A capability and depth matrix lists every capability an AI agent has and sets how far it may act on each: 1 read, 2 draft, 3 execute routine work within bounds, 4 move money or take hard-to-undo actions, 5 coordinate other agents. Each row records the limits and approver, against the agent's evidence and expiry date.
Why record the depth each capability needs?
Because one agent does many things with very different risk. A servicing agent might read balances, draft letters, update contact details and issue small refunds. Applying one autonomy level alike to all of that either holds back the safe tasks or over-trusts the risky ones. The case is made in why depth is earned per agent and needed per action.
The depth scale itself is defined in depth: what it means for AI agents. This template applies it to a single agent.
When should you use it?
Fill it in during design, update it at every certification, and attach it to the go-live approval. Change one row at a time; widening an AI agent's depth one capability at a time explains why.
Sector teams often keep a pre-filled version for their common capabilities. The bank operations depth matrix and the finance agent depth matrix are two examples. This is the general template behind them.
How do you fill in each field?
- Capability is a verb and an object. "Update customer email address" is a capability; "customer data" is not.
- Requested depth is what the business wants.
- Ceiling comes from your risk appetite for that class of action.
- Evidenced depth is what testing on your own cases supports.
- Depth in force is the lowest of the three. Never higher.
- Limits narrow the capability further: amounts, volumes, customer types, hours.
- Held for names the role that approves work the agent drafts or holds.
- Evidence and expiry reference the agent's test run and the date it lapses.
The template
| Field | Guidance |
|---|---|
| Capability ID | Reference, such as C-07 |
| Capability | Verb and object |
| Systems touched | Where it reads and acts |
| Requested depth | 1 to 5 |
| Ceiling | From risk appetite |
| Evidenced depth | From testing on own cases |
| Depth in force | Lowest of the three |
| Limits | Amount, volume, population, time |
| Held for | Approving role |
| Reversible? | Yes, compensable, window-bound, no |
| Evidence reference | Test run and date |
| Expiry | Date evidence lapses |
| Owner | Named person |
A filled sample
A hypothetical credit card issuer's servicing agent.
| ID | Capability | Requested | Ceiling | Evidenced | In force | Limits | Held for | Reversible? |
|---|---|---|---|---|---|---|---|---|
| C-01 | Read account and transactions | 1 | 1 | 1 | 1 | Verified customers only | n/a | n/a |
| C-02 | Explain a transaction | 1 | 1 | 1 | 1 | None | n/a | n/a |
| C-03 | Update email address | 3 | 3 | 3 | 3 | One change per 30 days | n/a | Yes |
| C-04 | Raise card dispute | 3 | 3 | 2 | 2 | Under 500 | Disputes analyst | Yes |
| C-05 | Refund a late fee | 4 | 3 | 3 | 3 | One per 12 months | n/a | Compensable |
| C-06 | Increase credit limit | 3 | 2 | 2 | 2 | n/a | Credit underwriter | No |
| C-07 | Block a card | 3 | 3 | 3 | 3 | On customer request or fraud signal | n/a | Yes |
| C-08 | Close an account | 2 | 2 | n/a | 1 | n/a | n/a | No |
Three rows show the rule "lowest wins" in action. C-04: evidence supports only drafting, so disputes are held. C-05: the business asked for 4, the ceiling is 3. C-08: no evidence yet, so the agent can only read until it is tested.
C-07 is depth 3 with no holds. Blocking a card protects the customer and is easy to reverse, which is why it can safely be executed immediately.
Where Pryme fits
This matrix is how Pryme thinks about every agent. Attest certifies each agent on your own cases; the agent's role sets a ceiling; your limits narrow further; the lowest wins at the moment of action, and the verdict names the deciding bound. Evidence lapses after 90 days. See agent governance is a bound, not a model.
Questions people ask
What if two capabilities need to happen together?
Treat each separately and let the lower depth govern the combined task. If one step must be held, the whole task waits for that step's approver.
Who decides the ceiling?
The risk function sets ceilings for each class of action, approved through your risk appetite. Agent owners cannot raise them.
How many capabilities should an agent have?
As few as the job needs. Most well-scoped agents have between five and fifteen. Many more usually means the agent is doing several jobs that would be easier to govern as separate agents.