The criterion
AI ticket resolution. A meaningful portion of tickets are resolved by AI without human involvement. “Meaningful” isn’t a fixed percentage, but it should show up in capacity metrics. Closing five percent of tickets with AI is an improvement whilst closing most of the Tier 1 volume with AI is a different business model.
What counts as a resolution
The definition is narrower than most dashboards assume. A ticket is resolved by AI when the requester's problem is fixed, the record is closed, and no technician took an action in between. Three common cases fail that test:
- Deflection. A knowledge base article surfaced in a chat window and the user gave up. The ticket closed. Nothing was fixed.
- Drafting. The agent wrote the reply and a technician read it, approved it and sent it. That is assisted resolution, and it is valuable, but a person was in the loop.
- Auto-close on silence. The requester stopped responding and a rule closed the record after seven days. Every service desk has these and none of them are AI resolutions.
What does count: a password or MFA reset executed against the identity provider after the requester is verified. A licence assigned or reclaimed. A stuck print queue or service restarted on the endpoint. A distribution list membership changed. A known error remediated by a runbook the agent selected, executed and verified. In each case something in the environment changed, the change was confirmed, and the requester was told.
Why there is no fixed percentage
The index deliberately does not publish a threshold. A ratio is only meaningful against a denominator, and ticket volume is not comparable across MSPs. An operation that has already automated away password resets at the source has removed its easiest wins from the denominator and will report a lower autonomous resolution rate than a competitor who has not, while running the better service desk.
The honest test is whether it shows up in capacity. If a material share of work is closing without technicians, the ratio of tickets to engineers moves, and it keeps moving as clients are added. That is why this criterion is read alongside margin and headcount: the resolution rate is the claim, the capacity curve is the evidence.
The second test is what the technicians now do. Where autonomous resolution is real, the human queue changes composition rather than just shrinking. Routine work leaves it and what remains is escalations, projects and the client conversations nobody wanted an agent to have. Where the claim is inflated, the queue looks exactly as it did before, only shorter during quiet weeks.
The safety boundary
Autonomous resolution is the one criterion where getting it wrong is visible to the client immediately, so the guardrails are part of the claim rather than a footnote to it. We expect an MSP to be able to say which action classes an agent may execute unattended, what identity verification precedes a credential action, what the agent is forbidden to touch, and what happens when a verification step fails.
An operation that cannot describe its boundary has not thought about it, and an agent with an unbounded action set in a production estate is a different risk profile to the one the client agreed to. That expectation connects directly to client transparency: the client should know which actions are AI-driven before one of them lands.
How the index assesses it
We ask for the resolution definition in use, the action classes an agent may execute, and how the MSP separates autonomous closes from deflections, drafts and silent auto-closes in its own reporting. Then we ask what the number has done over the last few quarters, because a rate that has not moved since the platform was deployed usually indicates a pilot rather than an operating model.
Redacted examples of closed tickets showing the executed action and its verification are the most useful evidence a submitter can provide.
Signals and anti-signals
| Signal | Anti-signal |
|---|---|
| A written definition of resolution that excludes deflection and auto-close on silence | The dashboard number includes anything an agent touched |
| Named action classes an agent may execute unattended, with verification steps | The agent can do whatever the API allows and nobody has enumerated it |
| Tickets per engineer has moved and kept moving as clients were added | A high resolution rate with an unchanged capacity curve |
| The human queue has changed composition, not just length | Technicians still work the same mix of tickets, only fewer of them |
Related criteria
- AI-first ticketingThis criterion is about order of operations, not outcomes. Whatever the channel, the first thing that touches the ticket is an agent.
- Margin and headcountAn MSP where the numbers are indistinguishable from a conventional shop is a conventional shop with better marketing.
- Client transparencyDisclosure is now the criterion with a regulatory floor underneath it, and the floor is lower than what a client will actually ask for.
All seven are listed on the criteria hub, and summarised on What is an AI-Native MSP?