Half of AI Service Desk Projects Will Be Abandoned. Good.

Gartner says that by 2027, 50% of AI projects at IT service desks will be abandoned because of unforeseen costs, risks, or failure to achieve the projected return on investment.

Good.

Most of them were never real transformation projects to begin with.

They were software purchases pretending to be operating model improvements. They were chatbot rollouts dressed up as autonomy. They were attempts to skip the hard, boring, necessary work of service design by putting an “AI agent” in front of a broken process and hoping the demo would become reality.

That is not transformation.

That is agentwashing.

Agentwashing is what happens when an organization uses AI language to avoid confronting operational disorder. It is the new version of buying a platform to avoid fixing ownership, workflow, intake, escalation, knowledge, measurement, and accountability.

Only now the promise sounds smarter.

What Agentwashing Actually Is

Agentwashing is not just marketing hype. It is not merely vendors exaggerating capabilities, although there is plenty of that. Agentwashing happens when a chatbot gets renamed an “agent” without gaining meaningful autonomy.

It happens when brittle workflow logic gets described as intelligence.

It happens when leadership promises labor reduction before anyone has stabilized intake, triage, ownership, queue management, knowledge quality, or exception handling.

It happens when a language model is expected to compensate for structural ambiguity that the humans themselves have never resolved.

A chatbot with a better publicist is still a chatbot.

And an “agent” operating inside a bad service model is not a breakthrough. It is a faster way to expose the mess.

Why Service Desks Are Especially Vulnerable

Service desks sit directly on top of operational reality. They are where strategy, process, tooling, people, politics, and user behavior collide.

That makes them attractive targets for AI. It also makes them dangerous places to fake maturity.

In many organizations, the service desk is already carrying years of operational debt. Portals force users to guess the right category. Request types are built around internal team structure instead of user intent. Knowledge articles are stale, fragmented, duplicated, or politically owned. Escalation paths depend on the one person who “knows how this really works.” Ticket queues become dumping grounds. Categories drift. SLAs measure motion instead of resolution. Leaders ask for ROI before the workflow is even stable.

Then AI gets layered on top.

The result is predictable.

The agent inherits the same bad taxonomy. It searches the same stale knowledge. It routes through the same unclear ownership model. It escalates into the same overloaded queues. It produces confident answers from weak inputs and then forces humans to audit, correct, explain, and clean up the fallout.

AI does not remove operational debt.

It compounds it at machine speed.

The Real Failure: Automating Ambiguity

The deeper issue is not that AI agents are useless. They are not.

The issue is that many organizations are trying to automate ambiguity.

Unclear demand in.

Unreliable output out.

If the service model is undefined, the agent becomes another intake layer, not a resolution engine. It may feel modern. It may improve the interface. It may create the illusion of progress. But underneath, the same unresolved work still has to go somewhere.

Unplanned and custom work are especially dangerous here.

In the FrontierOps model, work generally falls into patterns: standard/repeatable, planned, unplanned, and custom. AI performs best when the work is standard or repeatable, the decision logic is known, and the knowledge base reflects reality. It struggles when demand is vague, ownership is disputed, resolution paths vary by personality, and exceptions are the norm.

If your service desk lives in Unplanned + Custom work, adding AI does not create maturity.

It just gives chaos a user interface.

That is where risk increases. The agent is no longer just answering questions. It is acting inside a process nobody fully understands. It may classify, route, summarize, recommend, trigger, approve, or escalate based on an operating model that was never designed cleanly in the first place.

You are not deploying intelligence.

You are exposing your ambiguity.

Why the ROI Story Breaks Down

The business case usually looks clean in a spreadsheet:

• Reduce ticket volume

• Improve deflection

• Lower cost per contact

• Shorten time to resolution

• Reduce headcount dependency

• Increase employee satisfaction

On paper, it makes sense.

Then reality shows up.

The savings were modeled against ideal workflows, not actual work. Adoption assumptions were fantasy. Exception handling was ignored. Governance was underfunded. Knowledge cleanup was treated as a one-time task instead of an operating discipline. Risk controls were added late. Review queues appeared. Admin burden increased. Queue managers had to monitor weird routing. Knowledge owners had to rewrite articles. Service owners had to define policies they had avoided defining for years.

And my favorite: "Where's that spreadsheet Steve created in 2021?"

The support burden did not disappear.

It moved upstream.

Instead of agents resolving work independently, humans now babysit the agent, clean up its misses, maintain its knowledge supply, tune its workflows, review its decisions, and explain to leadership why the savings did not materialize.

The spreadsheet said “autonomy.”

The operating model said “babysitting.”

This is not artificial intelligence.

It is artificial confidence.

What Organizations Should Do Instead

The answer is not to reject AI in service management. That would be lazy.

The answer is to stop treating AI as a substitute for operational maturity.

Mature automation follows mature operations. It does not precede it.

Before deploying agents broadly, organizations should stabilize the service model first:

• Define the work

• Clarify ownership

• Clean up intake

• Separate standard requests from exceptions

• Design around user intent, not internal taxonomy

• Make the portal usable

• Remove duplicate categories

• Document resolution paths

• Assign knowledge ownership

• Decide what “good” looks like before asking AI to optimize it

Then constrain AI to narrow, high-confidence use cases:

• Let it summarize tickets

• Let it draft responses

• Let it recommend knowledge

• Let it classify where the categories are clean

• Let it automate requests that have clear inputs, clear approvals, clear fulfillment paths, and clear rollback procedures

Do not start with the messy edge cases and call that innovation. Measure the right things. Deflection by itself is not enough. A ticket avoided is not a victory if the user comes back angry, opens a second ticket, pings someone on Teams, or gets routed into the wrong queue.

Measure containment quality. Measure rework. Measure escalation accuracy. Measure exception rate. Measure time to correct. Measure knowledge freshness. Measure whether humans agree with the agent’s recommendation. Measure the cost of governance, not just the cost of the tool.

Most importantly, put ownership on the workflow, not just the platform.

The tool cannot own the operating model.

The vendor cannot own your service design.

The agent cannot own your accountability.

A Practical Test for Leaders

Before calling something an AI service desk initiative, leaders should ask a few uncomfortable questions:

• Can we clearly define the request types the agent should handle?

• Is the underlying knowledge current, trusted, and governed?

• Do humans agree on the correct resolution path?

• Are escalation rules explicit, or do they depend on tribal knowledge?

• Can we measure success beyond “tickets avoided”?

• Do we know what should happen when the agent is wrong?

• Is there a named owner for exception handling?

• Do we understand the risk of the agent acting inside this workflow?

  
  

If those answers are weak, this is probably not an AI initiative.

It is an operations problem wearing an AI badge.

And that is fine. In fact, it is useful. AI is forcing organizations to confront what was already broken. The problem is not the agent. The problem is the fantasy that the agent can redeem a service model the organization never had the discipline to design.

The Bottom Line

Gartner is not warning that AI has no future in service management. They are warning that fantasy has a half-life. The AI service desk projects that deserve to survive will be the ones built on stable operations, clean service design, governed knowledge, clear ownership, and disciplined measurement. The ones that fail will mostly be the ones that tried to skip those steps.

That failure is not a tragedy.

It is a correction.

An agent cannot save a broken service desk. It can only make the failure harder to explain.

Sources: Gartner, Magic Quadrant for Artificial Intelligence Applications in IT Service Management, September 2025; Gartner Newsroom, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027,” June 25, 2025.