A contact centre's AI agent competency matrix shows, for each person, their actual level against their role's target on eight skills: handovers, vulnerability, goodwill release, refusals and withheld actions, knowledge governance, AI disclosure, queue health and teaching the agent. Writing each cell as "actual/target" puts the gaps in plain sight. Template and a filled example below.
Why does a contact centre need a different matrix?
Because the skills are spread across more roles than in most teams. Advisers pick up handovers from the AI agent and release goodwill within limits. Team leads approve above those limits and review refusals. Quality analysts sample the agent's replies. A knowledge manager owns what the agent says. The centre manager reads queue health and decides how work flows. Each role needs a different mix, and a single target for everyone would be wrong for all of them.
The AI agent in a typical centre answers and resolves within bounds at depth 3, drafts replies at depth 2 for sensitive topics, and hands over to a person when it reaches the edge of its bounds. Customers must be told when they are talking to an AI system, an obligation set out in Article 50 of the EU AI Act for organisations in scope.
How do you fill it in?
- Set role targets on a scale of 1 (aware) to 4 (sets the standard).
- Rate each person with evidence: quality samples, practicum logs, observed handovers.
- Write each cell as actual/target, for example "2/3".
- Highlight every cell where actual is below target; those become the training plan.
- Read the result alongside the approver authority matrix for the contact centre, so nobody holds a right their level does not support.
The competency matrix template
Skills
| Code | Skill | Level 3 looks like |
|---|---|---|
| H | Handovers | Reads the AI agent's summary, checks it against the transcript, and resolves without making the customer repeat themselves |
| V | Vulnerability | Spots signals in the transcript and adapts, including exceptions within policy |
| G | Goodwill release | Releases held goodwill and refunds within limits, with reasons |
| R | Refusals and withheld actions | Explains why a refusal stands; knows withheld actions cannot be overridden by anyone in the centre and finds the alternative |
| K | Knowledge governance | Approves changes to the AI agent's knowledge only with a sealed or signed source |
| D | AI disclosure | Keeps the notice that customers are talking to an AI agent in every channel and script |
| Q | Queue health | Reads the workspace: held items ageing, refusal spikes, handover volumes |
| T | Teaching the agent | Submits corrections with sources; checks they took effect |
Who holds what (actual/target)
| Person | Role | H | V | G | R | K | D | Q | T | Credential | Expires |
|---|---|---|---|---|---|---|---|---|---|---|---|
A worked example
Take a sample energy supplier's contact centre. Values are illustrative.
| Person | Role | H | V | G | R | K | D | Q | T | Credential | Expires |
|---|---|---|---|---|---|---|---|---|---|---|---|
| L. Brooks | Centre manager | 3/3 | 3/3 | 3/3 | 4/4 | 3/3 | 3/3 | 3/4 | 2/2 | Operator certificate | 31 Mar |
| C. Diaz | Team lead | 3/3 | 3/3 | 3/4 | 2/3 | 2/3 | 2/2 | 3/3 | 2/3 | Operator certificate | 30 Jun |
| S. Okoro | Team lead | 3/3 | 4/3 | 4/4 | 3/3 | 3/3 | 2/2 | 2/3 | 3/3 | Operator certificate | 31 Dec |
| H. Lee | Knowledge manager | 2/2 | 2/2 | 1/1 | 3/3 | 2/4 | 3/3 | 2/2 | 3/4 | None | |
| R. Ford | Quality analyst | 3/3 | 3/3 | 2/2 | 3/3 | 2/3 | 3/3 | 2/2 | 3/3 | Foundations badge | No expiry |
| A. Baker | Adviser | 3/3 | 2/3 | 3/3 | 2/2 | 1/1 | 2/2 | 1/1 | 1/2 | Operator certificate | 30 Sep |
| N. Yusuf | Adviser | 2/3 | 3/3 | 2/3 | 2/2 | 1/1 | 2/2 | 1/1 | 1/2 | None |
The column that matters most is K. The knowledge manager, whose job is to own what the AI agent says, is at 2 against a target of 4. Every change H. Lee approves reaches every customer the agent speaks to. Until that gap closes, knowledge changes should be co-approved by S. Okoro, who is at target.
C. Diaz holds a certificate but is below target on goodwill release (G) and on refusals (R). That suggests a cap on C. Diaz's goodwill limit for now, not a removal of rights. N. Yusuf needs a supervised period on handovers and goodwill before any release rights at all.
How a manager reviews the agent's decisions day to day is covered in how contact centre managers review an AI agent's decisions. Market context changes the emphasis; see AI agent training for contact centre managers in insurance in the UK and in telecoms in Nigeria.
Where the Academy fits
Pryme Intelligence Academy's Operator track is aimed at reviewers, team leads and operations managers, with modules on human review, teaching the agent, keeping knowledge current and reading the workspace, which line up with columns H to T. Certified Pryme Operator adds a two-week practicum on real review decisions. More in AI agent skills by role; for a matrix built around credit rather than service, compare the competency matrix for a credit team.
Questions people ask
Should advisers be rated on knowledge governance?
At "aware" only. They should know where answers come from and how to flag a wrong one, not approve changes.
How often should a contact centre redo the matrix?
Quarterly, and after any change to the AI agent's depth, its knowledge sources or the centre's structure. Turnover in contact centres makes stale matrices common.
What if someone is above target?
Note it. People above target on V or T are natural coaches and candidates for team lead roles.