Managed services in the AI era: why the operational layer matters more, not less

Every quarter now brings a new survey confirming the same thing from a different angle: organisations are adopting AI quickly and struggling to get anything durable out of it. McKinsey’s 2025 State of AI survey put real numbers on it. 88% of organisations say they regularly use AI in at least one business function. 72% report using generative AI, more than double 2024’s 33%. And yet nearly two-thirds have not begun scaling any of it across the enterprise, and only around 6% report capturing more than 5% of EBIT (profit before interest and tax) from it.

Gartner has been more specific about why. In July 2024 it predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs and unclear business value. Gartner’s Rita Sallam summed up the underlying problem:

“After last year’s hype, executives are impatient to see returns on GenAI investments, yet organizations are struggling to prove and realize value.”

None of that is a story about AI being unable to build things. It plainly can. It is a story about what happens after the build, which is where most of the actual work of running software has always lived, and which AI adoption has not made any easier.

Building got cheap. Running did not.

An AI agent can scaffold an application in an afternoon. It cannot decide your access control policy, negotiate your uptime commitment, own an incident at 3am, or tell you honestly that a cost is about to spiral before the invoice arrives. Those are still, unglamorously, human and organisational problems. AI has made the first mile of software development dramatically faster without touching the rest of the journey at all.

That mismatch is why the operational layer (patching, monitoring, cost control, incident response) matters more now, not less. When building took months, the operational tail was a smaller fraction of the total effort. When building takes days, it is most of what is left.

The governance gap nobody priced in

The gap is not just technical. It is organisational, and it is happening whether or not leadership has approved it. A 2025 survey of 1,000 US employees who use AI in their jobs, commissioned by WalkMe and conducted by Propeller Insights, found that 78% admit to using AI tools not approved by their employer. That is not a rounding error. That is most of a workforce making its own decisions about which AI tools touch company data, with nobody accountable for what those tools do with it.

Put the two findings together. Most organisations have not scaled their sanctioned AI use, and most of their people are using unsanctioned AI regardless. The result is AI sprawl: more systems doing more things, with less oversight of any of them, at exactly the moment those systems are cheap enough for anyone to stand up without asking.

What “managed” has to mean now

The old model of managed services assumed a relatively stable, slow-changing set of applications that a provider patched, monitored and reported on. That model does not survive contact with a team that can ship a new AI-assisted feature, or a whole new internal tool, before lunch.

What has to change is the cadence, not the principle:

• Every material change gets checked again, not just the first release.
• Cost is watched continuously, because AI workloads have a habit of scaling spend in ways nobody budgeted for.
• Monitoring has a human behind it who can act, not an inbox nobody reads.
• Someone accountable actually knows what was built, which is difficult when it was built in an afternoon by someone outside the engineering team.

That is the job now: not just hosting an application, but continuously re-checking it as it changes, at the speed AI lets it change. Build fast. Then make sure something is actually watching what you built, for as long as it stays live.