AI is making IT faster, but isn't making it cheaper yet.

Analyst reports tell a tidy story about AI in IT operations. About a third of IT teams have it in production. Roughly one in five run it under formal governance.

But sit down with the people in the thick of it, ask them about the realities of AI, and you realize the story is far more complicated than most reports show.

Over the last few weeks, we've been traveling across the US, meeting IT leaders and talking about how AI is changing the way they work. We wanted to move past the usual AI clichés and get to the questions the reports don't answer. Who signs off? What happens when the AI gets something wrong? And what do you tell finance when they ask what you got for the money?

The conversations ran until the venue closed, and we took copious notes. Here’s what stood out:

1. The CFO isn't seeing ROI yet

AI is saving IT teams time. It isn't saving them money yet.

An IT leader we spoke to has AI handling user onboarding and permission changes. His technicians' utilization has dropped so far that, in his words, "y'all just coasted." But headcount and labor spend haven't moved.

"From an ops perspective, the ROI is there," he said. "From a CFO perspective, it's not hitting."

That gap came up again and again in our conversations. Teams can see hours saved in time entries. Few can turn those hours into lower cost, higher margin or new revenue, and that's all a CFO cares about. Even the hours lost are hard to pin down. Estimates of technician time lost to manual tracking and context switching ranged from 10% to more than 20%, and some teams don't track it at all. 

How leaders are closing the gap:

• Set the bar before you start. One IT leader at a regulated company won't approve an AI project unless it lowers cost, raises margin or drives sales. Ticket counts never reach the C-suite, so they can't justify the spend.

• Know your cost per service. Former Gartner analyst Kenneth ‘kengon’ Gonzalez compares IT to an all-you-can-eat restaurant with a shrinking budget. To show value, you need to know what each service costs, such as email per person per month, and be ready to retire what nobody uses.

• Count tickets prevented, not tickets closed. Users don't care how many tickets you close. They care about the ones they never had to open. One team reached 34% deflection by moving requests to the self-service portal.

• Track time as a by-product. Nobody will log lost time by hand. The tools have to capture it as people work.

2. Nobody agrees on the skills IT needs next

The skills debate is wide open. Leaders split on what their teams need to know now.

Some worry that junior staff will never learn the fundamentals that AI now hides from them. One leader wants people who can still work from the console when something breaks.

Others say critical thinking and talking with users now matter more than console skills. Kenneth Gonzalez puts it as treating AI as a partner, not a tool. AI handles the toil. People bring judgment, taste and relationships.

Both sides have a point. The open question is how teams learn the basics when AI takes over the work that used to teach them.

3. The winning pattern: AI does the work, a human approves it, trust comes from a human approval step

The examples the crowd trusted most all followed the same pattern. AI does the work, and a person clicks approve or deny before anything changes.

One person we spoke to built a custom model with access to his clients' Microsoft, Zoom and Google tenants. When an onboarding request comes in, it spots it, creates the accounts and sends credentials through a secure link. No technician holds admin rights. The AI asks for approval by email and writes a note back to the ticket when it's done. It also flags repeat tickets and suggests runbooks to stop them.

His bottleneck isn't the technology. It's getting each client to agree to it.

An IT leader at another company described a similar shift. His onboarding team went from four people to two, and the two who remain mostly review what the AI produces.

The AI setups leaders trusted most all worked the same way. AI does the work. A person approves or denies it. A note goes back to the ticket.

4. Rules plus freedom beat a big budget

The biggest win we heard about came from rules and freedom, not a big AI budget.

An IT lead's legal team wanted contracts from about 40 locations moved into SharePoint. His engineers estimated four months. He gave an AI model the requirements and the project charter, and it came back with a design for four sites and a migration plan for 700 GB of data. His team checked the plan, and the work was done three days later.

He credits two things. His team spent two years setting rules for how they use AI, and his CIO funded six months of open-ended experimentation with no fixed deliverable.

Not every team gets that freedom. One leader funds his experiments from unspent budget and spare hardware. He even tested AI on writing his own replies. It got the tone close but missed the relationship details only he would know.

5. AI has to prove itself and show its work too

More than half of the leaders who answered a question we had on governance said that they have rolled back or shut down an AI capability after deploying it. One has done it more than once.

One pulled a billing workflow before it could get expensive. Another saw a model drift beyond the job it was meant to do.

Compliance was the sharpest example. A HITRUST assessment rejected its Copilot-based AI architecture because it couldn't produce the evidence auditors needed: logs, screenshots and fact-checks. Attendees working under FDA rules and similar regimes described the same thing. Their audit trails are full of screenshots, and many AI workflows don't create that record on their own.

For teams in regulated industries, AI getting the job done isn't enough. It has to prove what it did.

6. Adoption is ahead of the benchmarks

60% of the leaders we polled have at least one AI capability in production, and two more are piloting. Analyst data puts AI in production at about a third of IT teams. 

Our learnings come from a small group, not a market survey, but lean mid-market teams under pressure to do more with fewer people may be moving faster than the reports show.

The benchmarks have their own blind spot. Analysts say about one in five teams run AI under formal governance, but nobody agrees on what "governed" means.

7. Data cleanup is the cost nobody budgets for

When we asked which AI cost wasn't in the sales pitch, knowledge base and data cleanup came out on top, ahead of retraining and license costs.

Leaders described messy data spread across ERP systems and SharePoint. AI built on that data inherits the mess. Garbage in, garbage out.

There's a skills cost too. Engineers have to learn to pick the right model for each job, not just the biggest one.

What IT leaders can learn and apply in their work

• Translate time into money. Hours saved don't convince a CFO. Work out your cost per service and show how AI lowers cost, protects margin or frees people for revenue work.

• Count the tickets that never get opened. Ken argued that customers don't care how many tickets you close. One MSP reached 34% deflection by sending a client's requests through the portal. Prevention is the better story.

• Put a human approval in the loop. Approve or deny steps, with notes written back to the ticket, earned the most trust in the room.

• Build the audit trail in. If you answer to HITRUST, FDA or similar rules, make sure every AI action leaves logs and evidence before you scale it.

• Clean your data first. Budget for knowledge base and data cleanup at the start, not after the first bad answer.

• Track time as a by-product. Nobody will log lost time by hand. The tools have to capture it as people work.

• Make room to experiment. The biggest win came from a team with clear rules and the freedom to try things.

Looking ahead

The playbook for running AI in IT operations is still being written, and the people writing it are the ones doing the work. 

If you're working through these questions too, join the conversation in the SuperOps AI community.