Most support teams give new agents access to an AI drafting tool on day one, show them where the button is, and move on. Six months later they wonder why adoption is below 20% and the agents who do use it are sending replies that read like they were written by someone who has never spoken to a customer.
The tool is not the problem. The onboarding is.
I've been responsible for getting new hires up to speed on AI-assisted support workflows at a top US-based company I work with. After a few rough early runs, I built a 30-minute session that runs on every new support hire's first week. Adoption after two weeks is consistently high, and CSAT on AI-assisted replies holds up the same as fully manual ones. Here is exactly what the session covers, and why each piece is in it.
Start With the Mental Model, Not the Interface
The first five minutes of the session have nothing to do with the tool. I spend them on a single idea: AI is a fast first-drafter. You are the decision-maker. The reply that leaves your queue is yours, not the tool's.
This framing matters because most agents arrive with one of two wrong assumptions. Either they think AI will "write the reply for them" and are disappointed when it needs editing, or they're suspicious of it and planning to ignore it entirely. Neither mindset produces good outcomes.
The right mental model is simpler: AI solves the blank-page problem. Staring at a new ticket and composing from scratch burns cognitive load you need for the parts only a human can handle: reading tone, knowing that this customer had a bad experience last month and is already on the edge. AI handles the structure and boilerplate; you handle the judgment and the humanity.
The goal is not to let AI speak for you. It is to give you a faster starting point so your actual intelligence is spent on what the draft gets wrong.
Three Prompting Patterns That Work in Support
After the mental model, I go straight to live practice. No slides. I open the AI tool and we work through three prompting patterns on real (anonymized) tickets from the previous week's queue.
Pattern 1: Help me respond
The most common use case. I paste the ticket and a one-line context note:
Support ticket: [paste message]. Context: [one sentence on customer situation or history]. Draft a professional, warm response that [specific outcome needed].
The context line is what most agents skip, and it is what makes the output actually useful. A bare paste gives generic output. Adding that the customer is on a legacy plan or that this is their third ticket on the same issue produces something that reads like a person wrote it.
Pattern 2: Help me find the policy
Support agents spend more time than anyone admits re-reading policies they have read a dozen times. AI is fast at this, but only if you ask correctly.
I teach agents to paste the relevant KB section directly into the prompt alongside the question:
Based on this policy: [paste KB section]. Customer is asking: [specific question]. What should I tell them, and is there any exception I should check?
This keeps the AI grounded in your actual documentation rather than inventing something that sounds plausible. The exception check at the end has saved a few escalations.
Pattern 3: Summarize this thread
When an agent picks up a long escalation chain, reading the whole thread carefully takes time they do not have. I use:
Summarize this support thread in three to four sentences. Include: what the customer originally wanted, what has been tried so far, and what is still unresolved. [paste thread].
The agent reads the summary, confirms it against the thread (always), and starts their response with full context in under two minutes.
The When-to-Skip Rules
This is the section most onboarding sessions leave out, and it is the most important. I spend roughly eight minutes here.
AI assistance is not right for every ticket, and the agents who overuse it are the ones whose CSAT drops. I give new hires four rules for when to skip the tool entirely and write the response themselves:
- The customer is visibly upset or in distress. Emotional tickets need emotional intelligence at the start. A draft reply from an AI, even a well-prompted one, can sound clinical at exactly the wrong moment. Start from scratch, write it yourself, then optionally use AI to check your tone.
- The answer requires live account data. AI cannot look up a customer's actual subscription status, payment history, or open orders. If the reply depends on something you need to look up in the CRM, look it up first, then decide whether AI helps you draft around that fact.
- You cannot explain the draft to your manager. This is the filter I use personally. Before sending any AI-assisted reply, I ask: if my manager read this reply and asked me to walk through why I wrote it this way, could I? If not, the draft is wrong. Edit until you can own it.
- It is a new edge case with no precedent. AI is good at patterns. Novel situations with no clear policy answer need a human judgment call first and AI assistance on the wording after, not the other way around.
The Pre-Send Checklist
The last five minutes of the session go to a four-point checklist I ask every agent to run before sending an AI-assisted reply. It takes under 30 seconds once it is habit:
- Name and pronouns correct? AI drafts sometimes use a wrong name form or a generic "they" where the customer signed with a different name.
- Anything invented? Read for any claim about a policy, a timeline, or a feature that you did not verify yourself. AI fills gaps with plausible-sounding text. That text may be wrong.
- Does it sound like you? Add one personal sentence, a real acknowledgment, or a specific callback to what the customer said. The draft is the skeleton; the human touch makes it a reply.
- Is the action clear? Every reply should tell the customer exactly what happens next: what you are doing, what they need to do, or both. Check that the AI draft did not leave this vague.
How I Know It Is Working
Two weeks after the onboarding session I do a quick check. I am not looking for perfect adoption; I am looking for honest use. A few signals tell me the session landed:
- The agent is using AI on a meaningful portion of their tickets, not just the easy ones where they could have typed the reply in 90 seconds anyway.
- Their first-reply time has tightened, even slightly. AI onboarding that works shows up in the queue metrics within two weeks.
- When I read their AI-assisted replies, I can hear their voice in them. The draft is the starting point, not the final product.
At the 30-day mark I do a lightweight CSAT spot-check on their tickets, comparing AI-assisted replies against fully manual ones. The agents who learned the checklist hold up fine. The ones who stopped editing before sending are the ones who need a follow-up coaching conversation, not a lecture about AI, but a conversation about reply quality.
What to Leave Out of the First Session
One mistake I made early: trying to cover too much. Prompt engineering best practices, token limits, model comparisons: all of it can wait.
The first session has one job: get the agent to a point where they can use AI safely on a real ticket by end of day one. Learning in stages, with practice between stages, beats a comprehensive dump that gets forgotten by lunch.
The Underlying Point
I run this session because I believe in the model: AI handling the structure and the boilerplate, human agents handling the judgment and the relationship. That blend, when it works, gives agents the cognitive headroom to be better at the parts of support that actually matter. They are not slower because they are being careful; they have more careful attention to spend on what the customer actually needs.
The 30-minute session is the on-ramp to that blend. Without it, the tool sits unused or gets misused, and neither outcome serves the agent or the customer.
If you are standing up an AI-assisted support team and want to compare notes on how your onboarding is landing, drop me a line. Happy to look at what you have built and point out where the gaps usually show up.
Comments