance of helping without inventing the missing parts.