How to choose the first process to give an AI agent

Most AI programmes start with a list. Twenty use cases, scored for value and feasibility, presented to the board. Few of them reach production. The ones that do usually began as a single process that one team owned and one person cared about.

That is why our closing line on almost every page says the same thing: start with one process. This guide is how to choose it.

Five tests for a first process

1. 01It costs you review time. Pick the process where skilled people spend hours checking routine work, because that is where an agent returns time without taking judgement away.
2. 02It runs often. A process that runs a few dozen times a month will not tell you much. Hundreds or thousands of runs give you enough cases to see where the agent agrees with your team and where it does not.
3. 03It is written down. If the way the work is done lives only in one person's head, write it down first. An agent cannot follow a rule nobody has stated.
4. 04It changes a system you can copy. The agent should be able to try every change on a copy of the ERP, CRM or queue before anything reaches the live one.
5. 05Someone already signs. There should be a person who approves the risky cases today. They become the approver the agent sends its exceptions to.

Processes that often pass

• Purchase bills arriving by email or WhatsApp, read, matched to the order and drafted in the ERP for a person to post.
• Invoice exceptions: price or quantity mismatches routed to the right person with the evidence attached.
• Customer tickets classified and sent to the right queue.
• Anti-money-laundering alerts triaged, with every outcome left to an analyst.
• Master-data changes such as a new tax code or a revised price list, tested against last month's transactions first.

Processes to leave for later

Some work should not be first, and some should never be automated at all. Payments, changes to a supplier's bank details and amendments to filed returns stay with people. Credit decisions, anti-money-laundering outcomes, clinical intake and hiring are recommend-only in elan1 Decide: the agent can rank and suggest, and a person decides. Start where a wrong answer is cheap to catch and easy to reverse.

Write a one-page brief

Before anyone builds anything, put these on one page and get the process owner to sign it:

• The process, in one sentence, and how often it runs.
• The system it changes, and whether a copy exists.
• The person who approves exceptions, and their deputy.
• What the agent may do alone, what goes to a person, and what is always refused.
• What your auditor will ask, and where the answer will be recorded.

That page becomes the rule set, and it is the most useful thing to bring to a first conversation with any vendor, including us.

Then prove it before you switch it on

Run the agent in shadow on your past cases first. It makes its own call, changes nothing, and you compare. Our guide to running an agent in shadow covers how. Only when it matches the way your team works today should any route be automated, and the cut-over decision stays with the process owner.