Write an internal service level agreement for each AI agent that other teams depend on. It states what the agent delivers, the service levels for the agent and for the people who handle its held work, how each is measured, what happens when one is missed, and who answers for it.
What does an internal SLA for an agent cover?
A service level agreement normally sets what a supplier promises a customer. Internally, it sets what one team can expect from another. For an agent, there are two providers: the agent itself, which handles work within its bounds, and the people who approve, review and complete what the agent holds or refuses. A customer's experience depends on both, so the SLA must cover both. Both sides change as the agent's depth widens through train, shadow, certify, widen.
That is the main difference from an SLA for a system. A system SLA measures uptime and response time. An agent SLA also measures how fast held items are cleared, how quickly handoffs are picked up, and how often the agent's output needs correcting.
When should you use it?
Agree it before go-live with the teams that send work to the agent and receive work from it. Review it after the first month in production and then quarterly. The method for choosing the levels is set out in how to set service levels for work done by an AI agent, with a short version in how do you set service levels for an AI agent?.
The AI agent runbook then explains how operations meets each level day to day.
How do you fill in each field?
- Service description says what work the agent does, for whom, in what hours.
- Agent service levels cover availability, time to complete in-bound work, and quality.
- Human service levels cover pickup of held items, handoff response and refusal review.
- Measurement names the data source for each level. Use the agent's own records where possible; surveys and estimates drift.
- Exclusions list what does not count against a level: source system outages, items waiting on the customer.
- Breach response says what happens when a level is missed: review, capacity change, depth change. Lowering depth is a valid response to repeated quality breaches.
- Owners name who answers for each level.
The template
| No. | Section | What to write |
|---|---|---|
| 1 | Parties | Agent owner; teams sending work; teams receiving held work |
| 2 | Service description | Work, customers, hours |
| 3 | Agent availability | Target and hours |
| 4 | Agent completion time | For work within its bounds |
| 5 | Agent quality | QA pass level, critical failures allowed |
| 6 | Held item pickup | Time for a person to act |
| 7 | Handoff response | Time for a person to respond to a handoff |
| 8 | Refusal review | How often refusals are sampled |
| 9 | Measurement | Source for each level |
| 10 | Exclusions | What does not count |
| 11 | Reporting | Frequency and audience |
| 12 | Breach response | Actions by level |
| 13 | Owners | Per level |
| 14 | Review | Date and frequency |
A worked sample
A hypothetical lender's document checking agent, serving the underwriting team.
| No. | Entry |
|---|---|
| 2 | Checks income and identity documents on mortgage applications, 7am to 10pm, seven days |
| 3 | Available in service hours except planned maintenance notified 48 hours ahead |
| 4 | Documents checked within 30 minutes of upload |
| 5 | Daily QA sample: no critical failures; at most 1 in 20 checks needs correction |
| 6 | Held documents (unclear, possibly altered) picked up by a document specialist within 2 business hours |
| 7 | Handoffs to underwriter acknowledged within 4 business hours |
| 8 | 20 refusals sampled weekly by underwriting team lead |
| 9 | Times from verdict record; QA from scorecards |
| 10 | Time waiting for applicant to resend documents |
| 12 | One critical failure: suspend affected capability and investigate. Completion time missed three days running: capacity review. Held pickup missed: approver cover review |
| 13 | Agent levels: head of lending operations. Held pickup: document specialist team lead. Handoffs: underwriting manager |
Row 6 is often missing from agent SLAs. Without it, the agent can meet every level of its own while applications sit for days in a held queue nobody owns.
Internal audit may test whether SLAs are monitored as part of an AI management system; see ISO/IEC 42001 internal audit of AI agents.
Questions people ask
Should the agent's SLA be stricter than the human team's was?
Not automatically. Set levels by what the receiving team needs. Agents are usually faster at completion, but held items still depend on people.
What if the agent meets its levels but customers still wait?
Look at the human levels. Most delays in agent-assisted processes sit in held queues and handoffs, not in the agent.
Is an internal SLA binding?
Not legally. It is a management commitment. Treat breaches seriously anyway, because customers feel them.