Skip to content
Rubatt

AI ticket triage for MSP service desks, with a technician on every client message

AI ticket triage for MSP service desks: what to automate, how to test on past tickets, where approval gates go, and the rules in the US, UK and Canada.

Rubatt engineering7 min read

AI ticket triage on an MSP service desk means an agent reads each new ticket and sets the fields a dispatcher would set: client, contract, category, priority, ticket type and queue, with a suggested owner. Done well, the agent fills in the reversible fields on its own, sends anything it is unsure about to a person with the reason, and never sends a client message or closes a ticket without a technician approving it.

This guide is for MSPs and IT service firms in the US, UK and Canada that want triage built around their own contracts and rules. It covers what to automate first, how to test the agent on your own past tickets, how to roll it out without risking the board, and what changes by country. If you are still deciding between your PSA's built in feature, a packaged product and a custom build, start with who builds custom AI agents for MSPs.

What does AI ticket triage actually decide?

Most of a dispatcher's first minute on a ticket is spent answering the same questions. An agent can answer them in the same order:

  • Who is this? Match the sender to a client and a contact, including people who write from a personal address or a shared mailbox.
  • Is it covered? Check the contract or agreement: managed, block hours, project work or out of scope.
  • What kind of ticket is it? Incident, service request, change, alert or a reply to an existing ticket.
  • How urgent is it? Priority from impact and urgency, using your matrix, plus client tier and any VIP contacts.
  • Is it already known? Link it to an open incident, a parent ticket or an RMM alert storm instead of creating dozens of separate tickets.
  • Where should it go? The board or queue, and a suggested technician based on skills and current load.

Each of these is a field in your PSA. That matters, because a wrong field is easy to see and easy to fix. The agent is not deciding how to resolve the problem. It is preparing the ticket so the right person opens it with the context already there.

What should the agent set alone, and what needs a person?

The useful rule is the one we use in every build: reads run free, and anything that cannot be undone waits for a person. On a service desk that splits cleanly.

Fields the agent can usually set on its own, because a technician can change them in one click:

  • Category, subtype and ticket type
  • Board or queue
  • Links to a parent ticket or open incident
  • An internal note with the context it pulled from documentation, past tickets and the RMM

Actions that should wait for a technician to approve:

  • Any message to the client, including the first acknowledgement if it promises a response time
  • Closing, merging or deleting a ticket
  • Marking work as billable outside the contract
  • Changing priority on a ticket that triggers an SLA clock or an after hours page
  • Anything that touches a client environment, such as a password reset or a license change

The second list is where an approval gate belongs. The technician sees exactly what the agent wants to do and approves, edits or rejects it. A bad suggestion becomes a rejected draft instead of a wrong email to a client. We explain the reasoning in more depth in why agents need approval gates.

How do you know it works on your tickets?

Demos use clean tickets. Your board does not. Before the agent touches anything live, test it against a set of your own past tickets.

  • Build the test set from history. Take a few hundred closed tickets across your clients, boards and seasons, including the messy ones: forwarded chains, screenshots with no text, tickets that changed category halfway through.
  • Label them with your best dispatcher. Use the fields the ticket should have had, not the fields it ended up with. Where two people disagree, write down the rule. That rule usually belongs in the agent's instructions.
  • Score each field separately. Client matching, contract, category and priority fail in different ways. One overall score hides the field that will hurt you.
  • Look at the misses by client. An agent that is right on average can still be wrong on your biggest client, whose naming and contract are unusual.
  • Keep the set. Run it again every time the prompt, the model or your rules change.

We call this evaluating before you connect. It is the step that turns "it looked good in the demo" into a decision you can defend to your service manager.

How should you roll it out?

Roll it out in stages, each one easy to reverse.

  • Shadow mode. The agent reads new tickets and writes its suggestions to a log, not to the PSA. Dispatchers keep working as they do today. You compare the two for a few weeks.
  • Internal notes only. The agent adds its suggested fields and context as an internal note on the ticket. Dispatchers accept or ignore it. You learn which suggestions they trust.
  • Reversible fields live. The agent sets category, queue and links directly. Priority changes and client messages still go through approval.
  • Gated actions. Drafted replies and contract flags appear in an approval queue for the technician.

Each connector ships behind a feature flag, so the PSA write access, the email inbox or the RMM link can be switched off on its own without a rebuild. See how we handle integrations.

What happens when the agent is unsure?

It should say so. Every suggestion carries a confidence level and a short reason. Below the threshold you set, the ticket goes to the normal dispatch queue with the agent's guess and why it was unsure, so the dispatcher starts from a draft instead of a blank form.

Some tickets should skip the agent's judgement entirely and go straight to a person:

  • Possible security incidents. Phishing reports, unexpected MFA prompts, ransomware notes or anything mentioning a breach.
  • VIP and named contacts. Client owners and anyone your contract treats differently.
  • Angry or escalated threads. Tone matters more than category here.
  • Anything the agent cannot match to a client. Guessing the client is how data ends up on the wrong ticket.

What can go wrong?

Plan for these before go live, not after:

  • Ticket text is untrusted input. A ticket can contain instructions aimed at the agent. The agent should treat the body as data, never as commands, and its write permissions should be narrow enough that a manipulated suggestion still has to pass a person.
  • Alert storms. One outage can produce hundreds of RMM alerts. Triage should group them under one incident instead of creating work for every alert.
  • Drift. New clients, new contracts and new products change what correct looks like. Rerun the test set on a schedule.
  • Language. If you serve clients who write in French, as many Canadian MSPs do, test French tickets separately.

Does triage handle personal data differently in the US, UK and Canada?

Yes. Tickets name your clients' staff and often contain screenshots, email addresses and device details, so the agent processes personal data. This is general information, not legal advice. Check your position with your counsel.

  • UK and EU. UK GDPR and the EU GDPR apply. As an MSP you are usually a processor for your clients' data, so your contracts and any data protection impact assessment need to cover the new processing and the model provider you use.
  • Canada. PIPEDA covers private sector organizations across most of Canada. Alberta, British Columbia and Quebec have their own private sector privacy laws, and Quebec's was updated by Law 25. Accountability stays with the organization using the tool.
  • US. No single federal privacy law covers this. State privacy laws and your clients' sectors, such as healthcare or finance, decide what applies, and your client contracts often go further than the law.

In practice: send the model only what triage needs, keep the data in your own accounts, and write down which client data the agent sees. More detail is in our guide to GDPR compliant AI agents and on security and data.

How we build triage agents for MSPs

We build custom AI agents for small and mid-size businesses in the US, UK and Canada, including IT service firms. One example is a Microsoft Teams app for a Swiss cybersecurity firm. Our team is in Pakistan, with working hours that overlap US, UK and Canadian business days.

A triage build with us starts with one board and one set of rules. We test on your past tickets, run in shadow mode, and put approval gates on every client facing or irreversible action. You own the code, the prompts and the evals from the first commit. See support agents and how an engagement works on solutions.

If you have a service desk workflow in mind, book a call. We will tell you plainly whether your PSA's built in feature or a packaged product is the better choice.

Questions

What is AI ticket triage for an MSP?

An agent reads each new ticket and sets the fields a dispatcher would: client, contract, category, ticket type, priority and queue, with a suggested technician and an internal note of context. A technician approves anything that cannot be undone.

Can an AI agent reply to clients on its own?
How do you test AI ticket triage before going live?
Which tickets should skip AI triage?
Does AI ticket triage work with our PSA?
Is AI ticket triage allowed under GDPR, UK GDPR and PIPEDA?

Rubatt engineering

We build agent systems for operations, sales, research and support teams. Everything here comes from something that broke in production first. Reach us at hello@rubatt.com.

Building something like this?

We are happy to talk architecture before there is any commercial conversation.