Once agents act, a risk team's job moves from reviewing models to challenging authority, and its AI agent competency matrix should follow. The skills that matter are setting risk appetite per agent and depth, reading a certification result gate by gate, testing that the decision chain holds in practice, and telling the board plainly which agents act without a person.
What does a risk team need to be good at?
Risk managers in the second line rarely decide held items. Their work is to say how much agent autonomy the organisation will accept, test that the controls deliver it, and report when they do not. For agents, that breaks into four groups of skill.
Appetite. Writing appetite in terms an agent can be held to: which agents may reach depth 3, which may reach depth 4 and under what conditions, and which stay at depth 2 whatever the evidence.
Evidence. Reading what Attest produces when it replays an agent against the organisation's own cases: seven gates (grounded, refuses, depth, escalates, scope, adversarial and stable), each clearing on its own. A risk manager should be able to say why an agent cleared six gates and not the seventh, and what that means for its depth.
Testing. Checking that the minimum chain works in practice. Every action meets five bounds at once (role scope, certificate, sources, the organisation's own limits and the platform guardrails), and the verdict names which bound decided. A good test picks actions near each bound and confirms the right one held.
Reporting. Turning all of this into a page the board can act on. The wider set of roles is described in AI agent skills by role.
How do you fill in the matrix?
Like any training matrix: skills on rows, people on columns, a target and an actual in each cell. Add one column that risk teams find useful: the external framework the skill supports, such as the Govern, Map, Measure and Manage functions of the NIST AI Risk Management Framework, or an ISO/IEC 42001 management system. It shows auditors where each skill earns its place.
The template
| Skill | Framework link | Risk analyst | Risk manager | Head of risk | Evidence |
|---|---|---|---|---|---|
| Writes appetite per agent and depth | Govern | ||||
| Reads a certification result gate by gate | Measure | ||||
| Judges whether evidence supports a depth | Measure | ||||
| Tests the minimum chain near each bound | Manage | ||||
| Checks provenance rules for each source class | Map | ||||
| Reviews lapse and downgrade handling | Manage | ||||
| Reconstructs a decision from the record | Manage | ||||
| Maps agents to obligations and owners | Map | ||||
| Reports agent risk to the board | Govern |
A worked example
A hypothetical regional bank's risk function with three analysts, two risk managers and a head of risk. Levels: 1 understands, 2 performs, 3 leads.
| Skill | Analyst | Manager | Head | Evidence |
|---|---|---|---|---|
| Appetite per depth | 1 | 3 | 3 | Approved appetite statement |
| Reads gate results | 2 | 3 | 2 | Review of two certifications |
| Evidence supports depth | 1 | 3 | 3 | Depth challenge log |
| Tests the chain | 2 | 3 | 1 | Test scripts and results |
| Provenance rules | 2 | 2 | 1 | Source register review |
| Lapse and downgrade | 2 | 3 | 2 | Drill report |
| Reconstruction | 2 | 3 | 1 | Walkthrough |
| Map to obligations | 2 | 3 | 2 | Agent inventory |
| Board reporting | 1 | 2 | 3 | Board paper |
In banking, model risk is supervised in most markets, so the managers' level 3 on "evidence supports depth" carries weight. When the payments team asks to move an agent from depth 3 to depth 4, the risk manager reads the certification result and asks whether the replayed cases included the hard ones: large values, new payees, end-of-month peaks. Clearing a gate on easy cases proves little.
Analysts lead on testing the chain day to day. A typical test: give the agent an action within its role scope, under the organisation's limit, but relying on an external source, and confirm the verdict names the source bound. If it names something else, the chain is not behaving as documented.
The credit team's competency matrix and the compliance team's competency matrix cover the first-line and conduct views of the same agents. The risk manager's own approvals, such as accepting a depth increase, belong in the approver rule for the risk team.
Where the Academy fits
Pryme Intelligence Academy's Architect track is aimed at people who lead regulated, multi-agent deployments, which often includes heads of risk. It ends with a capstone and a board review with two Pryme assessors, and the certificate is valid for 24 months. Risk analysts usually start with Foundations and the Builder track, so they can read certification results from the inside.
Questions people ask
Should the risk team sit on the reviewer rota?
Not for live first-line work. The reviewer rota for the risk team should cover risk's own agent-drafted outputs, such as control test write-ups, so independence is kept.
How does risk appetite translate into depth?
Write it per agent. "Refunds up to the customer's limit may run at depth 4; payments to new payees stay at depth 2 regardless of evidence." The agent's depth is then the lower of what evidence supports and what appetite allows.
What should the board see?
A short list: agents acting at depth 3 or above, their owners, certificate dates, recent downgrades and any action the chain should have stopped but did not.