Six weeks, four use cases, one moderately bruised ego. Here’s what I’d do differently if I were rolling out Now AssistServiceNow's generative AI features built into its products, such as incident summarization and knowledge article drafting. again from scratch.
What earned trust fast
Summarization on long incidents. This was the surprise winner. Agents were copying the AI summary into stakeholder updates within two weeks. Cheap to roll out, immediately useful, low blast radius if it gets something wrong.
Knowledge article drafting. Not the final article, the first draft. It cut KBThe library of articles that agents and employees search for answers, often shortened to KB. AI answers are only as good as the articles behind them. authoring time roughly in half on a sample of 30 articles. The trick was framing it as “you’re still the editor”, not “the AI writes your KB now”.
What felt like solutions looking for problems
Auto-routing low-volume queues. The model had nothing to learn from. We were better off with a hardcoded assignment ruleA rule that assigns a record to a group or person when the record meets fixed conditions. No model is involved. and three lines of business logic.
Free-text catalog search. It worked, but our catalog already had a curated taxonomy that was better than letting people search for “I need a thing for the new laptop”. Users didn’t want serendipity, they wanted predictability.
The one guardrail I’d add on day one
A simple “AI-generated content” disclosure on any record where Now Assist produced text that ended up customer-facing. We added it in week 4, after a stakeholder thought a summary was a human commentary. It should have been there from the start.
Adoption is a trust ladder, not a switch
The biggest mistake I almost made was trying to roll all four use cases out at the same time. We staggered them: summarization first, then KB drafting two weeks later, then the others. Each one got its own short demo, its own feedback channel, and its own decision: keep, tune, or kill.
The capabilities that survived weren’t always the most impressive ones. They were the ones where humans could see the AI’s work, edit it, and feel like they were still in charge.
What I’d tell my past self
- Pilot with one team that wants it. Not the loudest team. Not the most senior team. The one that actually has the pain.
- Measure something before turning it on. “Average handle time on long incidents” is a far better story than “users say they like it”.
- Be loud about what you turned off. Killing a feature you rolled out is a credibility deposit, not a withdrawal.