Troubleshooting ServiceNow Now Assist for ITSM: A Practical Guide

This guide helps ITSM teams work out why Now Assist for ITSM is misbehaving. It starts from the visible symptom, shows where to check feature status, active models, training data and integrations, lists common causes and fixes, and sets out when to roll back or raise a support case.

Symptom-Led Diagnosis: Where to Start

Start by identifying what is actually happening. Is Now Assist making predictions at all, or is it completely silent? Are its predictions usually wrong, occasionally wrong, or right but formatted incorrectly? Does it work for some request types but not others? Getting clarity on the symptom narrows the troubleshooting scope dramatically.

Where to Look in the Platform

Navigate to Now Assist configuration in the Service Desk module or ITSM module, depending on how it is deployed. Check that Now Assist is enabled. If it is disabled, enable it and test a request to confirm it starts making predictions.

Check the Now Assist models that are active. Different models handle different tasks: category prediction, priority prediction, knowledge article recommendation. If a specific capability is not working (for example, knowledge recommendations but category prediction is fine), identify which model that capability uses and check that model’s status.

Review the training data configuration. Now Assist learns from historical requests to make predictions. If the training data is stale, outdated or does not include the types of requests you are now receiving, model accuracy will suffer. Check when the model was last retrained and how many requests were used in training.

Check integration status. Now Assist may integrate with external knowledge bases or diagnostic tools. If those integrations are down, Now Assist cannot fetch knowledge articles or diagnostic data. Look at integration logs for errors.

Common Causes and Fixes

Now Assist makes poor predictions because the model was trained on insufficient or unrepresentative data. If you recently changed your request categories, business processes or terminology, the model is predicting based on old patterns. Fix this by retraining the model on recent data that reflects current business. You may need to manually categorise recent requests so the model has quality training data for new patterns.

Now Assist does not make predictions at all because it is not configured with a knowledge source. If it is supposed to recommend knowledge articles, it needs access to your knowledge base. Verify that the knowledge source is configured and that Now Assist has query permission on that table.

Now Assist recommendations are irrelevant because the knowledge base is not well-maintained. If knowledge articles are outdated, overlap with each other or cover the wrong subject matter, Now Assist will recommend poor articles because it is working with poor source data. Fix this by reviewing and updating knowledge articles, removing duplicates and improving article metadata.

Now Assist predictions are inconsistent across different users or devices because the model configuration differs between them. For example, if one Service Portal shows Now Assist suggestions and another does not, check that both portals are configured with the same model and knowledge source.

Testing and Validation Approach

Create test cases that cover the range of requests you expect. Submit a test request that should trigger a specific category prediction. Submit a request that should receive a knowledge article recommendation. Submit an edge case (unusual request type, vague description) and observe what Now Assist does. Document the results to establish a baseline for comparison.

Test with incomplete information. Submit a request with minimal description and see if Now Assist still makes useful predictions. Submit a request with industry-specific terminology to see if the model understands specialized language. Test with requests that are intentionally ambiguous to identify where the model struggles.

Involve actual users in testing. Have service desk staff try requests they typically handle. Ask them whether Now Assist recommendations are helpful, accurate or frustrating. User feedback often reveals issues that automated testing misses: recommendations that are technically correct but not practically useful, predictions that are consistent but off-topic.

Escalation Points and When to Call Support

If Now Assist is not making predictions despite being enabled and configured, and you have verified that the knowledge source is accessible, this is likely a platform issue that requires ServiceNow support. Escalate with information about when the predictions stopped and what platform changes were made around that time.

If predictions are consistently wrong across multiple request types and multiple users, the model itself may be degraded. ServiceNow support can review model performance metrics and recommend retraining or configuration adjustments. Provide examples of incorrect predictions so they can investigate the pattern.

If Now Assist works for some request types but not others, the cause is usually model training or knowledge base configuration. These you can typically fix without escalating. Retrain the model or update the knowledge source configuration and retest.

Match each visible symptom to the layer that usually causes it

Most assist problems look like one problem from the agent’s chair and turn out to be three different ones underneath. A suggestion that is missing, a suggestion that is present but wrong, and a suggestion that appears in the wrong place each point to a different layer: the data the feature reads, the configuration that decides when it runs, or the display on the form. Work from the symptom to the layer before changing anything, because a fix applied to the wrong layer often makes the original symptom harder to see.

The last row deserves emphasis. A suggestion that disappears for a restricted role is usually the intended result of a sound access design, not a fault. Treat it as a fault only if the restricted role should have been able to read the record, and check the access model in the security and access design guidance before raising it as a defect.

Gather the evidence a support case needs before you raise it

A support case that arrives with a clear reproduction is resolved faster than one that says the assist is not working. Build the evidence package before contacting the vendor, and keep it in the same location as the rest of the change records.

1. Record the exact time of the first failure, the instance and the release in use, and whether any platform change was made in the preceding few days, including changes to the data the feature reads.
2. Choose one failing record and one working record of the same type, and note the fields that differ between them.
3. Reproduce the failure with a test user who has the same role as the reporting agent. Capture the screen, the form view used and the time taken to show or fail to show the suggestion.
4. Check the logs the feature writes for the period of the failure, and export the entries that mention the record or the suggestion identifier. Remove personal data before sending the export outside the organization.
5. State what you have already ruled out, such as access, enablement and the queue configuration, so the first reply does not repeat those checks.

Know when to roll back instead of repairing

When a suggestion type starts producing wrong outputs that agents are acting on, repair is the wrong first move. Each day spent diagnosing is a day in which agents may follow bad guidance. Switch the affected suggestion type off, confirm that agents are told, and check the records that were changed while it was live. The roll-back is reversible, and the diagnosis can proceed at a calmer pace. Re-enable only after a test run has reproduced the fault and confirmed the fix, following the approach in the Now Assist testing guide.

Signs that the problem is not in the assist at all

Some apparent assist failures are caused by the systems around it. Slow forms can be a performance problem on the platform rather than on the feature, so compare form load times with and without the feature enabled before blaming it. Broken knowledge links can look like bad suggestions when the actual fault is an article that has been moved or retired. Search performance problems can make the knowledge panel appear empty. The guide to troubleshooting ServiceNow performance issues is a useful companion when the timings look wrong across the whole form, not only in the assist panel.

Questions raised during a live fault

Should the agent keep using the feature while we investigate?

Only if the outputs have been shown to be harmless for the record type in question. If there is any doubt, the service owner should decide on a per-queue basis, and record the time of that decision. Agents should be told in plain terms what to ignore and what still applies.

How do we tell a real fault from a change in ticket patterns?

Compare the failing period with the same weeks in the previous quarter, grouped by request type. If the shape of the requests has changed, such as a new product generating a new vocabulary, the suggestions may be behaving correctly on data they were not trained for. That is a retraining question, not a defect.

Who should be called first, the platform team or the vendor?

The internal platform owner, always. They can usually check the configuration and the access model quickly, and a clean internal result makes the vendor case faster to resolve. Involve the vendor once the internal checks are documented.

Can a single agent’s complaints be treated as a fault?

Treat them as a lead, not a verdict. Ask for the record reference and reproduce the case. A complaint that does not reproduce for a second tester is often a display issue in one browser, a training gap or a genuine but rare edge case that belongs in the test set.

When a fault will not clear

Send the timeline, the reproduction steps and the list of ruled-out causes, and the discussion can start from the point where the internal work stopped rather than from the beginning.