Plan for the work that remains, not the work that leaves. People still review held actions, sample its output, handle the harder cases it passes back, and must be ready to take everything back if its depth drops. Build the plan from those four loads, and change headcount only after depth holds through a certificate cycle.
Why is the usual capacity calculation wrong?
The tempting arithmetic is simple: the agent handles a share of the volume, so you need that share fewer people. It fails for four reasons.
First, the work the agent does not take is harder. Routine cases go to the agent; exceptions, complaints and unusual requests stay with people. Average handling time for the remaining work rises.
Second, the agent creates new work. Held actions need approvers. Quality sampling needs reviewers. Briefs and sources need maintaining.
Third, the agent's share is not stable. Depth is earned per agent and can be lost the same day a run fails. On that day, the work returns to people.
Fourth, the skills needed change. Reviewing an agent's draft is a different job from producing one from scratch.
The operations head owns the staffing plan, with finance and HR. They keep the decisions on headcount, redeployment and hiring. Capacity for the agents themselves, meaning compute, connections and throughput, is a related but separate question, covered in how to plan capacity for AI agents.
What is the method?
- Operations analyst: sets the baseline. Volume by work type, handling time, and staff hours, for the period before the agent.
- Agent owner: forecasts the agent's share by work type and depth. Depth 3 work is mostly removed from people. Depth 2 work is not: a person still reviews and sends every item.
- Agent, depth 1: estimates review time from its shadow and early live runs: how long approvers spend per held item and per draft.
- Operations analyst: re-estimates handling time for the work left with people, using a sample of the cases the agent passed back.
- Operations analyst: adds the new loads: held queue approvals, quality sampling, exception handling, brief and source maintenance.
- Operations head: sets the fallback floor. The number of trained people needed to take back each capability's work within a day if the agent's depth drops to 1. This is the minimum the team cannot go below.
- Agent, depth 2: drafts the staffing plan with the figures, assumptions and triggers for review.
- Operations head, HR and finance: decide on redeployment, hiring freezes or reductions, and record the decision with the evidence.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Baseline | Nothing | none | Gathering data |
| Agent share | Nothing | none | Forecasting by depth |
| Review time | Measures from records | 1 | Validating |
| Remaining work | Nothing | none | Sampling passed-back cases |
| New loads | Counts held and sampled items | 1 | Estimating effort |
| Fallback floor | Nothing | none | Setting it |
| Plan | Drafts | 2 | Approving |
| Headcount decision | Nothing | none | All of it |
What is the fallback floor, and how do you size it?
The fallback floor is the smallest team that could absorb the agent's work for a period if the agent stopped. Every agent will lose depth at some point: a certificate lapses, a source changes, a run fails. If the team has shrunk to the point where no one can do the work, the firm is either forced to keep a failing agent running or to leave customers waiting.
Size the floor per capability. For each, ask how long the firm could tolerate the agent being at depth 1, and how many trained people would be needed to clear the volume in that time. The floor is also a skills question: the people must still be doing the work often enough to do it well. Rotate them through it.
The procedure for actually taking work back belongs in the agent's rollback plan, which should state the floor and who is on it.
When should headcount change?
Not at go-live. Wait until the agent has held its depth through at least one full certificate cycle, including renewal, and through a peak period. Until then, the agent's share is a forecast.
Natural turnover is the gentlest lever. Pausing backfill while the agent proves itself lets the team shrink slowly without redundancies, and keeps the option to reverse. Redeployment into review, exceptions and agent analysis is often the better use of experienced people, who are exactly the ones who can tell when a draft is wrong.
Put the staffing assumptions in the AI agent deployment plan from the start, so the decision is not made in a hurry after go-live.
What does the plan record?
- Baseline volumes, times and hours.
- The agent's forecast share by capability and depth, and the actual share each month.
- New loads: approval, sampling, exceptions, maintenance, with estimates and actuals.
- The fallback floor for each capability and the named people on it.
- Triggers for review: a depth change, a certificate lapse, a volume change beyond a set band.
- Each headcount decision, its evidence and who made it.
Where does it usually go wrong?
Savings are booked before depth is earned. Finance takes the forecast as fact. When the agent stays at depth 2 longer than planned, the team is short and the queue grows.
Approvers are not counted. Held work is assumed to take no time. It takes real time, and approvers often sit in other teams.
The best people leave first. Experienced staff, uncertain about their future, find other jobs. The team keeps fewer people who can handle the hard cases. Communicate the plan early and plainly.
No one can do the work any more. After a year, nobody remembers how to process the routine cases by hand. Keep people on the fallback floor in practice, following the discipline in train, shadow, certify, widen.
Questions people ask
Does an agent at depth 2 save any staff time?
Often yes, because reviewing a good draft is faster than writing one. But the saving is smaller than the volume suggests, and it depends on how often drafts need substantive edits.
Should we tell staff about the staffing plan?
Yes. Uncertainty drives people away and makes them reluctant to help the agent succeed. Say what will change, what will not, and how decisions will be made.
How often should the plan be revisited?
Monthly at first, and whenever the agent's depth changes. Depth changes shift the work between agent and people overnight.