Now Assist Isn't Magic — It's Architecture

Every ServiceNow customer I talk to wants Now Assist. Far fewer have asked the harder question underneath it: is the platform actually ready to be assisted? The demos are seductive — summarise this incident, draft this knowledge article, resolve this case — and they create the impression that generative AI is a switch you flip. It isn't. It's an architecture you earn.

I've watched two kinds of Now Assist projects. The ones that ship treat AI as the last layer on a healthy platform. The ones that stall treat it as a shortcut around platform debt they were hoping no one would notice.

The data problem nobody wants to own

Generative features are only as good as the records they read. If your CMDB is half-populated, your knowledge base is stale, and your incident data is a free-text swamp, Now Assist will faithfully summarise the mess. The model isn't wrong — your inputs are. Before any AI work, I look at completeness, relationships and freshness. A trustworthy CMDB and a curated knowledge base aren't AI prerequisites that are nice to have; they are the AI.

Scope the skill, not the ambition

The temptation is to point AI at everything. The teams that succeed do the opposite — they pick one painful, high-volume, low-risk workflow and make the assistant genuinely excellent at it. A tightly scoped skill with clear inputs and a measurable outcome beats a sprawling assistant that's vaguely helpful everywhere. Narrow first, then expand from a position of trust.

Guardrails are a design input, not a patch

The question isn't only what the assistant can do — it's what it must never do. Which data can it surface, to whom, and under what role? Where does a human stay in the loop? Designing guardrails after the fact is how pilots quietly die in security review. Build them in from the first sprint and you ship with confidence instead of apologising for the model later.

Now Assist is genuinely transformative — but the transformation belongs to teams who treat AI as an architectural commitment, not a feature toggle. Get the data, the scope and the guardrails right, and the magic takes care of itself.