ServiceNow Now Assist for ITSM Best Practices

The AI features that Now Assist adds to incident and request work include summaries and drafted customer text. This page helps the reader decide which to switch on first, how to measure whether each one earns its place and when to retire a use.

Practice 1: Deploy Now Assist Incrementally for High-Value Use Cases

Now Assist provides AI-assisted features for incident and request workflows. Start with the highest-impact use case: typically incident summarisation, where the assistant condenses a long record for the agent.

Reason: Incremental deployment allows teams to adapt to AI assistance gradually and provides value quickly rather than attempting a comprehensive rollout that takes months.

What good looks like: Deploy Now Assist for incident classification in month one. When an agent creates an incident, the system suggests a category and assignment group based on the description. Agents can accept the suggestion or override it. In month two, add Now Assist for knowledge search: as agents work on incidents, the system suggests relevant knowledge articles. In month three, add recommendations for incident routing based on current workload.

What to avoid: Attempting to deploy all Now Assist features simultaneously. This creates change overload and makes it difficult to assess which features provide value.

Practice 2: Set Clear Expectations About AI Accuracy and Limitations

Reason: AI predictions are probabilistic, not deterministic. The system will sometimes get recommendations wrong. Setting expectations prevents agent frustration and ensures teams use AI as an assistant, not an authority.

What good looks like: Communicate that Now Assist is a tool to accelerate work, not replace human judgment. Agents should review AI suggestions and override when they believe the suggestion is incorrect. Frame AI as a helpful assistant, not an infallible oracle.

What to avoid: Claiming AI accuracy is higher than it truly is or suggesting agents should never question AI recommendations.

Practice 3: Monitor AI Accuracy and Refine Continuously

Reason: AI models improve with feedback. Monitoring agent overrides of AI suggestions reveals where the model needs refinement. High override rates for certain categories indicate the model is weak in that area.

What good looks like: Dashboard showing AI recommendation accuracy by category. Incidents where the agent overrides the AI suggestion are tracked. High-error categories receive additional training data or model refinement. Accuracy metrics improve over time as the system learns from feedback.

What to avoid: Deploying AI and never measuring its performance. Without monitoring, teams cannot tell if AI is providing value or creating work for agents to review and override bad suggestions.

Practice 4: Use Now Assist for Routing and Load Balancing

Reason: AI can optimize assignment by recommending assignment groups based on current workload, agent skills and incident type. This can reduce assignment rework and accelerate incident routing to the best handler.

What good looks like: When an incident is created and categorized, Now Assist recommends an assignment group considering current queue length, agent skills matching the incident type and historical performance. The assignment is automatic or the agent confirms the recommendation, improving initial routing accuracy.

What to avoid: Letting AI assignment override clear escalation rules or violate compliance requirements (such as segregation of duties). AI should support decision-making, not bypass governance.

Practice 5: Integrate Now Assist with Knowledge Management

Reason: AI can surface relevant knowledge articles as agents work on incidents, accelerating resolution without requiring explicit search.

What good looks like: As an agent reviews an incident description, Now Assist suggests knowledge articles that have resolved similar incidents. The agent can link the article directly from the incident, sharing the solution with the customer and creating learning for future similar incidents.

What to avoid: Surfacing knowledge articles that are low-quality, outdated or irrelevant. AI recommendations are only useful if the underlying knowledge is of high quality.

Practice 6: Use Now Assist for Incident Summarization

Reason: AI can summarize long incident descriptions and communication threads, saving agents time reading through extensive context.

What good looks like: An incident has received 20 comments as multiple teams work to troubleshoot. Now Assist generates a summary of the key issue, steps taken so far and current status, allowing a new agent to quickly understand the history without reading every comment.

What to avoid: Using AI summaries as substitutes for detailed investigation. Summaries can miss important details. Agents should review full details for critical incidents, not rely solely on AI summaries.

Practice 7: Set Up Escalation and Audit for High-Risk Decisions

Reason: For critical incidents or high-stakes decisions, ensure human review is required even when AI makes a recommendation.

What good looks like: For critical incidents with potential major impact, AI suggestions are presented for review but not automatically applied. Supervisors must confirm assignments before they take effect. For sensitive incident types (security, privacy), AI recommendations are advisory only; humans make final decisions.

What to avoid: Automating high-risk decisions based solely on AI predictions. Maintain human oversight for decisions with significant business or compliance impact.

Practice 8: Measure Impact on Agent Efficiency and User Experience

Reason: Now Assist should improve efficiency metrics. Track whether AI assistance is actually reducing time to resolve incidents or just adding another step agents must review.

What good looks like: Dashboard showing time to categorize and assign incidents before and after Now Assist deployment. If time decreases significantly, AI is providing value. Track agent satisfaction with AI assistance: do agents find recommendations helpful, or do they feel burdened by having to review and often override AI suggestions?

What to avoid: Deploying AI and assuming it improves efficiency without measuring. Some AI implementations add overhead (agents spending time reviewing bad suggestions) that offsets any time savings.

Running the feature as a service with its own change control

Once the assistant is live, it behaves more like a service component than a feature switch. Its outputs change when the knowledge base changes, when categories are restructured, when access is adjusted and when the underlying model is updated by the platform. Teams that handle these changes as ordinary configuration tend to find out about a problem from an agent’s complaint. A better practice is to give the assistant a change-control path of its own, with a named owner, a list of change types and a standard for what must be re-checked after each one.

A weekly rhythm that keeps the feature honest

Practices tend to work only when they have a rhythm. The routine below is short enough to run alongside normal service reviews, and it gives every role a defined piece of work so that nothing depends on one enthusiastic person.

1. Read a small sample of the week’s assistant outputs, chosen across the busiest queue and one quieter queue.
2. Log each unhelpful output with a cause from a fixed list: stale content, wrong category, missing context, or unclear wording in the agent view.
3. Assign each cause to an owner, and make the knowledge fix or the configuration fix the next action.
4. Check whether last week’s fixes changed the outputs they were meant to change.
5. Publish a one-page note to the team that names the causes found and the fixes made, so agents see that their feedback leads somewhere.

Communicating with customers about assisted replies

Where the assistant drafts text that reaches a customer, the team needs a clear position on disclosure and on who owns the words that are sent. A practical position is that the agent owns every sent message, regardless of where the draft came from, and that the team’s style standard applies equally to drafted and hand-written text. Brief the customer-facing leads on the standard before the feature is widened, because they will be the ones fielding questions about it.

Deciding when to retire a use rather than fix it

Not every weak use deserves repair. If a use has been reviewed through three cycles and the causes keep pointing to information that cannot be maintained, retire the use and record the decision. Keeping a poor feature alive for the sake of the roadmap erodes agent trust in the uses that do work.

Who should own the assistant overall?

A single named owner, usually from the service desk side, with the platform and knowledge owners as signatories for their own changes. Shared ownership with no named lead tends to produce no decisions at all.

How do we handle a platform update that changes outputs?

Run the baseline comparison for the most-used uses before agents see the change, and keep a note of any output that moved. If behavior changed materially, pause the affected use until the owner accepts the new behavior.

Should we publish performance results to the whole company?

Publish the findings and the decisions, not a headline score. Leadership needs to know what was changed and why, and a single number without its causes invites the wrong conclusion.

What if agents want the feature turned on for every queue at once?

Ask them which queue they would start with and what they would check. The answer often shows whether the request is about the feature or about frustration with a different problem in the same queue.

How long should the weekly rhythm last?

Keep it running for as long as the feature is live. When the causes logged each week become repetitive and small, the rhythm can move to a fortnightly cycle, but it should not stop.

Next step

The governance model guide covers oversight of the assistant, and the note on change calendar coordination covers how overlapping planned work is seen and coordinated.