AI Copilots for IT Ops: What 90 Days of Real Usage Data Shows
Ticket deflection rose, but the biggest gains came from faster incident write-ups, not automated fixes.
AI copilots for IT operations have been sold on a bold promise: incidents that diagnose and fix themselves. After ninety days of real usage data from three enterprise IT teams, the reality is more modest — and arguably more useful.
What actually improved
Ticket deflection rose, as expected, mostly for password resets, access requests and how-to questions. But the largest time savings came from somewhere less obvious: documentation.
Engineers used the copilot to draft incident summaries, post-mortems and change records. Work that used to take an hour at the end of a stressful incident now takes ten minutes of editing.
“Nobody joined IT operations to write post-mortems. Giving them that hour back changed how the team felt about the tool.”
What didn't
Automated remediation remained rare. Teams were reluctant to let the copilot execute changes in production, and most of its suggested fixes still needed an engineer's judgement. Where it did help was in narrowing the search — pointing responders to the right logs, recent changes and similar past incidents.
- Mean time to acknowledge improved more than mean time to resolve
- Junior engineers benefited most, especially on night shifts
- Quality depended heavily on the state of the existing knowledge base
The lesson for CIOs
Copilots amplify whatever operational discipline already exists. Teams with clean runbooks, tagged incidents and current documentation saw the biggest gains. Teams without them mostly got faster access to outdated information.
Before buying an IT operations copilot, it may be worth spending a quarter cleaning up the knowledge it will learn from.