How to build a skills matrix for an AI agent team

Start from the agents, not the people. List every capability each agent holds, the depth it needs and the depth the agent is certified to, then write down which human skill each capability needs: reading, reviewing, approving, teaching or certifying. Map every person against those skills at a stated level, with the certificate and expiry date that proves it.

What is a skills matrix for an agent team, and who owns it?

A skills matrix is a grid of people against skills, with a level in each cell. For a team that runs governed agents, the columns are not generic "AI skills". They are the specific jobs people do around agents: reading a verdict, clearing a held action, correcting the agent's brief, recognising a refusal that should stay a refusal, and checking a certificate.

The matrix needs one owner, usually the operations manager or team lead who runs the review rhythm. Learning and development can maintain the training side, but the person who decides who approves what must own the grid, because the grid is what those approval rights rest on.

It differs from a training matrix, which records which courses people have completed. Completion tells you someone sat a course. A skills matrix tells you whether they can do the job today, which is the question an auditor asks.

The steps to build it

1. Person: list the agents, their certified depth and the depth each action needs. Take each agent in the workspace and write its certified depth and its capabilities with the depth each needs: 1 read, 2 draft, 3 execute routine work within bounds, 4 move money or take hard-to-undo actions, 5 coordinate other agents.
2. Agent, depth 1: pull the current state. A reporting agent reads the workspace and the record and lists every agent, its capabilities, its certified depth, its certificate expiry and who approved its held actions in the last cycle.
3. Person: name the human skill each depth demands. Depth 2 needs a reviewer who can judge a draft. Depth 3 needs someone who can read exceptions and the record. Depth 4 needs a certified approver who knows the customer's limits and red lines.
4. Agent, depth 2: draft the matrix rows. The agent drafts one row per person from certificates, test results and review history. A person checks every row; the agent cannot confirm its own work.
5. Person: set the level in each cell. Use four levels: aware, can review with support, can approve alone, can teach or certify others. Each level needs evidence, not a manager's impression.
6. Person: attach the evidence. Link each cell to a certificate number, a test result or a practicum log. A certificate that can be checked at source is stronger than a PDF in a shared folder.
7. Person: find the gaps and single points of failure. Any capability that needs depth 3 or 4 with only one certified approver is a risk. Any person approving at a depth their cell does not support is a bigger one.
8. Person: set the review date. Agent certificates lapse after 90 days, so review the matrix at least every 90 days and on any day an agent's depth changes.

The columns that matter most are covered in what skills a team needs to run AI agents. If you want a grid to start from, the skills matrix template for an AI agent team has the fields ready to copy.

Which controls keep the matrix honest, and what evidence does each leave?

• Approval rights follow the matrix. A business can adopt the practice of granting approval rights depth by depth only to people whose certificate is current. Evidence: the workspace's list of eligible approvers matches the matrix.
• Certificates checked at source. Each certificate is verified by number or QR code with its issuer, which shows whether it is current, lapsed or revoked. Evidence: a dated check per cell.
• Every approval is in the record. The append-only record shows who cleared each held action. Evidence: a sample of approvals matched to cells.
• A second person signs the matrix. The owner drafts it; a risk or compliance lead reviews it. Evidence: a signed version per cycle.

What should you measure?

Measure coverage, currency and fit. Coverage is the number of certified approvers for each capability that needs depth 3 or 4; two is a minimum, three is comfortable for a team with leave and shifts. Currency is the number of cells whose evidence expires in the next 30 days. Fit is the number of approvals in the record made by someone whose cell did not support that depth, which should be zero.

Also watch how often a held action waits because no eligible approver is on shift. That figure shows a gap in the matrix faster than any survey.

Where does it usually go wrong?

The most common mistake is building the matrix around tools instead of decisions. "Can use the workspace" says nothing about whether someone can tell a correct refusal from a wrong one.

The second is treating levels as permanent. An agent that moves from depth 2 to depth 3 changes the skill needed to oversee it overnight, because the person stops checking every draft and starts reading exceptions. If the matrix is not updated, people approve at a depth they were never assessed for.

The third is leaving out the people who certify agents. Whoever changes an agent's brief, sources or bounds needs a column of their own, because their work decides what everyone else will review.

Where the Academy fits

Pryme Intelligence Academy gives the matrix its evidence. The Foundations badge covers aware level for anyone. Certified Pryme Operator covers review and approval, Certified Pryme Builder covers designing and certifying agents, and Certified Pryme Architect covers regulated multi-agent deployments. For teams, an admin view shows progress, results, certificates and expiry dates, which is the raw material for each cell. More on the whole picture is in the guide to AI literacy and certification for people who work with agents.