The criterion
AI-maintained documentation. Documentation is updated by AI. Client environment records, runbooks, and configuration notes are kept current by automated processes, not by technicians writing up what they did when they get a spare moment. The documentation reflects what the environment looks like right now.
Why documentation is load-bearing here
In a traditional MSP, documentation is the work that happens if there is time, which means it happens late, partially, or not at all. The cost of that has always been real but diffuse: slower onboarding of new technicians, longer resolution on unfamiliar estates, and key-person risk concentrated in whoever remembers how a client's network was actually built.
In an AI-native MSP the cost stops being diffuse. Agents reason over the environment record. If the record says a client runs one firewall and they run two, the triage enrichment is wrong, the remediation runbook targets the wrong device, and the autonomous resolution described under AI ticket resolution becomes an autonomous mistake. Documentation moves from a professional courtesy to a dependency.
That is why this criterion is worth assessing on its own even though it produces no revenue line of its own. It is the substrate the other criteria stand on.
What it looks like in practice
The criterion is that documentation is updated by automated processes, not by technicians writing up what they did when they get a spare moment. Concretely, that separates into three streams that behave differently:
- Environment state. Assets, identities, licences, network topology, backup coverage and configuration baselines are synchronised from the systems of record on a schedule. Nobody types them.
- Change history. When an agent or a technician changes something, the record of the change is emitted by the action itself rather than composed afterwards from memory.
- Prose. Runbooks, client-specific procedures and the reasoning behind a non-standard configuration are drafted from ticket and change data, then reviewed. This is the stream where a human still belongs, and saying so is not a weakness in the claim.
A useful discriminator is what happens when the environment drifts. In a maintained system, drift produces a diff: the record is compared against what discovery now finds, and the difference is raised as an exception. In an unmaintained one, drift produces nothing at all until someone opens a document during an incident and discovers it is two years old.
Staleness is the metric
Volume of documentation is not evidence of anything. The measure that matters is age against change. An MSP that can say when each class of record was last reconciled against the live environment, and can show that the interval is short, has a maintained system. An MSP that reports how many articles it has written has a library.
We are also interested in coverage of the awkward parts of the estate. Cloud tenancies and managed endpoints are the easy case because they expose APIs. Documentation quality is decided by what happens to the on-premises equipment, the vendor-managed line of business application, and the client's own shadow IT, which is the same boundary that shapes automated onboarding.
How the index assesses it
We ask which records are synchronised, from which systems, at what interval, and which are still authored by people. We ask what happens to a record when discovery and the document disagree. And we ask whether agents read this documentation in production, because documentation that no automated process consumes is documentation nobody will notice is wrong.
Screenshots of a reconciliation report, a drift exception, or a change record emitted by an executed action are all acceptable. A tour of a well-organised knowledge base is not, on its own, evidence for this criterion.
Signals and anti-signals
| Signal | Anti-signal |
|---|---|
| Records reconciled against live discovery on a stated interval | Documentation updated when a technician remembers, or during onboarding only |
| Drift raises an exception rather than sitting silently in a document | Nobody knows how old any given record is |
| Agents read the documentation in production, so errors surface quickly | The knowledge base is written for humans and consumed by nobody |
| Prose is drafted from change and ticket data, then reviewed | Prose is generated wholesale and published unreviewed |
Related criteria
- Automated onboardingOnboarding is the most labour-intensive month of an MSP relationship and the one clients judge hardest. It is also the easiest place to see whether the automation is real.
- AI ticket resolutionClosing five per cent of tickets with AI is an improvement. Closing most of the Tier 1 volume is a different business model. This criterion is about which one you are running.
- AI-first ticketingThis criterion is about order of operations, not outcomes. Whatever the channel, the first thing that touches the ticket is an agent.
All seven are listed on the criteria hub, and summarised on What is an AI-Native MSP?