Now Assist on ServiceNow: What Earns Its Keep and What Quietly Doesn’t

A CIO at a European insurer called me last month with a problem that is becoming familiar. They had switched on Now Assist for ITSM eight months earlier, paid for the SKU, sat through the keynote demos, and asked their service desk leads to “use the AI.” The dashboard told them adoption was high. The MTTR numbers had barely moved. Customer satisfaction was flat. Two of the agents were quietly turning the summarisation feature off because it kept hallucinating the wrong impacted CIs into resolution notes. The CIO did not want a vendor pitch. He wanted to know which Now Assist features actually pay for themselves on a real instance, and which ones the platform team should leave alone until the next release.

That conversation is the post.

The hidden cost of switching it all on

Most ServiceNow customers I work with treat Now Assist the way they treated Predictive Intelligence in 2019. They flip the toggles, run a demo for the steering committee, and assume that because the model produces output, the output is valuable. It is the same mistake people made with chatbots in 2017. The output looks confident. The screenshots are great. Whether anyone is actually faster or more accurate is a separate question, and almost nobody measures it properly.

The hidden cost of switching all of Now Assist on without a plan is not the licence. It is trust. Every time an agent reads a summary that flattens a nuanced thread into something subtly wrong, or accepts a suggested resolution that closes the wrong incident, the next thirty interactions get treated with more suspicion. After a quarter of that, the AI features stop being used even when they would help. You have paid for capability and bought scepticism instead.

So the question on any serious instance is not “is generative AI on ServiceNow worth doing.” It is “which Now Assist features do we trust, on which records, with which guardrails, and how do we measure the lift.” That is a small list, and most of it is not what the marketing leads with.

Where Now Assist actually pays back, in my experience

I will be specific. These are the features I have seen produce a measurable result on production instances across ITSM, HRSD and CSM in the last twelve months, in roughly the order of confidence I have in them.

The first is resolution note generation on closed incidents, written from the work notes and activity log. This is the unglamorous win. Agents hate writing close notes. Knowledge teams hate the inconsistent quality. Now Assist can take the full activity stream of a closed incident and produce a tight, factual resolution note that is good enough to publish to a knowledge article candidate queue. On the instances where I have measured this honestly, agents save somewhere between forty and ninety seconds per ticket on closure, and the knowledge intake quality improves enough that the knowledge manager actually approves three or four times more candidates per week. The guardrail is simple. The agent always edits before posting. The AI never closes the ticket. Closure remains a human action.

The second is incident summarisation for handover, especially across shifts and across regions. Not the “summarise this case” button on a five-comment ticket. That is a parlour trick. The real value shows up on a P2 that has been bouncing between three groups for nine hours with sixty work notes. A clean, ordered summary of what has been tried, what was ruled out, and what the current hypothesis is, written to the work notes by the AI and reviewed by the assignment group, prevents the classic “fresh group, same first three steps” pattern. I have seen MTTR on long-tail tickets drop by ten to fifteen per cent after this is bedded in, and the operational benefit shows up in the war-room logs before it shows up in any AI dashboard.

The third is virtual agent for the genuinely narrow, genuinely high-volume requests. Password reset. Group membership. Standard catalogue items where the conversational interface beats the form. These have been good on ServiceNow since the pre-Now-Assist conversational interface. The generative layer makes them better at handling messy phrasing, but the win is still about the underlying flow, not the model. If your virtual agent is not deflecting because nobody fixed the catalogue, no amount of Now Assist will rescue it.

The fourth, more cautiously, is knowledge article generation from resolved incidents and problems. Not auto-publish. Generate-as-draft, assign to the knowledge owner, let them edit and approve. The acceptance rate I see is around thirty per cent on a healthy instance. That is a great number when you consider the baseline is “knowledge owner waits for inspiration.”

Beyond that, the picture gets murkier. Agent chat reply suggestions are useful for new starters and a tax for senior agents. Field auto-population on case forms is helpful when the case is well structured and a liability when it is not. Code generation in App Engine is enjoyable and dangerous in roughly equal measure. The “ask Now Assist about this customer” panel is impressive in a demo and rarely changes what a good account manager would do anyway.

Generative AI on ServiceNow is a data problem before it is a model problem

This is the part of the conversation that vendors do not love. The quality of Now Assist on your instance is governed almost entirely by the quality of your knowledge base, your CMDB, your case categorisations, and your assignment data. NowLM is a capable model. It does not know your environment. The grounding it gets comes from the records you give it.

I have walked into instances where the knowledge base has eight hundred articles, three hundred of them are duplicates, two hundred are out of date, and the rest are written for an audience that left the company two reorgs ago. Then someone is surprised when knowledge-grounded responses are mediocre. The model is not the issue. The garden is overgrown. The same is true for CMDB grounding. If your impacted-CI data is fifty per cent right, the AI’s view of impact is fifty per cent right at best, and probably worse because errors compound.

Practically, the order of operations for any team starting on Now Assist looks the same. Clean the knowledge base first, even if it is a quarter’s worth of work. Audit the categorisations and rationalise the choice lists. Establish a CMDB data quality baseline you actually believe. Only then turn on the AI features that depend on those foundations. Skipping that step is the single most common reason Now Assist underperforms, and it is the part nobody puts on the project plan.

The measurement trap most teams fall into

There is a particular bad habit I want to call out. ServiceNow gives you a Now Assist analytics dashboard out of the box, and it is excellent at telling you how often the features are invoked. It is much weaker at telling you whether anything good happened as a result. “Summaries generated this week” is not a business metric. It is an engagement metric, and it tells you almost nothing about value.

The metrics that matter are boring and slow to move. MTTR on closed incidents, broken out by assignment group and severity. First-contact resolution on the service desk. Knowledge article acceptance rate by knowledge domain. Time-to-handover on P2 and P3 incidents that cross assignment groups. Agent edit distance on AI-generated content, which is a clean signal of whether the AI is producing something close to acceptable or something the agent has to rewrite from scratch. If you are not tracking those, you do not know whether Now Assist is helping. You know whether people are clicking the buttons.

The serious teams I work with bake those measurements into Performance Analytics before they switch any AI feature on, so they have a baseline to compare against. The teams that switch it on first and try to measure later end up arguing about anecdotes for six months.

Where to start, practically

If you are early on Now Assist, four moves earn back the time you spend on them.

The first is to pick one use case and own it end to end. Resolution note generation is the easiest first win. Define what good looks like, set the edit-distance target, train the agents in a single thirty-minute session, and measure week over week. Do not switch on five features in parallel and hope the dashboard tells you which one worked.

The second is to fix the data the chosen use case depends on before you go further. If resolution notes is the use case, then knowledge quality and work-note hygiene are your inputs. Get them to a known state. This is unglamorous and it is the step that decides whether the next year of AI work is a slog or a flywheel.

The third is to govern who can switch features on. The Now Assist admin role hands out broad capability. On a healthy instance, the platform team owns enablement, the process owners sign off on the use case, and nobody activates anything in production without a written hypothesis about the metric it should move. That sentence sounds like overhead. It is what stops you from explaining to the steering committee in three months why nine features are switched on and nothing has improved.

The fourth is to read the model’s outputs honestly, sample by sample, for the first month. Pull a random ten resolution notes per group per week and have a senior engineer mark them for accuracy and tone. The signal you get from forty samples a week beats the signal you get from any vendor dashboard.

The honest summary

Now Assist on ServiceNow is real and it is useful. It is also overhyped in the parts you would expect and undersold in the parts that quietly matter. The teams getting value out of it are the ones treating it like any other platform capability. Pick the use case. Fix the data. Measure the lift. Govern the rollout. The teams getting frustrated are the ones who treated the licence as the project.

If you are on Now Assist and the dashboard says adoption is fine but nobody on the floor can tell you what got faster, you have a measurement problem and a data problem. Both are fixable. Neither is fixed by switching on more features.

If you want a second pair of eyes on your Now Assist rollout, the Milic Media 10-Day Instance Health Report covers the AI surface as part of the platform-hygiene and customisations dimensions, and it produces a written, fixed-fee read on whether the foundation under your AI is solid enough to scale. If you want a broader view of how we help mid-market teams use the platform without paying enterprise overhead, the services page lays it out.

Mladen Milic runs Milic Media Kft, a boutique ServiceNow consultancy delivering implementation, health audits and HRSD work across the EU. Reach him at mladen@milicmedia.com.