Answer

How can I use an AI agent to sort customer requests?

We would start by having the agent suggest which service team should handle a request and show what information is missing. For a damaged lamp without an order number, a person checks that suggestion and asks the customer for the number. The original message stays visible. Compare this proposed approach with your existing routing rules before deciding whether you need an agent.[S1]

Illustrative example · damaged delivery

How the request reaches the right service team

The customer's message

My lamp arrived damaged. What should I do?

The AI agent's suggestion

Reported issue
Damaged delivery
Suggested team
Complaints
Still missing
Order number

Original message stays linked

A service representative checks and takes over

They compare the suggestion with the message and ask for the order number.

The agent does not approve a refund or close the case.

An illustrative example, not a client case. The arrow leads from the message to the suggestion. The request stops waiting for attention only when a person confirms they have taken over.
Last updated: | Next review: Illustrative service workflow Machine-readable record

The service team still needs to help the customer

Take an illustrative request: “My lamp arrived damaged. What should I do?” The order number is missing. Someone in the service team needs to identify the order and help the customer. We use this invented example to explain where an agent could assist that work.

In our example, the agent could suggest the complaints team because the message reports delivery damage. The receiving person would see that suggestion alongside the original message and the missing order number. They could then check the destination and ask the customer for the detail needed to continue. That is the limited job we would test before giving the agent more work.

Check whether your existing rules already find the right team

If customers select “damaged delivery” in a form, a fixed rule may already be enough. Free-text messages need more interpretation, but that does not automatically call for an autonomous agent. Microsoft’s unified routing separates classification from assignment and supports rules or machine learning for classification.[S1]

Compare the proposed role with the simplest workable option. Our guide to agents and workflow automation explains that choice. Agree with your team which cases go to complaints, billing or technical support, and where unclear requests wait for review.

If you hire an external team, use our questions for choosing an AI agent development partner. Ask them to demonstrate one ambiguous request and show who takes responsibility for it.

Show the reason for the suggestion and the details still missing

Before asking the customer for the order number, the representative needs to check that the request has reached the right team. We would therefore keep the following information together in the service system. The reason for the suggestion points back to the message; the representative does not need access to the agent’s internal reasoning.

What the team seesIllustrative request
Reported issueThe customer says the lamp arrived damaged.
Suggested destinationComplaints, because the message reports delivery damage.
Missing informationNo order number is given. The agent leaves it unknown.
Original requestA link back to the customer’s message.
Next person to actA named representative checks the suggestion and asks for the missing details.

In this starting setup, the agent can read approved requests and save a proposal in the service system. It cannot promise a replacement, approve a refund or mark the case resolved. Those permissions stay outside this role.

Confirm that a person has taken over the request

The lamp request is still waiting for help while the proposal sits in a queue. In the approach shown in the graphic, a person confirms that they have taken over, then follows up on the missing order number. Agree who watches the queue, who covers their absence and when an unaccepted request needs attention. If the connection fails, keep the request open so the team can find and handle it.

The same distinction matters in chat. In Copilot Studio, a chat can be transferred to staff using a connected customer-service system. A simple escalation message is a separate option.[S2] That documentation does not prove an email-ticket handover works. Test your actual channel: does the original request arrive with the proposal, and can the receiving person continue?

Keep the role only if it reduces the team's work overall

Try representative requests under approved conditions, including missing order numbers, two concerns in one message and an unavailable receiving team. Compare with the current process on equivalent work. For all reviewed requests, record wrong destinations, omitted details, time spent correcting the proposal and cases left without an owner.

A quick first assignment is little help if the team spends longer fixing it. Agree the review period and acceptable results before the trial. There is no universal threshold in the sources here. If rules work just as well with less upkeep, keep the rules.

Choose one approved or anonymised request with the person who manages your service queue. Work through it using the current rules, then describe what an agent would add and who would accept its proposal. That gives you a specific role to test. Our agent introduction guide explains how to bring that role into the team.

What this example establishes

  • The request, proposal and graphic describe MING’s suggested method. They show no actual customer or measured reduction in handling time.
  • Microsoft documents product-specific capabilities. Availability and integration must be checked for the system and channel you use.
  • Access to customer data, retention and service policies require decisions for your organisation before a trial.

MING's proposed method, informed by platform documentation. No client deployment or measured improvement is claimed.

Could an agent help with your service queue?

Bring one anonymised request and describe how it reaches the right person today. We can discuss what an agent could add and who would check its suggestion.

Discuss my service workflow

Tell us what you have in mind. We will respond personally.

Sources

[S1]
Overview of unified routing Microsoft Learn · 2026-02-03 Supports: Routing separates classification from assignment. Classification can use rules or machine learning. Product documentation checked 19 November 2026. It describes how classification works in this system.
[S2]
Hand off to a live agent Microsoft Learn · 2026-08-03 Supports: Connected conversation transfer requires configuration. Copilot Studio also offers an escalation message. Copilot Studio standard-harness documentation checked 19 November 2026. Its scope is conversational handover.

Frequently asked questions

Does sorting tickets require an AI agent?
No. If a form field or a clear rule identifies the destination, existing routing may be enough. A classification model can also sit inside a fixed workflow. Consider an agent when it needs to choose further checks based on what the request contains, and test whether that helps.
Can the agent automatically refund a damaged order?
Refunds are outside the example role. They require a separate decision about policy, authority, system access and review. A label such as damaged delivery does not establish that a refund is owed.
How should we handle a request with two different problems?
Keep both concerns visible in the proposal and give the case to a person who can coordinate them. Agree whether to keep one case or split it. Do not silently drop one concern just because the system expects a single category.
All Insights