
TLDR
Treat an edit to an agent’s rules as a change to policy, and route it to whoever owns that policy.
Inventory the AI your technicians actually use, including the tools nobody approved.
Review a use case whenever its data, audience or authority changes, even if the model stays the same.
You’ve approved the support platform, security has signed off on the integration, and the application owner has agreed that employees in one group get access automatically. Then a technician clearing a backlog adds a rule for contractors based on how similar requests were handled last month.
The rule works. The technician has also just changed an access policy, and nobody with the authority to do that was asked.
This is the question an IT service team should bring to ITIL AI governance, and it’s a lot more useful than whether your company “has an AI policy.” Editing a workflow is editing policy, and most service desks still review those edits like config tweaks.
What Does ITIL AI Governance (Version 5) Add?
ITIL® AI Governance (Version 5) is a separate module with its own certification, which PeopleCert launched in July 2026. It rests on two models. The ITIL AI Capability Model describes six capabilities that shape how organizations adopt and use AI. The ITIL AI Governance Improvement Model helps you choose a governance approach that fits your context, needs and risk.
The Capability Model is the 6C model from PeopleCert’s November 2025 white paper: Creation, Curation, Clarification, Cognition, Communication and Coordination. In a service desk that could mean Creation drafting a script, Curation flagging a stale knowledge article, Cognition spotting an odd request pattern and Coordination triggering the actual change in a connected system.
The paper also lays out four governance patterns (directive, guided, federated and autonomous) along two axes: where authority sits and how assurance works. They’re different arrangements and the paper doesn’t rank them, so you don’t need to reorganize before using them. You do need to know which decisions your team can make locally, which need a service owner and which rules apply to every department.
If you run ITIL 4 practices today, a lot of this will feel familiar, and that’s the useful part. An edit to an agent rule is a change, and it belongs in change enablement like any other, with a standard-change model for the low-risk, pre-approved kind.
Start With the AI Nobody Approved
Your inventory has to include the technician who drafts a PowerShell script with a personal AI tool. That tool never touches your ticketing system, and the script still ends up running in your environment.
Mark Bassham, a consultant who beta-tested the module, makes this point in a November 10 article on ITIL.com. Plenty of organizations already use AI beyond their approved systems, and visibility into that use has to come first. The gap goes well beyond IT: a 2025 global study led by the University of Melbourne with KPMG found that only 40% of employees said their workplace had any policy or guidance on generative AI use. It surveyed more than 48,000 people in 47 countries, which tells you about the wider workplace more than about your own team.
Nobody volunteers that they pasted an Active Directory export into a chatbot, so don’t run the inventory like an investigation. Run it as an amnesty or an anonymous survey, and check it against what your proxy and browser telemetry already show. For each use, record the owner, the data classification, where the output goes and whether the tool can act or only suggest.
Where the output goes matters too. A private summary becomes widely visible once someone pastes it into a shared incident channel, and approval to process the original ticket says nothing about that disclosure.
When you find a gap, give technicians an approved route that does the same job, like a sanctioned assistant for script drafting whose output goes through the change process. Then publish that route in the same channel where people ask for help, so nobody has to go looking for it.
Which Decisions Belong to Which Owner?
I don’t think a model or a tool is the right unit of AI approval. The unit that works is a defined piece of service work with a written limit on what it can decide. For an application-access workflow, that might be collecting the request, checking the eligibility rule and adding a qualifying employee to one group. Contractor requests go to the application owner until someone agrees on a rule for them.
Split the ownership before you automate:
The application owner approves eligibility and business exceptions.
The identity or security owner approves the access mechanism and its restrictions.
The service owner owns delivery, escalation and what the employee actually gets.
The workflow operator implements the approved behavior and handles failed runs.
Smaller teams combine these roles, which is fine as long as it’s deliberate. Permission to edit a workflow doesn’t make the editor the owner of every policy the workflow touches.
Then make the approval hard to exceed. Keep a short approval record with the live workflow (requester population, accepted evidence, target system, what the agent can decide alone, the exception owner and how to withdraw it), and scope the workflow’s service account in your identity provider to the approved groups only. If the service account can’t add anyone to the contractor group, a rule edit can’t quietly grant it either.
In Ravenna, that scope lives in the access policy’s eligible and ineligible groups, under a named policy owner (ideally the application owner). A contractor exclusion belongs there, well away from an agent rule someone can edit on a busy afternoon.
A Green Run Can Still Be an Unauthorized Outcome
Suppose a contractor gets access exactly as the edited workflow says. The run reports success, the ticket closes and nobody sees an error. You’ve learned the implementation works, and nothing about whether the application owner ever authorized contractor access.
Data sources cause the same problem more quietly. Add historical tickets to what an agent can retrieve and you’ve also added old exceptions, expired approvals and decisions made for someone else, all of which the agent may now treat as precedent. Treat anything it retrieves as evidence to weigh.
So for every meaningful change, ask whether it alters who can receive the service, what information the agent can use or what it can change. Any yes goes to the owner of that boundary, even if the model and the interface are identical.
Paul Burby, Global Service Management Lead at Mars, describes the healthy version in a bylined ITIL.com article on AI governance controls. In his change-management example, AI checks configuration-item relationships and surfaces previous incidents to suggest a possible wider impact, and the change owner keeps final approval.
You can still pre-authorize repeatable, low-risk work as a standard change, with recorded conditions and a named owner for anything outside them. A hundred clean runs is good evidence for that decision, and it still doesn’t widen the permission on its own.
Does Your Approval Step Let Anyone Say No?
A human approval step is weak when the reviewer only sees the agent’s recommendation. For the access example, show the requester’s identity, the policy version and the evidence behind the eligibility call, with anything missing shown right next to the proposed action. Ravenna sends approvals as a Slack DM or a Microsoft Teams Adaptive Card so the decision reaches approvers where they already work.
Then watch for rubber-stamping. Track each reviewer’s approval rate and time to decision, and make rejecting with a reason a single click. If rejecting takes three screens and approving takes one, your queue will approve its way to empty.
Employees need a way to challenge an outcome too. Someone denied access should reach the responsible team without arguing with the same agent, and the next person should see why it said no. Requests that slide into direct messages instead are the blind spot here, and they’re exactly where a chat-based service desk can see what a portal can’t.
Session three of PeopleCert’s Make AI Count masterclass series, set for November 20 for PeopleCert Plus members, treats when to review, challenge, override or stop AI-supported actions as workflow design, appeals and feedback included.
Plan the stop as carefully as the start. Decide who can pull the kill switch, where the runbook lives and how requests already in flight get reconciled against actual group membership, or you’ll find out mid-incident which employees got half a workflow.
Every Review Should End in a Decision
A governance review should end with one of four outcomes for the use case: keep it, narrow it, change the process or withdraw it, and someone’s name next to the choice.
For application access, measure time until the employee can actually use the app, and define corrections after closure precisely: reopens, access revoked shortly after it was granted and follow-up tickets from the same requester. A falling ticket count means very little if people have moved their unresolved requests into DMs.
Check the groups an average hides, like contractors, new hires and employees in another region, since they often face different eligibility evidence and different support hours.
Your auditors will care about this even if your governance group doesn’t yet. Logical access is a core area of IT general controls for SOX, and an agent granting access is a control that needs the same evidence as a person doing it. NIST’s AI Risk Management Framework includes outcomes for AI inventories, documented roles and responsibilities, and periodic review, and the service-management records you already keep can cover most of that.
For AI service management, this is where governance turns into a service-owner job. At Ravenna we start from the request you’re trying to resolve and the authority that work needs, and automating the request never moves ownership of its policy onto the platform.
Pick one live workflow and trace its last change in authority: find the approval, check who receives its exceptions today and follow one completed request through to what the employee got. If those don’t line up, settle it before the workflow grows.
If you’d like to see where approval records and exception owners sit in Ravenna, our team can walk through it using one of your own workflows as the example.
Frequently Asked Questions
What Is an AI Governance Framework?
An AI governance framework sets out who can approve AI-supported work, what limits apply and how outcomes get reviewed after deployment. ITIL AI Governance (Version 5) and the NIST AI Risk Management Framework are two examples. In a service desk, that means an owner, an approval record and a review cadence per use case.
What Is ITIL 5?
ITIL 5 is the common shorthand for ITIL (Version 5), the latest version of PeopleCert’s IT service management framework. ITIL AI Governance (Version 5) is a separate module with its own certification, launched in July 2026.
What Should You Consider Before Implementing AI in ITSM?
Start with which requests the AI will handle, what it may decide without a person and who owns the exceptions. Then check what data it can reach and how you’ll stop it if something goes wrong, including the requests already in progress.
How Can You Use AI in ITSM Without Losing Control?
Treat changes to an agent’s rules, data sources and permissions as changes, routed to the owner of each boundary. Scope its service account to what was approved, and review outcomes on a schedule so you aren’t waiting for a complaint.




