Ask a support leader what AI enablement for their team looks like and you get a date, a room, and a vague sense that somebody will demo prompting. Ask what the room is doing ninety minutes in and the answer runs out. That gap is measurable. Zendesk’s research puts it at 72% of CX leaders saying they have given their teams adequate training on generative AI tools against 55% of agents saying they received none.
This is a walkthrough of the workshop itself, for a support org of eight to twenty agents who already have licenses: the intake, the roster, four blocks, and what the team is holding when it ends. Which of the three products sold under this name you should buy, and what each costs, is answered in our comparison of the options for support and success teams. This one is about delivery.
The constraint nobody designs around
Support is the one function that cannot empty the room. Finance can close the books a day late and marketing can ship the campaign on Thursday. A support queue refills while the team is learning, and the backlog waiting at 2pm is why most sessions get cut to a lunch-hour demo and teach nothing. The first design decision is not the curriculum. It is who covers the floor.
Intercom’s support leads wrote up how they handled this in September 2024: formal training blocks set on Monday and Tuesday, team members rotating through in a pattern that kept ticket queues covered, and an explicit decision to accept a short-term hit to response times. That last part is the one most leaders will not say out loud, and the session fails without it.
So never train a support team in one sitting. Two half-cohorts, same facilitator, same material, a week apart; the second inherits the first cohort’s draft conventions and gets to attack them, which produces a better document than either group writes alone. Start after the morning spike, not at 9am, and pick the week with care: a hospitality group’s complaint volume lands on weekends and holidays, which is why AI enablement for hospitality and restaurant groups works around a different calendar than a software team’s.
The intake, a week out
A session on invented examples is a catalog course with a facilitator in front of it. What makes the half day worth more is that every example is yours, and that takes prep work the team does before the room opens. What we ask for:
- Twenty recent tickets spanning your five largest categories, two of which went badly. The failures teach more than the clean ones.
- Your five worst-rated saved replies, pulled from the macro library rather than chosen by a manager.
- The policy pages as agents see them: returns, refunds, warranty, account access. If the live page and the internal wiki disagree, that is the first finding of the day.
- The escalation matrix, however informal. If it lives in one senior agent’s head, say so.
- An administered business account with data terms in place, provisioned before the session. Teaching on personal consumer logins means nothing taught is usable on Monday.
- One senior agent’s last ten replies, as the quality bar the room is aiming at.
If a team cannot produce real tickets, we move the date. Reading scrubbed sample data to a support floor is the fastest way to lose it.
Who should be in the room
Seat choice is a budget decision with published evidence behind it, and the part almost nobody gets right.
Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked a generative AI assistant rolling out across 5,179 customer support agents and published the study in the Quarterly Journal of Economics in 2025. Issues resolved per hour rose 14% on average. Broken out by skill, that average hides everything: 34% for novice and lower-skilled agents, close to nothing for the experienced ones. Retention improved too.
The reading for a roster is direct. Weight the seats toward the newer half of your floor, because that is where both the gain and the churn sit. Put two seniors in the room as the source of examples rather than as students; their replies are the benchmark and their judgment about what never goes near an assistant becomes the rule the group writes down. How that split changes shape across other functions is mapped in our guide to running different curricula for different roles.
Two more seats earn their place. The QA lead, because their scorecard is the language the room will use for “good reply”, and a session that invents a second standard has made things worse. And whoever owns the macro library, because block three creates work that needs their approval.
One seat to leave empty: the VP who will treat the honest part of the session as a performance review. Nobody admits to pasting whole threads into a chat window while their skip-level is listening. Cap each cohort at ten. Past that, three people talk and the rest watch.
Four blocks, and what the room is doing in each
A support half day runs about three and a half hours with one break. Four blocks, in this order, because each one depends on the last.
Block one: everyone answers the same ticket cold
Forty minutes. Pick a real ticket of middling weight, give the room eight minutes and no guidance, and have everyone draft a reply with whatever assistant they have open.
Then read four of them out loud.
The spread is the diagnostic, and the part the team still remembers six weeks later. A typical room produces one agent who never opened the assistant, one who pasted a full thread with a second customer’s order number in it, one who got back a warm paragraph quoting a returns window that changed in March, and one who did it well and cannot say what they did differently. Nobody argues with the agenda after that.
Block two: sounding right and being right
Fifty minutes, and the heart of the session. This is what separates support fluency from finance or marketing fluency.
An assistant is good at tone and unreliable on policy. Those are two separate failure modes and they look alike on the page, because a confidently wrong eligibility rule arrives inside a sentence that reads better than anything the agent would have written. Support is the function where that pairing goes straight to a customer.
The exercise is crude and it works. Take a generated reply, give the room two colors, and mark every sentence as a tone claim or a policy claim. Tone claims you can accept on sight. Policy claims get traced to the page, every time, including the obvious ones. Then repeat it on a reply whose policy half is subtly wrong, and let the room find it.
Solidroad’s survey of 500 support agents names the right problem, even as a vendor’s own number: 82.5% said they felt prepared when they started handling interactions, while 53.5% said the hardest part of ramping was applying training to real customer situations. Agents rarely fail for lack of technique. They fail at the point of transfer, which is why this block uses your tickets rather than a scenario library.
Block three: the macro hour
Forty-five minutes on saved replies, not single answers. A faster one-off reply helps one customer. A rewritten macro helps every ticket in that category after it, which is where the compounding lives.
Pairs, two macros each, policy page open. Output goes to the macro owner as a proposal, never live from the session, because an unreviewed saved reply at scale is a worse problem than the one you started with.
Block four: writing the convention down
Thirty-five minutes, on screen, written by the team rather than handed to them. Three things:
- What may be pasted where, by data category and by account type.
- What never goes to an assistant, put as decisions rather than tools: a refund call, a warranty judgment, a contract reading.
- The plain sentence an agent uses when a customer asks whether a bot wrote the reply.
Expect this block to overrun. The disagreement is almost never about the tool. It is two experienced people finding they have been working to different rules about customer records all year, and the half hour it takes to settle that is the most valuable half hour of the day.
What the team walks out holding
A session that ends in applause and no artifacts has not changed Monday. Six things leave the room.
- One page of paste rules, with your own data categories named.
- A never-delegate list organized by decision, not by feature.
- Between eight and sixteen rewritten macros queued for the macro owner’s review.
- One shared verification habit for policy claims, phrased the same way by everyone.
- The disclosure sentence, agreed and rehearsed before a customer ever asks for it.
- A named owner with two hours a month and the follow-through in their objectives.
The sixth is the one that decides whether the other five survive. Licenses renew whether or not anyone’s work changed, which is the whole argument in why buying licenses is not the same as building capability.
Where these sessions fall apart
Five failure modes, in rough order of how often we see them.
No real material. The tickets never arrive, the facilitator falls back on generic scenarios, and the room concludes, rightly, that it could have watched a video.
Personal accounts. Agents are taught a workflow on a consumer login they are not supposed to use for customer data. Everything taught is unusable the next morning.
Audience inflation. The invite list grows to the whole department plus two VPs. A convention needs a group small enough to argue.
Wrong week. A retail or e-commerce support queue cannot absorb a training day in November, which is why peak season picks the date for retail operations teams rather than the L&D calendar.
Metrics left alone. Intercom’s write-up makes the point: as automation absorbs the simpler tickets, agents spend their day on slower work, so scoring them mainly on resolution speed punishes the behavior the session was meant to produce. If handle time is still the only number on the wall in week three, the training is over.
Why 27 catalog hours cannot replace a half day on your own queue
Two formats dominate AI training for customer support teams, and the hour counts show what each one buys.
| Format | Time | Price as published | What it is good at |
|---|---|---|---|
| IBM’s Generative AI for Customer Support on Coursera | Three courses, 27 hours, stated as four weeks at ten hours a week | Included with Coursera Plus | Shared vocabulary, prompting basics, a certificate |
| NobleProg’s prompt engineering course for support and chatbots | 14 hours, live, onsite or remote | From £3,200 plus VAT per private group | Instructor contact, though much of the outline is chatbot building rather than agent fluency |
| Facilitated half day on your own tickets | Three to four hours per cohort | $2,000 to $15,000 market range | A written convention the team keeps using |
Prices and outlines were checked against the vendors’ public pages in October 2026; catalogs change quarterly, so recheck before quoting anything internally.
The difference is not quality. It is whose examples are on screen. A course cannot know your returns window, your escalation matrix, or that your warranty macro has been wrong since spring.
The sixty days that decide whether it took
Baseline before anyone is trained. A rough number taken up front beats a precise one rebuilt after the fact.
- Reply quality on twenty tickets, scored by a senior agent against the QA rubric that was in the room. Same rubric, same scorer, same sample at day sixty.
- Macro movement. How many saved replies and help-center articles were revised since. Zero means the compounding work was never picked up.
- One consistency check. Ask three agents, one at a time, what may be pasted where. Three answers mean the convention needs a thirty-minute reset, not a new session.
- A ten-reply audit. Trace every policy claim in ten AI-assisted replies to its source.
Leave out self-reported confidence and completion rates. Both rise without any work changing.
Frequently asked questions
How long should an AI session for a support team run?
Three to three and a half hours per cohort, so two mornings for a floor of sixteen. A full day on one group empties the queue and loses the room after lunch.
Can we train the whole support team at once?
Only if you will close the queue, which almost nobody will. Split into two cohorts a week apart and rotate coverage. The second group improves the first group’s conventions rather than repeating the exercise.
What do we need to prepare beforehand?
Twenty real tickets across your main categories, your five worst macros, the policy pages agents open, the escalation matrix, and a provisioned business account with data terms. About two hours of someone’s time, and the session is worth much less without it.
Should an internal champion run it or an outside facilitator?
An internal champion is better on fit and worse at getting two senior agents to settle a disagreement in front of their team. Intercom’s experience with peer-led sessions was that enthusiasm travels well; the refereeing is the part that gains from someone with no stake in the org chart.
Do managers belong in the room?
The QA lead and the macro owner, yes, because the session produces work that is theirs. Skip-level executives, no. Agents will not own up to what they have been pasting into a chat window if the person who writes their review is watching.
Does this work over video?
Well enough for the cold draft and the macro hour, less well for writing the convention, where the useful argument happens. If it has to be remote, cap it at six people and use breakout pairs for block three.
When in onboarding should a new agent get this?
After two or three weeks on live tickets, not during week one. The exercises depend on knowing your own categories, and the study of 5,179 agents puts the gain with newer agents, so schedule it rather than waiting for the next annual session.
What if our tickets contain data we cannot hand to a facilitator?
Standard practice is that the team redacts names and account numbers and keeps the tickets on their own screens, with the facilitator working from category and structure. Nothing in these four blocks requires an outsider to hold your customer records.
How is this different from an AI role-play simulator?
A simulator trains the conversation: an AI plays an upset customer and scores the agent on handling. This session trains the tool: what to paste, what to verify, what never to delegate, which saved replies to fix. Teams that buy one often still need the other.
Key takeaways
- Design around the floor first: two cohorts a week apart, after the morning spike, with an accepted short-term hit to response times.
- The intake is the product. Twenty real tickets, five bad macros, the live policy pages, a provisioned business account. Without them you have bought a catalog course with a human reading it.
- Weight the seats toward newer agents. The study of 5,179 support agents found 34% more issues resolved per hour for novices and close to zero for the experienced; seniors belong in the room as the source of examples.
- Block two is the one that matters: separating tone claims from policy claims, then tracing every policy claim to the page. A warm paragraph with a wrong refund rule is this function’s characteristic failure.
- Judge it on artifacts, not applause, and give the six that leave the room a named owner with hours in their objectives.
If you are mapping out what a session for your support floor should cover, book a free AI readiness call. Bring your ticket categories and your macro library and we will tell you which of these four blocks your team needs most.
Arthur Wandzel