ServiceNow Predictive Intelligence Assignment Group: Incident Assignment Prediction, Training Records, and Unable to Predict Fixes
Predictive Intelligence reads the short description of a new incident and guesses which group should own it. When the model is trained on enough clean history and the thresholds are sane, that guess is often right. It still stops at the group. No out-of-box solution predicts the person, weighs workload, or checks who is on shift, so a correctly predicted incident can sit in a queue with an empty Assigned to field. This page covers how the prediction runs, why it goes quiet, and where person-level assignment fits beside it.
In short
ServiceNow Predictive Intelligence predicts the assignment group of an incident with an out-of-box classification solution called Incident Assignment, which a before insert business rule calls when the record is created. It needs at least 10,000 training records, trains on the most recent 300,000, drops any group with fewer than 30 records, and writes nothing when the prediction confidence is below the class threshold. It predicts a group, not a person, and does not consider agent availability or workload.
Why it fits
This page is for a ServiceNow admin, ITSM process owner or service desk manager who has turned on, or been asked to turn on, Predictive Intelligence for incident assignment, and whose incidents are either landing in the right group with nobody assigned, or not being predicted at all.
It predicts the group and stops there
The out-of-box incident solutions predict assignment group, category, service and configuration item. None predicts Assigned to. A ServiceNow community FAQ puts the split plainly: Predictive Intelligence uses machine learning to predict the category and assignment group from history, while Advanced Work Assignment decides the agent by availability and capacity.
Small groups are never predicted
The documentation excludes classes with fewer than 30 records in the training data. A new team, a specialist group or a group renamed last quarter cannot be predicted at all, so tickets for them keep landing on the service desk or on the nearest large group.
Below the threshold it writes nothing
Each class has a precision and coverage setting. If the confidence of a prediction is lower than that class threshold, the field stays empty. The troubleshooting docs say exactly this is the root cause when a prediction shows in a REST test but not on the record.
A 100 percent precision setting turns it off
A thread with an accepted answer shows a class threshold of 100 against a confidence of 99.4: nothing is predicted. Setting precision to 100 and coverage to 0 is also the documented way to exclude a class on purpose, so the same setting both hides a class and looks like a broken model.
It does not train on a personal instance
A community developer article states that Predictive Intelligence solutions cannot train on a PDI, and a personal instance shows only the 100 and 0 precision option. A proof of concept built on a developer instance proves the configuration, not the prediction quality.
Every prediction needs a reason you can read
Predictive Intelligence reports precision and coverage per class, not a sentence per ticket. Assigner records the rule behind each person-level assignment, so when an incident lands on the wrong engineer you read one line instead of retraining a model.
Side by side
How Predictive Intelligence assigns an incident, step by step, and where each step leaks
| Step | What the platform does | Configured in | Skipped or wrong when | What you see |
|---|---|---|---|---|
| Solution definition | Classification solution Incident Assignment maps input fields to the assignment group output | Predictive Intelligence, Classification, Solution Definitions | Fewer than 10,000 training records, or input fields that do not describe the problem | Training fails or precision is low for most classes |
| Training | Trains on the filtered history, the most recent 300,000 records at most | Solution definition, training frequency from Run once to Every 180 days | A group has fewer than 30 records, or history was assigned badly | That group never appears as a prediction |
| Prediction | A before insert business rule calls the solution on the short description | Out-of-box business rule on the incident table | Short description is empty or generic when the record is inserted | Low confidence, no value written |
| Threshold check | Compares confidence to the precision and coverage setting of the predicted class | Class Confidence, Precision Coverage Lookup | Class threshold set to 100 or above the model confidence | Assignment group left empty |
| Rules afterward | Business rules or assignment rules that set the group take effect | Assignment Rules, business rules, flows | A rule overwrites the prediction | An info message that PI updated the group, then a different group |
| Person assignment | Nothing out of box; the group queue waits for a member or another mechanism | Advanced Work Assignment, On-Call, scripts, or an external router | No member picks it up | Correct group, empty Assigned to |
Output fields must be a choice field or a string field with a finite set of values; inputs can be string, reference, choice or HTML fields. The Update Prediction Results business rule updates solution precision and coverage when an incident closes. The ITSM Predictive Intelligence Workbench is deprecated since Yokohama and replaced by Task Intelligence for ITSM, which can auto-fill, recommend, monitor only, or turn predictions off.
In detail
What the ServiceNow documentation and community actually say about assignment group prediction
How ServiceNow Predictive Intelligence predicts the assignment group
Predictive Intelligence for Incident ships classification solutions named Incident Assignment, Incident Categorization, Incident Service and Incident Configuration Item, and needs the com.glide.platform_ml and com.snc.incident.ml_solution plugins. Incident Assignment learns from closed incidents which words in the short description tend to end up with which group.
The prediction runs once, at creation. A ServiceNow employee explained in a community thread that the out-of-box classification solution for the assignment group uses a before insert business rule to trigger the prediction, and that if the short description changes before the record is inserted, the prediction uses the changed text. An incident created from an email with a subject line of Help, or from a phone call logged with a placeholder description, is predicted from almost nothing, and that shows up as low confidence and an empty field.
If you need prediction at a different moment, the community pattern is a custom business rule calling sn_ml.ClassificationSolutionStore.get() and then predict() on the solution version. That is also how teams run prediction on catalog tasks or custom tables.
Training requirements: how many records Predictive Intelligence needs
The solution definition documentation sets the floor: the minimum number of records required for classification solution training is 10,000. If you submit more than 300,000, the most recent 300,000 are used. ServiceNow guidance recommends 30,000 to 300,000 high-quality records, and retraining can be scheduled from Run once up to every 180 days.
The number that matters more is per group. Classes with fewer than 30 records in the training dataset are excluded from training. A service desk with 60 groups often has a long tail of small ones: database, identity, a vendor queue, a new team. Those groups cannot be predicted, which is why a model can report good overall precision while specialists never receive a predicted ticket. A Now Assist community article on multi-method prediction makes the same point from the field: groups with limited training data perform poorly, and that is why its authors do not rely on Predictive Intelligence alone.
History quality is the other input. The model learns the assignment you actually made, including the incidents the service desk reassigned three times. If groups were renamed or merged, filter the training data to the period after the change, or the model predicts groups that no longer take work.
Why Predictive Intelligence is unable to predict the assignment group
Most reports of prediction not working come down to three causes, in this order.
The class threshold is above the model confidence. The troubleshooting documentation says that if predictions appear in REST API tests but not on records, the confidence may be below the threshold, and that a threshold higher than the prediction confidence is the root cause of seeing no prediction. In one accepted answer the threshold was 100 and the confidence 99.4; the advice was simply that at 100 percent you will not see any predictions.
The group was excluded. Either it had fewer than 30 records, or someone excluded it deliberately by applying the precision 100, coverage 0 combination under Class Confidence. Both look identical from the incident form.
The model is not trained where you think. Solutions do not train on a personal developer instance, retraining frequency changes need Update and Retrain rather than Update, and a thread titled unable to predict ended with the only available advice: retrain the solution definition.
Tuning has a trade-off you cannot escape. A ServiceNow tuning article describes precision above 70 percent with coverage above 80 percent as typically more effective than agents, and notes that raising recall lowers precision. Lower the thresholds and more incidents get a group, and more of those groups are wrong.
Predictive Intelligence predicts a group, not an assignee
This is the gap most teams discover after go-live. All out-of-box incident solutions output a group, a category, a service or a configuration item. None outputs Assigned to. When a community member asked how to use Predictive Intelligence for the Assigned to field in HR, the only answer was to adapt the assignment group example and build your own.
A ServiceNow community FAQ on Advanced Work Assignment draws the line directly: AWA routes work to queues and agents based on conditions admins define, while Predictive Intelligence uses machine learning to predict the category and assignment group from history. AWA is the piece that determines the agent best suited for assignment based on availability, capacity and, optionally, skills. Predictive Intelligence knows nothing about who is on shift, who already holds eight P2s, or who is on vacation this week.
So a successful prediction produces an incident in the correct group queue, and then the old problem starts: who picks it up. Groups that rely on members to pull from the queue end up with the easy tickets taken first and the hard ones aging.
Licensing: which ServiceNow tier includes Predictive Intelligence
There is no public price list. Under legacy packaging, community members describe Predictive Intelligence as requiring ITSM Professional or Enterprise and not included in Standard, and the Task Intelligence FAQ says it is available with CSM Pro or ITSM Pro. A community article on the April 2026 repackaging lists Predictive Intelligence in the new Foundation tier alongside Virtual Agent and Now Assist, and says Advanced is the new baseline for customers coming from ITSM Professional. That article is a community summary, not an official price sheet, so confirm the entitlement on your own order form before planning around it.
On the Now Assist question, ServiceNow employees have answered the same way twice: Now Assist does not replace Predictive Intelligence for the next few years, and neither capability replaces the other. What was deprecated is the ITSM Predictive Intelligence Workbench, since Yokohama, in favor of Task Intelligence for ITSM.
Where Assigner fits beside Predictive Intelligence, and where it does not
If your problem is getting the incident to the right group, keep Predictive Intelligence, train it on clean history, and set sensible thresholds. Assigner does not classify free text into groups and does not replace that model.
Assigner picks up at the step Predictive Intelligence leaves empty. Once an incident has a group, it assigns a person inside that group by round robin, weighted rotation, current workload, skills and live availability, using rules you write and can read. Every assignment records the reason, so a disputed ticket is explained in one line rather than by a precision chart. It also applies the same rules to work outside ServiceNow, such as leads in a CRM or tasks in an ops tool, and costs $12 per user per month billed yearly. If you already run Advanced Work Assignment and it covers your channels, you may not need another tool; the ServiceNow Advanced Work Assignment page covers where it stops.
Questions buyers ask
Frequently asked questions
What is Predictive Intelligence in ServiceNow?
It is the ServiceNow machine learning capability that trains classification, similarity, clustering and regression solutions on your own records. For incidents, it ships classification solutions that predict the assignment group, category, service and configuration item of a new record from its short description, learned from closed incident history.
How does ServiceNow Predictive Intelligence predict the assignment group?
The Incident Assignment classification solution is called by a before insert business rule when the incident is created. It reads the short description, predicts a group with a confidence score, and writes the group only if the confidence is at or above the precision and coverage threshold set for that group.
What is the minimum number of records for Predictive Intelligence?
The documentation sets 10,000 records as the minimum for classification training and uses the most recent 300,000 when you submit more. ServiceNow recommends 30,000 to 300,000 high-quality records, and any group with fewer than 30 records in the training data is excluded from training.
Why is Predictive Intelligence unable to predict the assignment group?
Usually because the class threshold is higher than the model confidence, so nothing is written. Other causes are a group with fewer than 30 training records, a group excluded with precision 100 and coverage 0, a generic short description at insert, or training on a personal instance, where solutions do not train.
Can Predictive Intelligence assign an incident to a specific user?
Not out of box. The incident solutions predict the group, category, service and configuration item, and none predicts Assigned to. Person-level assignment needs Advanced Work Assignment, On-Call Scheduling, a custom script, or an external assignment engine that picks a member of the predicted group.
Does Predictive Intelligence consider agent availability or workload?
No. It learns from historical assignments and predicts the group that usually handled similar text. A ServiceNow community FAQ contrasts it with Advanced Work Assignment, which is the component that decides the agent by availability, capacity and optionally skills.
Do assignment rules override Predictive Intelligence?
A business rule or assignment rule that sets the group can take effect after the prediction, and users report seeing the info message that Predictive Intelligence updated the group followed by a different group. Decide which mechanism owns the group for each category, or the two will fight.
Which ServiceNow license includes Predictive Intelligence?
Under legacy packaging community sources say ITSM Professional or Enterprise, not Standard, and the Task Intelligence FAQ says CSM Pro or ITSM Pro. A community summary of the April 2026 tiers lists it in Foundation. ServiceNow publishes no price list, so confirm on your order form.
Is Predictive Intelligence replaced by Now Assist?
No. ServiceNow employees have said Now Assist does not replace Predictive Intelligence for the next few years and that neither replaces the other. The ITSM Predictive Intelligence Workbench was deprecated in Yokohama in favor of Task Intelligence for ITSM.
How often should the assignment group model be retrained?
Training frequency can be set from Run once up to every 180 days. Retrain after any group reorganization, merge or rename, and filter training data to the period after the change, or the model keeps predicting groups that no longer take work.
Can Predictive Intelligence work for HR cases?
Yes. HR Service Delivery has an out-of-box solution called Auto assignment group for HR Cases, enabled through the sn_hr_core.case_auto_assignment property, and it shows the predicted group on the case form for the agent to accept or reject.
What precision and coverage should I use?
A ServiceNow tuning article describes precision above 70 percent with coverage above 80 percent as typically more effective than agents. Raising coverage lowers precision, so start per group: strict thresholds for groups where a wrong guess is costly, looser ones for high-volume generic groups.
Keep reading
Related ServiceNow routing pages worth reading next:
- ServiceNow Advanced Work Assignment - The component that picks the agent by availability and capacity.
- ServiceNow assignment rules - The rule model that can overwrite a predicted group.
- ServiceNow assignment group - The group layer every prediction targets.
- ServiceNow round robin - Why normal assignment has no rotation field.
- ServiceNow HR case assignment - Matching rules, skills and the HR auto assign rule.
- Auto populate the assignment group - The rule and script options when a model is overkill.
Stop hand-sorting your incoming work
Let Assigner route it to the right available person by skill, workload, and availability, using rules you control. Nothing sits unassigned and no one gets cherry-picked. Every assignment shows why.