AI training for customer service: replies and handoffs

One conversation, two pieces of writing

Suppose tracking shows a package as delivered, but the customer says it never arrived. They have included an order number and a tracking screenshot in the conversation. No one has explained the discrepancy yet. In this exercise, the company's approved guidance directs these cases to the order fulfillment team. The agent should pass along the order number and screenshot. The guidance gives no response deadline.

A weak summary says, “Customer is asking about delivery.” The agent taking over cannot tell what the customer has already sent. They may start by requesting the order number and screenshot, even though both are in the thread.

The participant first decides what the customer needs to hear: the materials were received, the delivery needs investigation, and the case will go to the fulfillment team. They use AI to turn those points into a draft, then edit it into a short reply. A sample response from this exercise could read:

I understand the package hasn't arrived, even though tracking says “delivered.” Thank you for sending the order number and screenshot. I'll pass them to our order fulfillment team and ask them to look into what happened to your package.

The customer can see that the agent read their messages. The reply describes the planned next step without adding a deadline that the guidance does not provide. The participant can compare it with a longer AI draft and decide which sentences help the customer and which simply repeat the thank-you.

Next comes the internal note. A colleague needs different information from the customer:

The customer says the package hasn't arrived, although tracking shows “delivered.” The order number is in their message, and the tracking screenshot is attached. Both need to go to the order fulfillment team to investigate what happened. The case hasn't been forwarded yet.

A reference to the message containing those materials helps the next agent open them and continue without asking the customer to send them again. In this exercise, the participant has only prepared the handoff text, so the note makes clear that forwarding the case is still pending.

Feedback that changes the next draft

An instructor can spend time on one specific editing choice. If a participant shortens the note to “Delivery problem, needs checking,” the instructor can show what the reader lost: the information about the order number and attachment. That shorter note forces a colleague to reread the conversation. The participant restores those details while removing repeated greetings, for example. They learn to shorten a text around what its reader needs.

A second person reads the revised note and identifies the materials they can use to take over. This is also a useful point to discuss the tone of the customer reply. A long assurance about the team's commitment can become a sentence explaining what happens next. The instructor works with the participant's own draft, and the participant makes the edit and compares the versions.

In the next exercise, the customer again reports a package that has not arrived despite being marked delivered. They supply the order number but say they could not attach the screenshot. The agent now needs to write a short request for that screenshot without asking for the order number again. The internal note should make clear what the team is waiting for. It is a small change, but it requires reading the conversation and adapting the response instead of repeating the previous answer.

Using the team's knowledge and tools

The company guidance in this example gives the agent a basis for explaining the next step. In a real program, approved service policies or knowledge-base articles could serve the same purpose. Participants should practice turning those documents into language a customer can easily understand. The company determines which guidance applies and what solutions its agents are authorized to offer.

The team may already have writing tools in its support platform. Zendesk documents features for expanding and simplifying comments and tells users to read the revised text before submitting it. Those functions can support an exercise in comparing two versions and choosing the one that gives the customer the most useful information with fewer unnecessary sentences.

For summaries, the team needs to know how much of the conversation the tool can see. According to Zendesk's ticket summary documentation, including internal notes requires the Copilot add-on and the relevant setting; side conversations are excluded. Availability depends on the plan and settings. Before the session, check the account participants will use and agree on which materials they can work with.

Statistics and a workload scenario for 2026

Brynjolfsson, Li and Raymond (2025) found a 15% average increase in issues resolved per hour with AI assistance.

Less skilled and less experienced agents improved their hourly resolution rate by 30% in the same study.

The study covered 5,172 support agents at one software company and its subcontractors. Human agents received suggestions; training or a new handoff-note format requires separate evaluation.

The planning example covers writing a support reply and checked handoff note. Assume 400 cases per month. Manual work takes an assumed 10 minutes per case. After the workflow change, assume 7.5 minutes per case, including source checks, the employee decision and corrections. Allow another 2 hours per month for exception review and administration.

Under these assumptions, the current workload is 66.67 hours per month.

Total human effort after the change is 52 hours per month.

The team recovers 14.67 hours per month.

Formulas: current hours; hours after change; recovered hours.

• 400 × 10 / 60 = 66.67 hours
• 400 × 7.5 / 60 + 2 = 52 hours
• 400 × (10 - 7.5) / 60 - 2 = 14.67 hours

Validate these assumptions on comparable cases. Recovered time is capacity for other work. Tool and implementation costs need a separate calculation.

The supporting source provides context for this workflow. The scenario volumes and timings are separate planning assumptions.

Source and calculations checked: 21 October 2026.

Fitting the program around customer service

A beginner may need help producing the first draft. Someone who uses AI daily could work with a longer thread or practice shortening a wordy reply. A shared discussion is useful when everyone has written their own text first. The team lead can help assess whether the reply fits the company's service policies and usual language.

For teams working shifts, the workshop could run separately for several groups or be divided into shorter sessions. The schedule needs to leave time for revisions after feedback, with service coverage arranged so participants can finish the task. Any work between sessions and its review should be part of the agreed program.

Syntalith can tailor team training to the cases, tools, and experience of your staff. A proposed workshop could follow a case from understanding the conversation through writing a customer reply and handing it to a colleague, with time for instructor feedback. For teams working with product documentation, our article on AI training for technical service teams develops a separate example.

We also build knowledge applications, automations, and AI agents. If the difficulty is access to current information, we can discuss work on the system separately. For simple, repetitive messages, improving existing reply templates may be enough. Training makes sense when employees need practice adapting their writing to different cases.

For an initial conversation, describe a situation where customers have to repeat information or replies often need revision. You can start without sharing customer histories or preparing a complete training specification. See Syntalith's pricing for service options.