Now Assist ServiceNow: What Actually Works and What Doesn’t After a Year of Real Deployments

A head of platform at a European insurer calls me on a Tuesday. They have bought Now Assist as part of a Pro Plus bundle. The executive sponsor wants a demo in six weeks. Sales has been shown a video with a chat window that resolves incidents in one sentence, and the CIO has already told the board that AI will cut ticket handling time by forty percent. The head of platform reads the release notes, opens the Now Assist configuration, and realises he has no idea which of the twenty-plus skills to enable, in which order, or whether any of them will survive contact with the messy reality of his instance.

That call is the reason I am writing this. There is a large gap between the marketing story around Now Assist and what actually lands in production. After a year of watching real Now Assist deployments across mid-market and enterprise customers, I can tell you which capabilities are genuinely worth turning on this quarter, which ones need six months of data cleanup first, and which ones you should quietly skip until the next release.

The Hidden Cost of Turning On Everything

The biggest mistake I see with Now Assist is the one that costs the least on day one and the most by month three. Someone with sc_admin toggles on every skill in the Now Assist Admin console because they were told it comes free with the SKU. The summarisation skills fire on every case, the generative resolution notes appear on every incident, the virtual agent gets pointed at the knowledge base, and a proud email goes out to the business. Then the tickets arrive.

The summariser produces bland recaps that fusion teams do not trust and end up rewriting. Resolution note generation writes convincing paragraphs that reference the wrong CI because the CMDB relationship is stale. The knowledge search pulls a five-year-old KB article that contradicts current policy. Users lose confidence in the first two weeks, and once they stop trusting it they do not come back.

Generative AI on ServiceNow is not a feature you switch on. It is a product you launch. That means picking a narrow use case, cleaning the data that use case depends on, measuring a baseline, and turning it on for one team before you tell the CIO it worked. Everything I am about to describe assumes that framing.

What Actually Works Today

Three Now Assist capabilities have earned their keep in the deployments I have supported this year. They are unglamorous, they are narrow, and they all rely on data you already have.

The first is incident summarisation for L2 handoff. When an incident bounces between assignment groups, the receiving analyst usually reads the last three work notes, misses the middle four, and asks the previous group to repeat itself. Now Assist writes a paragraph summary of the incident thread on demand. It is not perfect, but for a fifteen-comment ticket it saves ninety seconds and closes a real communication gap. The reason this works is that the input is structured, bounded and short. The model is not being asked to reason about your business. It is being asked to compress text you already wrote.

The second is knowledge article draft-back from resolved incidents. After an analyst closes a ticket with a substantive resolution, Now Assist offers to draft a KB article in your knowledge base template using the incident description, work notes and resolution. The analyst reviews, edits, publishes. This is genuinely useful because it removes the friction that has kept your knowledge base at eighty stale articles for the last three years. You will get more KB articles per month in the first sixty days than you did in the previous twelve.

The third is virtual agent for password reset and simple catalog item lookup. Not for anything that requires reasoning. Not for troubleshooting. For deflection of the four or five request types that make up sixty percent of your L1 volume. Point it at Active Directory via a well-scoped IntegrationHub spoke, protect it with SSO, and it will handle a lot of your Tier 0 traffic without complaint. The gain is not that AI is doing something clever. The gain is that a single well-integrated bot replaced the twelve buttons a user used to have to find in the service portal.

Everything else in the Now Assist catalogue is at some point on the maturity curve between promising and dangerous. Which brings me to the next section.

What Needs Serious Preparation Before You Turn It On

There is a family of Now Assist skills that will demo beautifully and fail quietly. Case summarisation for complex customer service cases. Now Assist for code generation in Flow Designer. Now Assist for change management risk scoring. Anything that pulls from your CMDB. Anything that pulls from your HR profile data.

The pattern is the same in every case. The skill depends on the quality of a data set that most instances have neglected for years, and the AI cannot tell that the underlying data is wrong. Now Assist will confidently tell your change manager that a low-risk change touches only two CIs when in reality the CMDB relationships are ten releases behind and the change will take down half of production. It will summarise an HR case using an employee profile that has the wrong manager because the SuccessFactors sync has been failing silently since March.

Before you turn on any Now Assist skill that reads from CMDB, run a real CMDB data quality review. Not a health score dashboard glance. A sample-based audit of the CI classes and relationship types the skill touches. Same for HR profile data if you are running HRSD skills. Same for knowledge base freshness if you are running knowledge search. Generative AI on ServiceNow amplifies the state of your data. If your data is good, it makes people look competent. If your data is stale, it makes them look confidently wrong, which is worse than not knowing.

Generative AI ServiceNow Governance and the NowLM Question

The second-order question everyone asks in the first month is which language model is actually doing the work, where the data goes, and how to reason about it under GDPR. This is where NowLM and the wider Now Assist model routing come in, and it is worth understanding before you sign anything with your DPO.

Now Assist skills can be served by different models depending on the skill and the tenant. Some run on ServiceNow’s own hosted language model, some route to hyperscaler-hosted commercial models, and enterprise customers can bring in dedicated model endpoints. NowLM is ServiceNow’s own family, tuned for ServiceNow-shaped tasks like summarisation and text-to-code. For most European mid-market customers, the practical trade-off is that NowLM has a tighter data residency story and slightly lower ceiling on open-ended reasoning, while third-party models are more capable and require more contractual work to satisfy privacy review.

The governance move I recommend to every client is boring and effective. Write a one-page policy that names, for each Now Assist skill you have turned on, which model serves it, where the inference happens, what data flows to that model, and what the retention is. Get your DPO to sign it once. Renew it at each release cycle. Do not try to invent a corporate AI ethics framework from scratch. Just document the boring facts about which skill uses which model, and put a change process around switching them.

The other governance layer people forget is prompt injection protection. Users can and will type instructions into ticket descriptions that attempt to manipulate what Now Assist does with that text. This is not a theoretical concern in customer service management if you take email or portal input from anyone outside your company. Turn on the content moderation controls that ship with Now Assist, monitor the audit log for the first month, and treat the AI output as something a human must review before it enters a work note.

The Cost Conversation

Now Assist is not cheap. The Pro Plus and Enterprise Plus bundles are meaningful uplifts on the Pro price, and the assumption inside those SKUs is that you will consume a reasonable volume of AI actions per user per month. If you turn on ten skills across a large user base and generate summaries and drafts on every ticket, you will burn through your entitlement and end up in a mid-cycle commercial conversation with your account team.

I tell every client the same thing. Model the AI action volume before you enable the skill. Do not enable summarisation on every incident. Enable it only on incidents that meet a filter, like priority two and above, or incidents that have crossed an assignment group boundary. Do the same for resolution note generation. This is not a cost-saving hack. It is basic engineering discipline. Every AI call has a latency cost, a token cost and a review cost. The skills earn their keep when they fire on tickets where the human time saved is greater than the review time added.

Where to Start, Practically

If you are deciding what to do with Now Assist this quarter, here is the sequence I would follow.

Pick one skill that reads from data you know is clean. Incident summarisation is usually the right answer because it reads from the incident record itself, which is data your team owns. Pick one assignment group as the pilot. Do not use the flagship team. Use the team that has enough volume for the summariser to fire twenty times a day and enough patience to give you honest feedback.

Baseline the metric you care about before you turn the skill on. For incident summarisation the useful metric is time from reassignment to first meaningful work note by the receiving analyst. Measure it for two weeks. Turn the skill on. Measure it for another four weeks. Publish the number, honest, including the misses. This is the evidence you will use to justify the next skill and to defend against the CIO who has been told AI cuts handling time by forty percent.

Only after that pilot is running with credible numbers should you enable a second skill. Knowledge draft-back is a good second because it does not touch live production tickets. Virtual agent for password reset is a good third because it is a self-contained integration decision, not a change to how L2 analysts work.

If your CMDB has not been audited in the last twelve months, do not enable any AI on ServiceNow skill that reads relationship data. Fix the data first, or you will spend the second half of your fiscal year explaining to change managers why they cannot trust the risk score.

The Real Question Is About Your Instance, Not About AI

Every Now Assist conversation I have had this year eventually turns into a conversation about the instance underneath. Are the incident forms clean enough that a summariser has good input? Is the CMDB current enough that risk scoring means anything? Is the knowledge base curated enough that virtual agent responses are trustworthy? Is the HR profile data current enough that case routing is safe? If the answer to any of those is no, the AI layer is not the fix. The instance is.

This is why I run the 10-Day Instance Health Report before I let any client sign off a Now Assist rollout plan. Two weeks, a fixed fee, and a document at the end that tells you exactly which AI skills are safe to turn on now, which need cleanup work first, and roughly how much of that work you are looking at. If you are staring at a Pro Plus renewal and trying to decide whether to actually turn any of it on, that is the document you want on the table. Book the 10-Day Instance Health Report here and we can have a clear scoping call inside the week.

If you would rather start with a broader conversation about what a boutique ServiceNow consultancy looks like on your account, our services page covers the full range of what we take on. Either way, do not turn on ten AI skills in a Friday afternoon and hope. The half-life of a lost user’s trust is much longer than the release cycle that made you think this was easy.

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.