Let us be honest: most IT operations audits are either so long nobody finishes them or so vague nobody learns anything. This one aims for the middle. Grab a colleague and a free afternoon, and work through it together. It targets the everyday function that keeps services running, known as ITOps. For every line, pick an answer: done, not done, or unclear. Unclear is scored as not done until somebody produces proof.
The checks follow the structure of the Ops Atlas ITOps page, which lists the practices, the common mistakes and a few small lab exercises you can mimic with your own data. Tick as you go.
Part 1: requests and incidents
- Is there one place where people ask for help, and do they know where it is?
- Can you tell a service request (a routine ask) from an incident (something broken) in the queue?
- Does each ticket carry a priority from a rule that two people would apply the same way?
- Does the lowest priority openly carry no promise, rather than a target nobody honours?
- Do you keep a list of workarounds with dates?
If you ticked fewer than three, start here. The queue is the foundation of everything below it.
Part 2: problems and repeat failures
- When the same fault returns three times, does someone open a problem record?
- Does that record have an owner and a review date separate from the daily queue?
- Has a problem record ended with a permanent fix in the last quarter?
- Do you review every major incident for causes, not only for blame?
A healthy score here is the strongest single sign that the rest is real. Teams without it tend to look busy while never getting ahead.
Part 3: change enablement
- Do you have at least three lanes: pre-authorised routine, assessed normal, and urgent?
- Does the routine lane suspend itself after repeated failures? The Atlas gives a sample limit of no more than one failure in the last twenty uses.
- Is there a written risk score, with weights, deciding what goes to review?
- Is there one forward schedule that shows clashes and freeze windows?
- Are change records created by your pipeline but approved by people, never by the pipeline itself?
- Does every failed or emergency change get a review?
- If a regulator applies to you, have you checked what it says about independence of approval and after-the-fact sign-off for emergencies?
That last point matters in financial services, where a European regulation has applied since January 2025. Read the source rules or ask a qualified adviser; this checklist only reminds you to look.
Part 4: the map of what exists
- Does every important service have a named owner in your records?
- Are there entries not confirmed in more than sixty days? The Atlas health check flags that age as stale.
- Can you spot orphans, items nobody owns?
- Do any two entries share a serial number?
- Does the record hold what decisions need, rather than everything that could be captured?
As a sanity check, count the items in a small slice, say ten. In the Atlas sample, nine rows looked complete, yet two had gone stale, a third belonged to nobody and a fourth clashed on its serial number. If your slice looks tidier than that, check how you measured it.
Part 5: monitoring and events
- Is the alert path treated as a production service, with an owner?
- Have you tested it in the last month by triggering a harmless alert?
- Are related alerts grouped so one cause creates one ticket? Tools such as Prometheus Alertmanager or the event module in ServiceNow can do this, and the Atlas lists both.
- Do you know what happens to alerts arriving after a maintenance window closes?
The Atlas lab on that last point is instructive. Thirteen alert rows produced two incidents at 08:20. After the window closed at 08:30, three stragglers raised three further incidents, bringing the tally to five opened with one clearing on its own. Muting rules need an end time, and the team needs to know what happens at the edge.
Part 6: service levels and clocks
- Are customer promises split into smaller internal and supplier commitments, each with an owner?
- Do you know which calendar each clock follows and when it pauses?
- Can you explain, to a customer, why a ticket open for days was counted in hours?
- Do reports show the pauses, instead of hiding them?
The Atlas offers a worked example where an eight-hour promise is a chain of desk triage, network diagnosis, a carrier and a buffer. It also contains a ticket that spanned 92.5 wall-clock hours yet counted just 4.5. A team that has not examined its own pause rules will be surprised the same way.
Part 7: traps
- Are you using more process than your team size justifies?
- Do you treat all changes alike?
- Have you added a rule after a failure without removing an old one?
- Does a shadow spreadsheet exist that people trust more than the official tool?
How to run the session
Block three hours and invite one person from each side of the work: someone who answers requests, someone who changes systems and someone who gets the complaints. Print the checklist or put it on a shared screen, and read each line aloud before answering. The aim is to surface disagreement. If two people give different answers to the same line, that line is already a finding.
Ask for evidence on every yes. “We review failed changes” should be followed by “show me the last one”. It takes seconds and removes most wishful ticking. Where evidence is missing, mark the line as no and move on without debate; you are collecting facts, not trying cases.
A worked line, start to finish
Take the item about the alert path. Suppose the team answers yes without hesitation. You ask who owns it, and three people point at three different colleagues. You ask when it was last tested, and the answer is “after the migration, I think”. That yes is really a no, and the action writes itself: name an owner, schedule a monthly harmless test, record the result somewhere visible.
Do the same for the change calendar, the owner column in the records and the review of failed changes. Four such conversions typically fill the improvement list for a quarter, and that is plenty.
Quick wins, section by section
- Requests and incidents: add a single “type” field to your queue and fill it for the next fifty tickets.
- Problems: pick the most repeated fault from last month and open exactly one problem record for it.
- Change: write the three lanes on one page and pin it where changes are logged.
- Records: add an owner to the ten most important services before anything else.
- Events: trigger one test alert this week and time the response.
- Service levels: pick one customer promise and draw it as a chain of smaller promises.
None of these needs a budget line or a new platform. They need a couple of hours each and a person willing to be named. If a quick win stalls, ask why; the reason, whether skills, time or ownership, is itself a useful audit finding.
Scoring yourself
Count the yes answers out of roughly thirty. Over twenty-five means you are doing the disciplined version already, and you should look at refinements. Between fifteen and twenty-five is the common zone: solid habits with gaps, usually in problems, the alert path or the map. Under fifteen means do not panic. Pick one section and fix it this month.
Resist the urge to turn the result into a score for the dashboard. What you want from this is the list of specific gaps, each one paired with a person and a due date.
Where to go from here
If your team is considering machine-assisted alert handling, our comparison of AIOps and traditional IT operations sets out the trade-offs, and a practical tour of monitoring, alerting and root cause analysis use cases fills in the detail for the event layer. Come back and rerun this audit in three months. The second pass is usually more revealing than the first.