Article May 2026
Not hype. Not demos. Not vendor slides. The real signs of an ITSM environment where AI consistently delivers value, and why most teams never get there.
Most teams have seen AI do something, generate a summary, suggest a category, surface a knowledge article. Very few have seen it consistently deliver value. Not occasionally, not in demos, but in day-to-day operations, on real tickets, with agents who'd notice if it stopped working. This is what that looks like.
What does it look like when AI works in ITSM?
AI works in ITSM when incident data is consistent, knowledge is structured, and workflows are connected, allowing the AI to produce reliable, repeatable outcomes that agents come to trust. The features themselves are broadly the same across vendors. What separates working environments from failing ones is the foundation underneath: predictable data, usable knowledge, and consistent process. When those are in place, AI doesn't guess; it recognises.
The tools are broadly the same. ServiceNow Now Assist, Halo's AI suite, Freshservice Freddy, Zendesk's AI tools, they're all built on similar underlying models, trained at similar scale, exposed through similar agent interfaces. A team using one isn't getting fundamentally different AI than a team using another. And yet outcomes vary wildly. The same vendor product produces transformational value at one customer and quiet disappointment at the next.
The difference sits underneath the AI, not inside it. In environments where AI works, data is predictable, knowledge is usable, and workflows are consistent. The model has coherent material to work with, so it doesn't guess, it recognises. The output isn't probabilistic and uncertain; it's recognisably correct often enough that agents stop double-checking. That moment, when agents stop second-guessing the AI, is the operational signal that the foundation underneath is doing its job.
The four signals below describe the operational state of an ITSM environment where AI delivers. Not vendor metrics. Not demo conditions. The things an agent or service desk lead would notice on a Tuesday morning, in production.
1 Data
Categories used consistently across teams. Resolution notes that describe what actually happened, in enough detail that another agent could read them and learn. Duplication is minimal, the same issue isn't logged five different ways across five different inboxes. Noise is filtered before the data reaches the AI.
The AI classifies tickets accurately on first pass. Pattern detection produces clusters that match how the team thinks about its work. Suggested next steps fit the situation. Agents don't have to correct every other suggestion.
2 Knowledge
Working environments don't necessarily have more knowledge. They have better knowledge. Articles follow a standard format. Steps are written so a Tier 1 agent can follow them without context. Articles get reviewed and updated as part of resolving issues, not in a quarterly cleanup that never happens. Ownership is named.
AI knowledge surfacing returns articles agents recognise as useful. Suggestions get clicked, read, and acted on rather than dismissed. Knowledge use rises measurably after AI rollout, because the AI is exposing existing good content, not exposing gaps.
3 Workflows
AI needs context to act meaningfully. In environments where it works, ownership is clear at every step. States are used consistently, not "Open" meaning three different things to three different teams. Handoffs are documented and followed.
AI outputs align with how the work actually happens, not just how the system was originally configured. Routing recommendations land in the right place. Escalations follow predictable paths. The AI isn't making sensible-but-wrong suggestions because it doesn't understand the team's actual operating model.
4 Trust
This is the real test, and it's the only one that compounds. Not "is AI enabled", that's a configuration question. Not "are people using it", that can be performative. The signal is whether agents stop double-checking. Whether suggested replies get sent without modification. Whether KB articles get clicked through to. Whether the AI's read of a situation is the read agents adopt.
Suggestion acceptance rates above 30% sustained over months, not the spike-and-decay pattern most rollouts produce. Agent feedback shifts from "the AI keeps suggesting wrong things" to "the AI saves me time on the routine stuff so I can focus on the hard tickets." Adoption happens organically, not because management is mandating it.
The contrast isn't subtle. Teams in the right column outnumber teams in the left column by a wide margin. The difference between the two is rarely the AI itself, it's what the AI is being asked to work with.
Most teams don't reach this state because they start at the top. AI features get implemented first, the procurement is approved, the licences are paid, the project plan begins at "enable Now Assist" or "deploy Freddy." Data quality, knowledge maintenance, and workflow consistency are addressed later, if at all. The result is a system that looks modern in the dashboards but behaves unpredictably in operation. The features are switched on; the foundation underneath them never quite caught up.
The pattern is so consistent across mid-market and enterprise ITSM that it now arrives faster than most teams expect. Vendor demos look perfect because vendor demos run on curated data. Pilots produce promising numbers because pilots are small, well-supervised, and run on the team's best agents. Then production lands, the data gets messier, the supervision drops, and the metrics drift. The honest answer is almost always upstream of the AI itself.
Teams that get this right do one thing differently. They stop asking "what can AI do for us?" and start asking "what does AI need from us?" That single reframing is the turning point. The first question makes AI a feature to be evaluated. The second makes it a system that requires investment to deliver. The teams getting real value have made that mental shift, usually after a first rollout that didn't work. The teams still struggling are still on the first question.
Working teams don't have better AI. They have a better answer to "what does the AI need from us to actually deliver?", and they invested in providing it.
The shift is easier to see in concrete operational terms than in abstract framing. The two snapshots below describe the same service desk, six months apart, running the same vendor product the whole time.
An IT director enables Now Assist for the service desk team. First two weeks, suggestion acceptance hovers around 22%. By month three, it's 14% and falling. KB articles surfaced by AI are the same five evergreen articles regardless of ticket content, because the underlying KB has only those five worth surfacing. Agents have quietly developed habits to ignore the AI panel. Management dashboards show "AI usage up 40%" because invocation count is rising, but reply acceptance is dropping. The IT director doesn't know which number to believe.
Following a focused programme to standardise categorisation, enforce resolution-note quality, and rebuild the top-50 KB articles with named ownership. Suggestion acceptance is sustained at 34%. New starters reach competent independent handling four weeks earlier than the previous cohort because AI suggestions encode tribal knowledge that used to live with senior agents. KB views are up because the articles being surfaced are actually useful. Two of the team's senior agents have shifted to higher-value problem management work because the AI is reliably handling the routine layer.
The change is not in the AI. The AI is the same product the team turned on six months earlier. The change is in everything underneath, and the change is what determines whether the AI in the dashboard is producing value or producing noise.
Working AI in ITSM isn't a feature. It's a foundation.
The teams getting consistent value from AI in their service desks haven't found a better tool. They've done the unglamorous work underneath, categorisation, knowledge, workflows, governance, that lets the same vendor AI everyone has access to actually deliver. That work is more boring than the demos, more useful than the strategy decks, and more durable than any single AI feature. It's also the only path to the outcomes the AI was bought for.
Distill scores your service desk data across the categories that determine whether AI delivers, categorisation, resolution quality, knowledge coverage, completeness, and noise. Five-minute free assessment, no signup, runs entirely in your browser.
Built for ServiceNow, Halo, Freshservice, Zendesk, and TOPdesk, or any CSV.