A skills matrix for an AI agent team lists each person against the skills governed agents demand, such as reading a held action, checking provenance, handling refusals and reconstructing a decision from the record. Each cell carries a level from 0 to 3, and the last columns show the certificate each person holds and when it expires.
What is the matrix for?
It answers three questions a team lead faces every week. Who can decide this held action? Who can cover when that person is away? Where is the team thin? An ordinary skills grid covers the first two for normal work. The agent version adds skills that did not exist before, and it ties levels to certificates that lapse, so the matrix goes out of date by itself if nobody renews.
If you have not built one before, what is a skills matrix? explains the format, and how to build a skills matrix for an AI agent team walks through the process of agreeing skills and assessing people.
Which skills belong in it?
Pick skills tied to what the agent actually does. A useful core set:
- Shared vocabulary: depth 1 to 5, the four verdicts (allowed, held, refused, withheld), lapse.
- Reading a held action: what the agent drafted, why it held it, which bound decided.
- Provenance checks: telling sealed, signed, declared and external sources apart, and knowing an external claim cannot authorise a payment.
- Refusing with a reason: a reason a builder can act on.
- Spotting pressure and injection: instructions hidden in documents or messages, customers pushing for release.
- Calibration and coaching: deciding consistently and helping others do the same.
- Writing and changing the brief: role, sources, bounds.
- Reading certification results: what each gate result means for the work.
- Incident response: suspending, falling back, telling the right people.
- Reconstructing a decision from the record.
Not everyone needs all ten. Reviewers need 1 to 6; builders need 1, 3, 7 and 8 at depth; risk and audit need 1, 3 and 10.
How do you score each cell?
Use four levels and define them with evidence, not self-rating.
- 0: not trained.
- 1: trained; has passed a test on the concept.
- 2: practises it on real work, checked by someone else.
- 3: can teach it and decides the hard cases.
Level 2 and above should rest on something in the record: a run of marked decisions, a certified agent, an audit trace. Sector versions with role-specific skills are the AI agent competency matrix for a claims team and the AI agent competency matrix for a payroll team.
The template
| Person | Role | 1 Vocabulary | 2 Held actions | 3 Provenance | 4 Refusals | 5 Pressure | 6 Calibration | 7 Brief | 8 Certification results | 9 Incident | 10 Record | Certificate held | Expires | Approval rights tied to it |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Add a final row that counts, for each skill, how many people are at level 2 or above. Any skill below two is a single point of failure.
A worked example
A hypothetical supplier payments team of five, with one agent that matches invoices and prepares payment runs.
| Person | Role | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | Certificate held | Expires | Approval rights |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Person A | Team lead | 3 | 3 | 3 | 3 | 2 | 3 | 1 | 2 | 2 | 2 | Certified Pryme Operator | Next April | Depth 4 up to 50,000 |
| Person B | Reviewer | 2 | 2 | 2 | 2 | 2 | 1 | 0 | 1 | 1 | 1 | Certified Pryme Operator | Next September | Depth 4 up to 10,000 |
| Person C | Reviewer | 2 | 2 | 1 | 2 | 1 | 1 | 0 | 0 | 1 | 1 | Foundations badge | No expiry | None until certified |
| Person D | Builder | 2 | 1 | 3 | 2 | 2 | 1 | 3 | 3 | 2 | 2 | Certified Pryme Builder | Next June | Brief changes, not payments |
| Person E | Finance controller | 2 | 2 | 2 | 1 | 1 | 1 | 0 | 1 | 2 | 2 | Certified Pryme Operator | Next February | Second approver above 10,000 |
| Count at 2+ | 5 | 4 | 4 | 4 | 3 | 2 | 1 | 2 | 3 | 3 |
Read the bottom row. Only Person D can write the brief, so if D leaves the agent cannot be changed safely. Person C has passed the shared test but is not certified, so C works in shadow and approves nothing. And because Person E's certificate expires first, E's renewal goes into the plan now. The matrix for a team lead's own priorities is discussed in which AI agent skills matter most for support team leads?, and the full role map in AI agent skills by role.
Where the Academy fits
The certificate columns map to Pryme Intelligence Academy credentials. The Foundations badge is free and does not expire; Certified Pryme Operator and Builder last 12 months and Architect 24. Each can be checked at source through VerifyDoc.ai by number or QR code, and for teams an admin view shows progress, results, certificates and expiry dates in one place.
Questions people ask
How often should the matrix be updated?
Monthly for certificate status, and after every calibration cycle for skill levels. A matrix updated once a year will show people as certified long after their authority has lapsed.
Should people score themselves?
They can propose a level, but a lead confirms it against evidence. Self-assessment tends to overstate the skills people use rarely, such as incident response.
What do we do about single points of failure?
Train a second person to level 2 before the first person's next leave or certificate renewal. Until then, record the gap in the team's risk log.