Back to blog

What you need before building an AI agent: a checklist

Before an AI agent is built you need four things: real examples of the work it will do, access to the systems it has to read and write, one named person who owns the process, and a measure that shows whether it worked. Without all four, the build either stalls while someone chases the missing piece or ships an agent nobody trusts enough to use.

This checklist is what we ask for before an agent build starts. None of it needs technical knowledge. Most of it is a few days of work for someone who knows the process well.

1. Real examples of the work

Collect a set of recent real cases with the right outcome for each: the inquiry and the reply that was sent, the supplier offer and the decision that was made, the invoice and where it was filed. Use the actual inputs, not a cleaned-up summary. If the work arrives as scanned PDFs and messy emails, the examples should be scanned PDFs and messy emails.

These cases do two jobs. They are how the decision rules get written, and they become the test the agent has to pass before it goes live.

2. Access to the systems

List every system the agent needs to read from and write to: the mailbox, the CRM, the ERP, a shared drive, a supplier portal. For each one, find out whether it has an API, who can create a login or a test account, and whether the agent needs to read only or also write.

Access is the most common delay. The project clock starts on the day access arrives, not on the day the contract is signed.

3. One person who owns the process

Name one person whose judgement the agent will copy. It is usually the person who does the work best today, not their manager. During the build they answer "what should happen in this case?" After launch they review the cases the agent hands back.

4. A measure of success

Agree on a number before the build starts: hours spent per week on the task, time to reply to an inquiry, share of cases handled without a person. Write down the acceptance test in the scope, so everyone knows when the agent is working. After launch, add the escalation rate: how often the agent handed a case back to a person.

5. A list of what the agent must not decide

Write down which decisions stay with people: anything above a certain value, anything involving a complaint, any case that doesn't match the usual pattern. This list is written before anything else is built, because it is what makes an agent safe to leave running.

6. Enough volume to be worth it

Check the numbers honestly. If the process involves fewer than about 50 decisions a month, an agent will not pay for itself. If the work takes less than about 5 hours a month, it is not worth automating at all. Our guide to choosing which processes to automate goes through this test in more detail.

7. The data protection check

Note which personal data the agent will see, and whether the use falls into a high-risk category under the EU AI Act, such as hiring or credit decisions. Most business agents do not, but it should be settled before the build. Our EU AI Act guide for small businesses explains the categories.

The checklist in short

• A set of recent real cases, with the correct outcome for each
• A list of systems, with API or login access arranged
• One named owner who can answer questions during the build
• A number that says whether it worked, and an acceptance test
• A written list of decisions the agent must hand back
• A volume check: at least about 50 decisions a month
• A note of the personal data involved

If you are earlier than this and still choosing a first project, start with our step-by-step guide to starting an AI project.

Frequently asked questions

What do I need before building an AI agent?

Four things: real examples of the work the agent will do, access to the systems it has to read and write, one named person who owns the process, and a measure that shows whether it worked. A written list of what the agent must not decide on its own comes next.

How much data do I need for an AI agent?

Most business agents are not trained on your data. They need a set of real past cases with the correct outcome, so the rules can be written and the agent can be tested against them. The cases have to be the real inputs, such as emails, PDFs or records, not a cleaned-up summary.

Is my process big enough for an AI agent?

If the process involves fewer than about 50 decisions a month, an agent will not pay for itself. If it takes less than about 5 hours of work a month, it is not worth automating at all.

Do I need to deal with GDPR and the EU AI Act before building?

You need to know which personal data the agent will touch and whether the use falls into a high-risk category such as hiring or credit decisions. Most business agents do not, but it should be checked before the build, not after.