AI agent skills by role: what each person needs to know

Each role needs a different slice of skill once agents do part of the work. Reviewers judge held actions and refuse with written reasons. Team leads keep those judgements consistent. Builders brief, bound and certify agents. Owners answer for each agent's depth. Risk teams test the controls, and the board asks whether all of it holds.

Why should skills be set by role rather than by team?

Because the risk sits with individuals. One reviewer approving a depth 4 payment carries more consequence than a whole team reading dashboards. An AI agent team is a small team with a big blast radius: a handful of people decide what thousands of agent actions may become, so a gap in one person's skill shows up at scale.

Setting skills by team also hides the people who matter most. A finance team may have twenty people, but only three approve held payments and one owns the agent. Training all twenty to the same level wastes time on the many and under-trains the few. Start from the decisions each person makes, then work back to what they must know to make them well.

What does a reviewer need to know?

Reviewers sit at the point where the agent stops and a person decides. Their work is the held queue: actions the agent drafted that need a named person before they go anywhere.

A competent reviewer can do five things. They read the verdict and see which bound held the action, whether the certificate, a source, the customer's own limit, the role scope or a platform guardrail. They check provenance, so an amount backed by a sealed ledger entry is weighed differently from one copied out of an emailed invoice. They know what their choices mean, including what edit and approve means: changing the agent's draft and then approving it, which makes the edited version theirs. They refuse when the case is wrong and write a reason a colleague could follow. And they know the red lines that are never approved, whatever the queue looks like.

What stays with the reviewer is judgement. The agent can assemble the case; it cannot decide that this supplier's change of bank details smells wrong.

What does a team lead or approver need?

Team leads run the rhythm around the queue. They do not need to review faster than anyone else; they need to see the pattern across reviewers.

That means four skills. Planning cover so the queue never waits on one person, which the guide to building a reviewer rota for an accounts payable team works through shift by shift. Running calibration, where reviewers decide the same cases separately and compare. Watching for drift, such as one reviewer whose approval rate climbs while their written reasons shrink. And acting on lapse: when an agent's certificate lapses after 90 days or a run that clears no depth drops it to depth 1 the same day, the queue fills with work that was routine yesterday. The lead must expect that and staff for it, not treat it as an outage.

What does a builder need to know?

Builders design, train and certify agents. They do not need to be engineers, a point made in whether you need technical staff to run AI agents: the best builders are often the people who wrote the process.

A builder must be able to write the job in plain language so Agent Studio can assemble the brief, choose sources and give each the right provenance class, and set bounds per capability. They must understand qualification, not training: the business never changes model weights; it changes the brief, the sources, the bounds and the certification. And they must read an Attest result. An agent clears seven gates (grounded, refuses, depth, escalates, scope, adversarial and stable), each on its own, against real cases from the business's own operation. A builder who can read a failed gate knows whether to fix a source, narrow a bound or accept a lower depth.

What does an agent owner need to know?

Every agent needs one named owner, usually the person who owns the process it works in. The owner decides what depth to seek, accepts the risk at that depth, and answers when something goes wrong.

The owner's core skill is knowing what depth means for their process. Depth is earned, not configured: the role sets a ceiling, evidence from Attest sets the depth, and the lower wins. An owner who wants depth 3 for invoice matching must understand what evidence earns it and what changes for their team when it arrives. Owners also change. When one leaves, the agent ownership handover template carries the open risks, the certificate dates, the bounds and the reviewers' known weak spots to the next person, so the agent is never ownerless.

What do risk, compliance and the board need?

Risk and compliance owners need to map each agent to the obligations it touches and test that the controls work in practice. They should be able to pick any decision from the record and reconstruct it: what the agent proposed, which bound decided, who approved it and on what sources. If they cannot, the control exists on paper only.

The board needs less detail and sharper questions. Which agents act at depth 3 or above? Who owns each? What happens when a certificate lapses? Are the people approving held actions certified, and is that current? A director who knows what depth, held and withheld mean can tell a confident answer from an evasive one.

In what order should these skills be built?

From the shared base outward. Everyone who touches an agent learns the same vocabulary first: depth, the decision chain, the four verdicts, provenance and lapse. Then reviewers and team leads, because they meet the agent first. Then builders and owners, then risk. The board can learn alongside the owners.

This follows the deployment order in deploying an AI agent: train, shadow, certify, widen. Reviewers practise while the agent shadows the work, so by the time its depth widens, the people handling held cases have already seen hundreds of them. A business can also adopt the practice of linking approval rights to a current certificate, depth by depth, so nobody approves at a depth they have not shown they can handle.

Where the Academy fits

Pryme Intelligence Academy maps its four tracks to these roles. Foundations, about four hours, serves owners, executives and board members. Operator is for reviewers, team leads and operations managers, and certifies through an exam and a two-week practicum of real review decisions. Builder certifies whoever designs and trains agents, with a capstone agent certified in Attest to depth 2 or higher. Architect is for people leading regulated, multi-agent deployments.