Four things, before anyone on the team starts: what depth an agent has and how it is earned, which limits the manager sets, how held actions reach people and who approves them, and how to read the record. Managers set the bounds, staff the review queue and answer for outcomes. If they learn second, the team learns their gaps.
Why should managers go first?
Because the team will copy them. A manager who treats every held action as an annoyance will produce a team that approves without reading. A manager who treats every refusal as a fault will produce a team that works around the agent by hand and leaves no record. The first weeks with an agent set habits that last.
Managers also own decisions no one else can make: how much the team may approve, who covers the queue on a busy day, and what to say upwards when something goes wrong. None of that can be delegated to a team that has not yet been trained.
What must a manager understand about depth?
Depth is how far an agent may act: 1 read, 2 draft for a person to send, 3 execute routine work within bounds, 4 move money or take hard-to-undo actions, 5 coordinate other agents. It is earned, not configured. The agent's role sets a ceiling, evidence from certification sets the depth, and the lower wins.
A manager needs to know this so they stop asking for "more automation" as a setting. They should know which of the team's agents are at which depth, and the depth each task needs. They should know that an agent's certificate lapses after 90 days, and that a lapse, or a run that clears no depth, drops it to depth 1 the same day. That last fact matters on a Tuesday afternoon when the agent stops sending and the team turns to the manager.
Which limits does the manager set?
Every action meets five bounds at once: role scope, certificate, sources, the business's own limits, and the platform guardrails. The lowest wins, and the verdict names which bound decided. The manager usually owns one of those five: the business's own limits for their area.
| The manager decides | The manager does not decide |
|---|---|
| Amounts above which actions are held | The agent's certified depth |
| Which cases always go to a person | Which actions are withheld at the platform floor |
| Who on the team may approve each kind of held action | Whether the agent's sources are sealed or external |
| How the queue is covered across shifts | The platform's own guardrails |
Knowing the split prevents two common mistakes: trying to loosen a platform guardrail, and forgetting that the team's own limits are theirs to tighten.
What should a manager know about held actions and the record?
A held action is drafted and waiting for a named person. The manager needs to know where those appear, how quickly they must be cleared, and what a good approval looks like: the reviewer opens the sources, checks the bound that held it, decides, and records a reason. Every verdict is filed in an append-only record, so any decision can be reconstructed later.
That record is also the manager's best coaching tool. Reading ten of the team's approvals a week shows who is rubber-stamping and who is refusing everything.
An example: an accounts payable lead
Take an accounts payable team lead in a mid-sized distributor. Before the team sees the invoice agent, she learns that it is certified at depth 2, enough to draft payment runs, and that its role lets it only read supplier changes. She sets the team's limit at 5,000, so the agent may only draft any payment above it or to a new supplier, decides that only two senior clerks may make those payments, and learns to read the record.
When the team starts, she can answer every question they ask. When the agent drops to depth 1 after a run that clears no depth, she already knows the fallback.
How should a manager then brief the team?
Short and practical. A lunch-and-learn works well as an opener if it is built around real held actions from the team's own work; see the lunch-and-learn session plan template on AI agents and how to run a lunch-and-learn on AI agents that people remember. If the team is a pilot, read what a pilot team should learn before an AI agent pilot. The full approach is in training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy's Foundations track, about four hours and free, covers why agents must be governed, the decision chain and depth. Managers who run a live workspace day to day go on to the Operator track, whose modules include reading the workspace, when something goes wrong, and running the rhythm.
Questions people ask
Does a manager need to be certified if they never approve?
They need to understand depth, limits and the record. Certification matters if they approve held actions or sign off others' approvals.
How long should a manager spend learning before the team starts?
A day is enough for the essentials, plus time to read the team's first week of verdicts closely.
What if the team learns faster than the manager?
Let them teach. A manager who asks a certified reviewer to walk through a held action is modelling the behaviour the programme wants.