A casework team's AI agent competency matrix sets out six skills caseworkers need when an agent summarises cases and drafts letters and assessments, describes what each skill looks like at four levels, and shows which caseworker holds which level. The level descriptors do most of the work: they make ratings consistent between team leaders. Template and a filled example below.
What changes for caseworkers when an agent arrives?
The caseworker stops assembling and starts checking. In a typical casework deployment, whether benefits, housing or social support, the agent summarises the case file at depth 1, drafts evidence requests, letters and eligibility assessments at depth 2, and books appointments within rules at depth 3. Any decision that changes what a person receives is held for a caseworker. Public bodies carry extra duties of transparency and fairness, and some public uses of AI are high-risk under the EU AI Act, so the skill of checking a draft properly is not optional.
The matrix captures that skill and the ones around it. It is not a RACI matrix, which says who is responsible for a task; nor an approver authority matrix, which says who may approve. It says who can.
How do you fill it in?
- Agree the level descriptors in Table 1 with all team leaders before anyone is rated.
- Rate each caseworker using evidence: observed cases, a practicum log, an assessment.
- Record credentials and expiry next to the ratings.
- Count how many people are at "proficient" or above for each skill, and how many of those are on each shift.
The competency matrix template
Table 1: skills and level descriptors
| Skill | Learning | Practising | Proficient | Expert |
|---|---|---|---|---|
| C1 Checking a drafted decision | Reads the draft | Checks the draft against the evidence with help | Finds unsupported findings and declared facts treated as verified, alone | Teaches others; reviews contested decisions |
| C2 Plain-language reasons | Knows reasons must be clear | Rewrites with feedback | Writes reasons the person would understand, every time | Sets the team's standard wording |
| C3 Safeguarding signals | Knows the policy | Spots obvious signals | Spots signals in notes and routine contacts; refers the same day | Leads safeguarding review of the agent's rules |
| C4 Data disclosure | Knows who may receive what | Checks with a colleague | Verifies identity and authority before any disclosure | Handles complex third-party requests |
| C5 Reading verdicts and refusals | Recognises held and refused | Explains a verdict with help | Explains the deciding bound and whether a refusal should stand | Feeds patterns back to the agent's owner |
| C6 Teaching the agent | Knows corrections exist | Flags errors to a lead | Submits corrections with the source | Reviews corrections from others |
Table 2: who holds what
| Caseworker | Role | C1 | C2 | C3 | C4 | C5 | C6 | Credential | Expires |
|---|---|---|---|---|---|---|---|---|---|
| Proficient or above |
A worked example
Take a sample council's housing support team of eight. Values are illustrative. L = learning, Pr = practising, P = proficient, E = expert.
| Caseworker | Role | C1 | C2 | C3 | C4 | C5 | C6 | Credential | Expires |
|---|---|---|---|---|---|---|---|---|---|
| H. Bell | Team leader | E | E | E | P | P | P | Operator certificate | 31 Mar |
| O. Adams | Senior caseworker | P | P | E | P | P | Pr | Operator certificate | 31 Aug |
| K. Ray | Caseworker | P | P | P | P | Pr | Pr | Operator certificate | 30 Sep |
| N. Shah | Caseworker | P | Pr | P | Pr | Pr | L | None | |
| B. Cole | Caseworker | Pr | P | P | P | Pr | L | Foundations badge | No expiry |
| V. Lund | Caseworker | Pr | Pr | Pr | Pr | L | L | None | |
| G. Moore | Caseworker | Pr | P | Pr | Pr | L | L | None | |
| I. Quinn | Caseworker (new) | L | L | Pr | L | L | L | None | |
| Proficient+ | 3 | 5 | 5 | 4 | 3 | 2 |
The pattern is clear. Safeguarding (C3) is well covered. Checking drafted decisions (C1) is not: only three people are proficient, and all three are in the senior half of the team. If those three are on leave together, decisions either queue or are approved by people still practising. Teaching the agent (C6) is thin everywhere, which means errors in the agent's drafts are corrected case by case instead of at source.
The team's plan: N. Shah and B. Cole complete a supervised fortnight on C1; H. Bell runs a session on C6 for the whole team; the approver rule for the casework team makes adverse decisions require "proficient" on C1 plus a current certificate; the reviewer rota puts at least one C1-proficient caseworker on every shift. Collections teams face a similar pattern with hardship in place of safeguarding; compare the competency matrix for a collections team. The broader role picture is in AI agent skills by role.
Questions people ask
How many skills should a casework matrix have?
Six to eight. More than that and ratings become guesswork; fewer and important skills, such as disclosure, get folded into general ones.
Should new caseworkers be rated at all?
Yes, from day one. A new starter at "learning" across the board is useful information for the rota.
Can caseworkers rate themselves?
They can propose ratings, which helps the conversation, but a team leader confirms them against evidence before they count.