Who does what when AI joins the work?
Decide which tasks AI can perform, where people check, and who can stop an action.
How work changes when AI joins it
Lesson 1 of 3 · Four ways AI can take part in a step
Adding AI to an old process does not make the process better by itself. If nobody decides who does what, the AI does whatever it was set up to do, and people assume someone else is checking.
For every step in a piece of work, there are four possible answers:
- A person does it. The AI is not used for this step.
- AI suggests, a person decides. The AI drafts, compares or ranks. A person chooses.
- AI acts after a person approves. The AI prepares the action. Nothing happens until a named person says yes.
- AI acts alone. The AI does the step without waiting. A person reviews afterwards, or not at all.
The question for every step: which of these four is it, and who decided that?
Lesson 2 of 3 · Where a person checks, and who signs off
For each step, name two people: the person who checks the AI's work, and the person who signs off before anything leaves the team. They can be the same person. They cannot be "the team".
The check should match the consequence. An AI that drafts a reply needs fewer checks than an AI that pays a supplier.
| Step | What the AI may do | What a person must check | Who signs off |
|---|---|---|---|
| Compare supplier quotes | Collect the quotes and build a comparison | That every quote covers the same work | Category manager |
| Recommend a supplier | Rank suppliers on the agreed criteria | The evidence behind the ranking and the contract conditions | Budget holder |
| Flag unusual invoices | Mark invoices that differ from the contract | Each flagged invoice, before any payment is stopped | Finance owner |
If you cannot write a name in the "Who signs off" column, the step is not ready for AI.
Lesson 3 of 3 · When the AI is an agent that acts
Some AI tools now act, not only advise. They send messages, update records or place orders. These are often called AI agents. A step where an agent acts needs four more answers:
- Which systems and tools the agent may use.
- Which actions need a person's approval first, for example any payment above a set amount.
- What the agent records, so that a person can see what it did and why.
- Who can stop it, and how quickly.
Then test the process with a wrong answer. Give the AI an incomplete or wrong input on purpose, and see who notices, how quickly, and what it costs.
A process is ready for AI when a wrong AI answer is caught by a named person before it costs money.
Map one workflow
Choose one piece of work your team does every week where AI is used or planned. Write up to four steps. Leave a box empty if you do not know the answer: the empty boxes are the result.
When one person cannot fill the blanks
If your map has open points, you have found the next problem. You can see where AI should fit, but the people who own the open points must agree: often Finance, Operations and IT. Someone must decide who changes the process, what happens when something goes wrong, and how to test the new workflow.
Process readiness briefing
A working session with the people responsible for one workflow. Together you agree the roles, checks, approvals, what happens when the AI is wrong, and how you will measure the result. You leave with a reviewed workflow map and an action list. Building or connecting the AI tools is separate work.
The next free programme
You have decided who checks and approves the work. But what happens when the AI uses the wrong records?
AIR Academy free programmes
AIR Academy is the training arm of AIR APAC, the Center for AI Readiness in Asia Pacific, a social enterprise company based in Singapore. This programme is free, needs no account and stays free. Your answers in the exercise stay in your browser: nothing is sent to AIR APAC.
The three free programmes: People: From Answers to Decisions · Process: Who does what when AI joins the work? · Data: Get your data ready for AI