Move people from doing the work to deciding on it, one capability at a time, as far as the agent's certificate supports. Keep three people apart for anything that moves money: the one who sets the limit, the one who builds the agent, and the one who approves. Train people before the old roles disappear.
What changes in the team, and who owns the change?
Before agents, most of a team's hours go on routine work: chasing, matching, drafting, updating records. With agents, routine work at depth 1 to 3 shifts to the agent. People move to three kinds of work: reviewing what the agent drafts or holds, handling exceptions it refuses or escalates, and improving it through corrections. A smaller group approves actions at depth 4.
The process owner owns the reorganisation, with HR and the second line involved from the start. Our guide to AI agent skills by role explains what each new role needs. The point of this playbook is order: the team should never be in a state where the agent acts and nobody competent is watching.
What are the steps?
- Process owner: lists the team's work by capability and marks each with the depth an agent could reach on evidence, not the depth anyone hopes for.
- Builder: certifies the agent. Only capabilities within the depth it clears on your own cases move. A run that clears no depth drops the agent to depth 1 the same day, so plan for that.
- Process owner: assigns new roles for each moved capability: reviewers for depth 2, approvers for held depth 3 and 4 actions, and an exceptions desk for refusals.
- Second line: checks separation. For each depth 4 capability, the limit-setter, the builder and the approver are different people.
- Learning lead: trains people before the switch. Reviewers practise on real held actions while the old process still runs.
- Process owner: switches one capability at a time. The old way stays available for that capability until the reviewers have a clean run.
- Process owner: sets a fallback. If the agent drops to depth 1, who does the work by hand, and how fast?
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| List work | Nothing yet | none | Process owner |
| Certify | Replays the team's own cases | per agent | Builder runs; evidence decides |
| Assign roles | Drafts, executes and holds | 2 to 4 | Reviewers, approvers, exceptions desk |
| Separate duties | Nothing | none | Second line |
| Train | Produces real practice material | 2 | Learning lead and trainees |
| Switch | Takes on one capability | certified depth | Process owner decides timing |
| Fallback | Drops to depth 1 if a run clears no depth | 1 | Named people do the work |
How do you keep separation of duties when the team shrinks?
Separation is easiest to lose when roles merge. A credit control team of eight becomes four people overseeing an agent, and the most senior one ends up setting limits, owning the agent and approving write-offs. Every individual decision looks sensible; together they remove the check.
Write the separation down per capability. For a team chasing overdue invoices, that might read: the finance director sets the write-off limit; the agent owner maintains the brief and sources; two named controllers approve held write-offs, neither of whom set the limit. The approver rule for the credit control team and the matching approver authority matrix show what that looks like in full.
Remember where control sits. As the reviewer is the control, not the agent argues, the agent's bounds are necessary but the people deciding on held actions are the control a supervisor will ask about.
Which controls keep you in control, and what evidence do they leave?
Capability-by-capability switch. Evidence: the switch log, with the agent's certificate reference on the day each capability moved.
Separation checked per depth 4 capability. Evidence: the role sheet compared with who approved each held action in the verdict record.
Approval tied to a current certificate. A business can make each approver's rights depend on a current certificate for that depth. Evidence: approver list against expiry dates.
Fallback tested. Evidence: a dated drill in which the team handled one day's volume with the agent at depth 1.
Exceptions desk staffed. Refusals land with named people. Evidence: refusals with an owner and a closing note.
What should you measure?
- Held actions waiting longer than target, by role. Rising waits mean the review layer is too thin.
- Refusals per day and time to close. These are the exceptions people now own.
- Corrections per hundred drafts, which should fall as banked corrections improve the brief and sources.
- People approving outside their certified depth. The target is none.
- Coverage across skills, using a competency matrix for the credit control team or your own team's equivalent.
Where does it usually go wrong?
Headcount moves before competence does. People leave or move on day one, and the remaining staff are reviewers in name only.
The review layer is treated as temporary. It is the job. Reviewers who think the agent will soon need no oversight stop reading carefully.
Exceptions get no owner. Refusals pile up in a shared inbox because everyone assumed the agent would handle them.
Model changes are ignored. Switching model provider is a change to the agent, and needs re-certification; see switching model provider without losing control of an AI agent.
Where the Academy fits
Pryme Intelligence Academy can train the team before the switch rather than after it. The free readiness assessment for teams of up to 25 people shows scores by concept and the gaps, and suggests the first agent to start with. Reviewers and team leads then take the Operator track, which includes a two-week practicum on real review decisions.
Questions people ask
Should we reduce headcount when agents take on routine work?
That is a business decision. From a control point of view, do not reduce the people who review and approve until the agent has a clean record at its certified depth and the fallback has been tested.
What if nobody wants the exceptions work?
Make it a named role with clear decisions and enough time. Exceptions are where the judgement is, and people who handle them well usually make the best approvers later.
How long should the old process run in parallel?
Until reviewers have a clean run on that capability and the fallback drill has passed. For most capabilities that is weeks, not days.