Frontline support staff should approve only what an AI agent holds at low depth and low value: sending drafted replies, small goodwill credits and verified contact-detail changes. This matrix lists each held action, the depth the agent runs at, the limit a support agent may release, the certificate they need, and where anything above that limit goes.
Why does a frontline team need its own matrix?
Because the frontline is where most held items land. A support agent sees dozens of drafted replies an hour, and the temptation is to treat "approve" as "send". The matrix makes clear that each approval is a decision someone answers for, and that the authority behind it is narrow on purpose.
It also keeps the frontline separate from leadership. Anything beyond a support agent's limit routes upward under the approver rule for the support leadership. Without a written boundary, held refunds drift down to whoever is free, and the record fills with approvals by people who had no authority to give them. The concept itself is covered in what is an authority matrix?.
How do you fill in each column?
- Held action. Copy the capability name from the agent's capability list.
- Depth. Most frontline rows are depth 2 (the agent drafts, a person sends) or depth 3 (routine work within bounds, held only when it meets an exception).
- Who may approve. A role plus the condition, for example "support agent with a current operator certificate".
- Limit. A value, or "n/a" for non-financial actions.
- Must check before approving. The one or two checks that make the approval meaningful, such as identity verified in this contact or the source behind a claim.
- Above the limit, route to. The leadership role that takes over.
- Never approve at this level. Actions the frontline must pass on regardless of value, plus any withheld actions, which no one can release.
Keep the matrix short. If it runs past a dozen rows, frontline staff will not remember it, and the routing should be doing the work instead. The workspace rule that enforces each row is set out in how to set an approver rule for the support team.
The template
| Held action | Depth | Who may approve | Limit | Must check before approving | Above the limit, route to | Never approve at this level |
|---|---|---|---|---|---|---|
Below the table, record the owner, the sign-off date, and the next review date. Review it at least once a year and whenever the agent's depth changes.
A worked example
A hypothetical home broadband provider runs a support agent that drafts replies, applies small credits and updates contact details. Values are samples.
| Held action | Depth | Who may approve | Limit | Must check before approving | Above the limit, route to | Never approve at this level |
|---|---|---|---|---|---|---|
| Send drafted reply | 2 | Any support agent with a current certificate | n/a | Reply matches the policy source the agent cites | Team lead if the reply promises compensation | Replies to a formal complaint |
| Goodwill credit for an outage | 3 | Support agent | 30 | Outage confirmed in the network status source | Team lead up to 250 | Credits to a closed account |
| Change of email address | 3 | Support agent | n/a | Identity verified in this contact, not a previous one | Team lead if verification failed once | Changes requested by a third party |
| Change of bank details for direct debit | 4 | Not at this level | n/a | n/a | Support manager with a second approver | Always route upward |
| Refund of a settled payment to another card | Refused by customer limit | No one | n/a | n/a | Manual process | Always |
Two rows show why training matters. The email change looks routine, but the agent holds it when the customer's identity was verified in an earlier session, and a support agent who approves without re-verifying has turned a held action into an account takeover. The bank details row is never frontline work, because it is depth 4 and a classic target for social engineering.
The pattern is the same in other teams. Compare the approver authority matrix for the IT team, where the risky rows are access grants, and the approver authority matrix for the compliance team, where they are sign-offs on customer communications. What each person needs to know to hold these rights is mapped in AI agent skills by role.
Questions people ask
Can new starters approve anything?
Not until they are certified for it. A new starter can work the queue in shadow, deciding drafts that a certified colleague then releases, and the comparison becomes their coaching material.
What should a support agent do with an item outside their limit?
Pass it on without editing it into something they can approve. Splitting one refund into several smaller credits to stay under a limit is a breach, and the record will show it.
Do frontline limits need to match across shifts?
Yes. Limits belong to the role and the certificate, not the time of day. If night cover is thin, route items to an on-call lead rather than raising frontline limits after hours.