Break the team's work into tasks and test each one. A task stays with people if it needs judgement nobody can write down, if a named person must answer for it, if a mistake would seriously harm someone, or if it means delivering hard news. The rest can move to the agent at a depth it earns.
Why decide this deliberately, and who owns the decision?
Without a deliberate decision, work drifts. The agent takes whatever it can technically do, people keep whatever is left, and nobody notices that the team now approves things it no longer understands. Deciding up front gives each person a clear job, gives the agent a clear scope, and makes the hand-overs between them explicit.
The team's manager owns the decision, with the process owner if they differ. Team members contribute, because they know which tasks hide judgement. HR joins where roles change. Risk reviews tasks with regulatory or customer-harm consequences.
What are the steps?
- Person, team manager, with the team: lists the tasks in each role, at the level of "check the payee name against the invoice", not "process payments".
- Agent, depth 1: adds volume and time to each task from system data and past cases, so effort is visible.
- Person, team manager: applies the four tests to each task: judgement, accountability, harm, hard conversations.
- Person, risk lead: reviews tasks with regulatory or harm consequences.
- Person, team manager: places each task in one of three groups: the agent does it within bounds; the agent drafts and a person decides; a person does it.
- Person, team manager: chooses skill-retention work: tasks people keep doing, or sample, so they stay able to check the agent.
- Person, team manager: designs the hand-overs between agent and people for each task in the middle group.
- Person, team manager: writes up the result as a task map, shares it with the team, and reviews it after the first month.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| List tasks | Nothing | none | Describing the work |
| Add volume and time | Measures each task from data | 1 | Nothing |
| Apply tests | Nothing | none | Judging each task |
| Risk review | Nothing | none | Regulatory and harm view |
| Place tasks | Nothing | none | Assigning to the three groups |
| Skill retention | Nothing | none | Choosing what people keep |
| Hand-overs | Nothing | none | Designing the flow |
| Write up | Nothing | none | Sharing and reviewing |
Which tasks usually stay with people?
Judgement that resists rules. Deciding whether a customer's story about a missed payment suggests financial difficulty. Experienced staff often cannot say how they know; that is a sign the task should stay.
Named accountability. Signing off a regulatory return, approving a credit decision above a limit, or confirming a control worked.
Serious, hard-to-undo harm. Releasing large payments, closing accounts, declining claims.
Hard conversations. Bereavement, hardship, complaints about the firm's own mistakes. An agent can arrange the call; a person makes it.
Exceptions. Cases the rules do not cover. The agent recognises and routes them; people resolve them.
Which controls keep the split working?
A written task map. Everyone can see who does what. Evidence: the map, versioned.
Skill retention. People keep doing a share of the work the agent does, or review samples in depth, so they can still judge the agent's drafts. Evidence: rotation schedules and sample reviews.
Explicit hand-overs. Each hand-over has a trigger, a destination and what travels with the case. How do you hand work between AI agents and people? covers the patterns. Evidence: routing records.
Changes go through the manager. Moving a task from people to the agent is a decision, not a drift. Evidence: task map changes with dates and reasons.
Which signs show it is going wrong?
- Approvers cannot explain their approvals. They no longer do the work, so they cannot judge it.
- Approval fatigue. Some people spend all day approving held items and start to click through.
- Customers cannot reach a person when they need one.
- Tasks move to the agent informally, because someone found it could do them.
- The team feels deskilled or sidelined. Morale falls, and the people you need for exceptions leave.
What should you measure?
The share of each person's time on judgement work against approvals, held items per approver per day, exceptions resolved and how long they take, sample review findings, and staff feedback on their roles. Decide in advance what freed time will be used for, as argued in decide what to do with the time agents free up before they free it.
Where does it usually go wrong?
The split is decided by the vendor. Vendors describe what their agent can do, which is not the same as what should stay with your people.
Roles become approval-only. A job made entirely of approving an agent's drafts is tiring and erodes skill. Mix approval with judgement work.
The decision is not revisited. As the agent earns depth, tasks move. Review the map at each re-certification.
The next team copies the map. A second team's work differs. Repeat the exercise when you extend an AI agent to a second team.
Where Pryme fits
In Pryme, agents and people work in the same Workspace, and each capability has its own depth, so the task map maps directly to what the agent may do, what it drafts and what it holds. Managing the combined team is covered in how do you manage a team of AI agents and people?, and the rollout path in deploying an AI agent: train, shadow, certify, widen.
Questions people ask
Should staff decide which of their tasks the agent takes?
They should shape it, because they know where judgement hides. The manager decides, and explains decisions that go against the team's view.
What if the agent could do a task well but we keep it with people?
That is a legitimate choice, for skill retention, customer preference or accountability. Record the reason, so the decision is revisited deliberately rather than overturned quietly.
How often should the split be reviewed?
After the first month, then at each re-certification or whenever the agent gains a capability. Tasks should move in both directions when evidence supports it.