AI ticket routing for any ITSM tool, trained on your own tickets

Your service desk already knows where every ticket goes.

A ticket arrives: "Outlook keeps asking for my password since this morning." Someone reads it, picks a category, picks an assignment group, sets a priority. Thirty seconds, maybe a minute. Then it happens again, a few hundred times a day.

Most of those choices are ones your team has made thousands of times before. And when a ticket lands in the wrong queue, it waits, gets reassigned, and waits again. Ask any service desk how many tickets are filed under "Other".

Your ticket history already holds the answer. Every resolved ticket records what was reported, and where it really ended up. That's exactly what a model needs to learn from.

One model can answer several of these at once: category, group and priority together, with one confidence for the whole answer. Or train one model per decision, so each has its own confidence.

Keyword rules work until someone writes "can't get in" instead of "password", or until the rule list reaches 400 lines and nobody knows which rule fired.

A general AI assistant knows what a password reset is. It doesn't know that, at your company, VPN problems from the Austin office go to Network Ops and everyone else's go to the Service Desk. Your categories are your team's conventions, and conventions are learned from examples, not definitions.

We measured that difference on BANKING77: 13,083 real customer-service messages sorted into 77 overlapping intents. That's the same shape of problem as ticket categories, though not tickets themselves. A model trained on about 10,000 labeled examples got 93.7% of 3,080 unseen messages right. Claude, given a definition of each category, got 84.7%: the trained model made fewer than half as many mistakes. Read the comparison →

The model sits inside the workflow you already have; it doesn't replace it.

1. A ticket is created. Your ITSM tool sends its subject, description and any fields you choose to your StayCharted model, through the API.
2. The model answers with a category, group and priority, and how sure it is.
3. Your rule decides what happens:
   • Confident answers fill the fields, or appear as a suggestion the agent accepts in one click.
   • Unsure answers go to a person, as today. On Essentials and Business, answers below the model’s cutoff also enter StayCharted’s Review Queue. Choose one place to review; corrections made in your ITSM tool must be sent back to StayCharted if you want to use them for retraining.
4. Corrections recorded in StayCharted become training examples. Retrain monthly, and the model follows your current groups and categories, not last year's.

In practice it looks like this (from our test on a ServiceNow developer instance; your categories and confidences will differ):

INC0010007 · "Outlook keeps asking for my password"

Work note   StayCharted: Category Software, Assignment group Service Desk, 80% sure.
            Filled in: Category, Assignment group.

StayCharted never connects to your ITSM tool. It answers the question your workflow sends, and your workflow and your people act.

Any tool that can send an HTTPS request when a ticket is created can ask the model. Where to add the call:

What's available can depend on your edition; your administrator can confirm. We've tested the ServiceNow business rule end to end on a ServiceNow developer instance: step-by-step ServiceNow guide →. The other rows come from each vendor's documentation and haven't been tested by us yet. The classification call can send the fields used during training:

curl -s "https://amt.staycharted.com/api/v1/models/$MODEL_ID/classify" \
  -H "Authorization: Bearer $STAYCHARTED_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input":{"Subject":"Outlook keeps asking for my password","Description":"Since this morning, on my laptop and phone","Location":"Austin"}}'

For a no-code action, follow the StayCharted Zapier guide to map ticket fields to your published model and route the answer.

1. Export 6 to 12 months of resolved tickets from one queue or one service: the subject and description, the fields known when the ticket was opened (such as the caller's department and location), and the final category, group and priority.
2. Use the final values, not the first ones. The first group a ticket went to is often the mistake you're trying to avoid.
3. Leave out anything filled in later, such as resolution notes or "resolved by". They make the test score look great and the model useless on new tickets.
4. Train, and read the report by category. An overall 95% can hide a group the model gets right only half the time.
5. Choose a review cutoff by what a misrouted ticket costs you, then connect it to new tickets.

If the numbers work for your team, you're live. If they don't, you've spent an afternoon and learned how consistently your team categorizes. Our best practices guide covers each step.

For requests arriving by email, see AI email routing.

• ServiceNow REST integration
• Jira automation actions
• Freshservice Web Request node
• Zendesk webhooks
• TOPdesk action sequences
• Halo custom integrations and runbooks
• ManageEngine custom actions

ServiceNow, Jira, Freshservice, Zendesk, TOPdesk, HaloITSM, ManageEngine ServiceDesk Plus, BMC Helix, Ivanti Neurons, SysAid and SolarWinds are trademarks of their respective owners. StayCharted is not affiliated with them.