Now Assist on ServiceNow: An Honest Look at What Actually Works in 2026

A head of platform at a European insurer emailed me last month with a screenshot. It was a Now Assist summary of a P2 incident, generated inside the agent workspace. The summary read like it had been written by a competent junior analyst. Concise, factual, decent grouping of the timeline. He wanted to know two things. First, was the platform team crazy to be excited about it. Second, why the SecOps team was still refusing to enable it on their queue.

Both questions have the same answer, and it is the answer that anyone actually working with Now Assist ServiceNow deployments in 2026 already knows. The technology works. The rollout does not.

What Now Assist Actually Does Today

Set the marketing to one side. Now Assist is a set of generative AI capabilities baked into the ServiceNow platform, powered by NowLM and, where the customer chooses, by third-party frontier models routed through ServiceNow’s own inference layer. The four capabilities that customers touch most often are case summarization, resolution note drafting, agent chat sidekick, and virtual agent topic generation. There are more. Flow authoring assistance, code generation for scripted actions, playbook suggestion, and a growing set of domain-specific skills for HRSD, CSM, and ITSM. But those four are the ones every customer trials first, so those are the ones worth being honest about.

Case summarization is the workhorse. Point it at a long HR case or a noisy incident thread and it produces a five-line rollup. It is genuinely useful for tier-two agents picking up an escalation. The quality is roughly what you would get from a decent human analyst who had two minutes to skim the record. Not brilliant, but not embarrassing, and consistent across thousands of records in a way no human is.

Resolution note drafting is where the wheels start wobbling. The output is grammatically clean and the format matches the customer’s template. But it hallucinates specifics. It will confidently write “restarted the mid-server node and confirmed sync resumed” when nothing of the sort happened. If your resolution notes feed compliance reports or vendor SLAs, this is a problem, not a productivity boost. The fix is to constrain the model to a strict template with placeholder fields the agent must fill, which cuts the value proposition in half but keeps you out of trouble.

Agent chat sidekick is real. It surfaces likely-related KBs, similar past cases, and suggested next actions inside the agent workspace. On the OMV work we saw response time on tier-one HR cases drop by roughly 15 to 20 percent after go-live, which matches what other mid-size deployments have reported. The gains are real. They are also smaller than the six-figure quotes in the ServiceNow sales deck.

Virtual agent topic generation is the one to be sceptical about. It generates conversational flows from a natural-language prompt. In demos it looks magical. In production it produces flows that need so much manual cleanup a decent bot developer would have been faster from scratch. Wait one more release cycle before betting anything meaningful on this one.

The Real Cost of NowLM Licensing

Now Assist is licensed per user, per module, per year. As of Q2 2026, the list price for the ITSM Pro Plus SKU with Now Assist bundled is roughly 60 to 90 percent above the ITSM Pro price you were probably already paying. HRSD, CSM, and SecOps each have their own uplift. This adds up quickly.

The pricing conversation nobody in the sales cycle has is the one about consumption. NowLM has token-based consumption limits attached to each user license. The limits are generous for typical case-summarization traffic. They are not generous if you decide to run resolution-note drafting on every case, agent sidekick on every open ticket, and virtual-agent generation across your bot inventory. Customers hit the ceiling faster than they expect, and the overage pricing is not friendly.

Before signing a Now Assist expansion, get the following in writing from your account team. Consumption per user for each capability at your current ticket volume. Overage pricing per million tokens. What happens if a rogue Flow Designer subflow triggers a summarization on every record in a table. That last one has happened to at least three customers I know of in the past twelve months. The bill was memorable.

Where Generative AI ServiceNow Deployments Actually Fail

The technology is not the failure point. The rollout is. Three failure modes come up over and over.

The first is enablement without governance. Platform teams turn on Now Assist across the estate, agents love it in week one, and by week six there are hallucinated resolution notes in the audit trail, compliance is asking pointed questions, and the SecOps team is refusing to let it near their tickets. The fix is to enable per-module, with a written acceptable-use policy for each capability, and to sample outputs weekly for the first quarter.

The second is trusting the model on records it was never trained to handle. Now Assist ships with strong performance on English-language ITSM records at a certain level of formality. The performance drops when the records are in German with heavy technical jargon, or in mixed Hungarian-English HR case notes, or in short-text SecOps records where context lives in linked CIs the model cannot see. Test with your own records, in your own languages, before making rollout decisions. The reference customer figures in the sales deck were generated on English-only, clean, mid-market ITSM data. Your data probably looks different.

The third is skipping the CMDB and identity work. AI on ServiceNow only produces useful answers if the data underneath it is clean. If your CMDB has stale CI relationships, if your user records are inconsistent between SuccessFactors and ServiceNow, if your KB articles have not been reviewed since 2022, then Now Assist will happily summarize garbage and hand it to your agents. The output looks confident. The output is wrong. Nothing about generative AI ServiceNow rollouts removes the requirement for the foundational data work. If anything, it raises the stakes.

The Practical Sequence for Rolling Out AI on ServiceNow

If you are a platform lead being asked to have a plan by the next steering committee, here is the sequence that has worked for the deployments I have been close to.

Start with case summarization on one module, one queue, for one team. ITSM tier-two is the classic first choice. Give it three weeks. Sample the outputs. Ask the team what they would trust it for, and what they would not. That conversation, more than any metric, tells you where the honest boundary sits.

Add resolution note drafting only after you have a template with locked fields. Do not let the model generate freeform resolution text into your audit trail. The convenience is not worth the risk.

Add agent sidekick to the same team once summarization is stable. The gains here are real and cumulative. Track handle time, first-call resolution, and reopens against the pre-Now Assist baseline. If you did not capture that baseline before switching on, you will not be able to prove the case internally when the licensing renewal conversation comes.

Delay virtual agent generation until the next platform release, or until you have a bot developer with time to clean up the output. Neither of these is a criticism of the capability. It is a call about where your effort earns the most return in the next six months.

Do the data work in parallel. CMDB hygiene, KB refresh, user record reconciliation. This is not glamorous, and it is not what the sales deck talks about, but it is where the difference between a good AI on ServiceNow rollout and an embarrassing one is decided.

Where to Start, Practically

If you are twelve months in with Now Assist licenses already purchased and quiet doubts about the return, three moves.

First, ask your account team for a usage report per capability per module. If half the licences are dormant, you have a rollout problem, not a technology problem. That is fixable without more spend.

Second, run a two-week output-quality sample on the capabilities you have live. Have team leads score fifty summaries and fifty resolution notes each. If the accuracy score is below 85 percent on any capability, pause the rollout on that capability and go back to prompt tuning and data cleanup before adding more users.

Third, before the next renewal, benchmark the licensing spend against what the same money would buy in senior agent capacity or in a proper instance health audit. Now Assist is not a shortcut around platform debt. On a healthy instance it is a genuine force multiplier. On a shaky instance it amplifies the noise. Which one you have is a question worth answering before you write the renewal check.

For customers who want a second opinion on where their platform actually sits before doubling down on AI spend, that is exactly the ground our ServiceNow consulting services team walks through in the first two weeks of any engagement. The answer is often less exciting than the vendor roadmap and more actionable than the internal debate.

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.