A ServiceNow instance is ready for Now Assist when it has clean and well-categorised historical data, a maintained knowledge base, defined AI governance, and at least one high-volume use case with a measurable baseline. Most organisations that buy Now Assist licences and see disappointing results have licensing and models that work perfectly — the gap is data quality and governance, not technology.
Why AI pilots stall on ServiceNow
The pattern is consistent across the industry: enterprises buy Now Assist entitlements, run a pilot, and struggle to demonstrate value from features they are already paying for. The reflex is to blame the model. In practice the model is usually the only part of the stack functioning as advertised.
Generative features summarise what exists, retrieve from what is indexed, and recommend based on what has happened before. When incident categorisation is inconsistent, resolution notes read "fixed", and the knowledge base was last curated three years ago, the output reflects that faithfully. The AI is not underperforming — it is reporting your data quality back to you.
The seven-point readiness check
1. Is your historical ticket data categorised consistently?
Now Assist's summarisation and similar-incident features learn from your history. If half your incidents sit in a generic "Other" category, or categorisation changed twice in three years without back-fill, the model has little reliable signal to work from.
Check: what proportion of incidents from the last 12 months carry a specific category and subcategory? Below roughly 70% and categorisation work should precede the pilot.
2. Are resolution notes usable?
This is the single most under-appreciated readiness factor. Resolution summarisation quality is bounded by resolution note quality. An estate where a meaningful share of closures read "resolved", "done" or "user contacted" cannot produce useful summaries regardless of model capability.
Check: sample 50 recently closed incidents. How many contain enough detail for a new engineer to understand what was actually done?
3. Is your knowledge base current and structured?
Retrieval-augmented features depend entirely on the knowledge base. Stale articles, duplicates and unowned content produce confidently wrong answers — considerably worse than no answer.
Check: what percentage of articles have been reviewed in the last 12 months, and does every article have a named owner?
4. Do you have a defined use case with a measurable baseline?
"Deploy Now Assist" is not a use case. "Reduce average handling time on password and access incidents, currently averaging X minutes across Y tickets per month" is. Without a baseline captured before deployment, you cannot demonstrate improvement afterwards — which is how pilots quietly lose sponsorship.
5. Is AI governance defined before deployment, not after?
Who approves what an AI agent may act on? What is logged? What is the escalation path when output is wrong? Which data may leave the instance? These questions arrive eventually. Answering them before go-live costs a workshop; answering them after an incident costs considerably more.
6. Do you understand your actual licence entitlements?
Now Assist capabilities are packaged and entitled differently across ITSM, HRSD, CSM and other workflows, and packaging changes between releases. We regularly find organisations planning against capabilities they have not licensed, or unaware of ones they have. Confirm your specific position with ServiceNow before designing around it.
7. Is your instance on a supported release with the required plugins?
Now Assist capabilities depend on recent family releases and specific plugins. An instance two or three releases behind may need an upgrade before AI work can begin — and that upgrade belongs in the plan as a prerequisite, not a surprise.
Scoring your readiness
| Points met | Position | Recommended next step |
|---|---|---|
| 6–7 | Ready | Proceed to a scoped pilot with baseline measurement in place |
| 4–5 | Nearly ready | Close the specific gaps first — usually knowledge base or governance |
| 2–3 | Foundation work needed | Data quality remediation before any AI investment |
| 0–1 | Not ready | Address core hygiene; an AI pilot now will fail and damage sponsorship |
What good remediation looks like
If you scored low, the work is unglamorous but finite and it improves operations whether or not you deploy AI:
- Categorisation remediation. Rationalise the category tree, then use Predictive Intelligence to back-fill historical records — ML is genuinely good at this specific job.
- Resolution note standards. Introduce a minimum standard with template prompts at closure. Quality improves within a quarter.
- Knowledge base triage. Retire the unowned, merge duplicates, assign owners and review dates to the rest. Usually smaller than feared once dead content is removed.
- Governance workshop. Half a day to define approval gates, logging, escalation and data boundaries.
Start where the answer is verifiable
The best first Now Assist use cases share three traits: high volume, low variance, and an outcome you can check. Password and access requests, routine incident categories with established fixes, and knowledge-answerable questions all qualify. Deploy there, measure weekly against your baseline, and expand only where the evidence supports it.
⚡ Key takeaways
- Most Now Assist disappointments are data quality problems, not model problems.
- Resolution note quality bounds summarisation quality — sample 50 tickets before you pilot.
- Capture a measurable baseline before deployment or you cannot prove improvement afterwards.
- Define AI governance before go-live; retrofitting it after an incident is far more expensive.
Certified ServiceNow consultants and AI practitioners sharing what we learn delivering implementations, agentic AI, and managed services for US enterprises.