To train customer-facing teams with AI, split every skill by where it's performed: conversations get AI roleplay, software workflows get supervised practice in the real product, and each rep gets one blended agenda across both. Teams that use a single mechanism for everything train half the job and wonder why the other half never improves.
Step 1: Inventory skills by habitat
List what your support reps, CSMs, and sellers actually do in a week, then sort each item by performance context. Conversational: discovery, objection handling, renewal negotiations, difficult-customer de-escalation. Procedural: ticket triage in the helpdesk, account setup in the admin panel, quote building, report pulling. Hybrid items split cleanly on inspection: "run a QBR" is a conversation plus a dashboard workflow, train each half in its habitat. This sorting step is the whole method; everything after is execution.
Step 2: Assign the mechanism per habitat
Conversations: AI roleplay with scoring rubrics. The category matured fast, Synthesia's July 2026 move from training video into live avatar roleplay sessions is the marker, and for dialogue skills the practice-performance match is right. Procedures: a live training agent in the product, demonstrating on the rep's screen, then supervising attempts in practice mode on realistic cases. Keep the boundary honest in both directions: rehearsing with an avatar practices the conversation; practicing in the product builds the skill. Neither substitutes for the other, and pretending otherwise is how teams end up with reps who can pitch beautifully and can't build the quote.
Step 3: Build one agenda per person, not two programs
The failure mode of blended training is two disconnected tools with two dashboards. Run it instead as one curriculum per rep: this week, the de-escalation roleplay and the refund workflow; next week, the exception path and the harder objection. Short sessions, spaced, resumable, because skill consolidates through repetition over days, not marathon workshops. An agenda-driven training agent holds this structure natively: it knows what each rep has demonstrated, what's next, and what needs a refresher.
Step 4: Verify by performance, on both halves
Completion tells you nothing here. Certify conversations by scored roleplay against your rubric; certify procedures by workflows performed unassisted in the product, witnessed and recorded. Both become per-rep facts a manager can act on: who's ready for the queue, who needs two more reps of the billing flow, which team never mastered the new process.
Step 5: Feed real work back into the curriculum
Your escalations, lost-deal notes, and repeat tickets are a free needs analysis: every recurring failure names the next scenario for roleplay or practice mode. Close the loop monthly. (Disclosure: we build Skippr; Skippr AI training handles the in-product half, live sessions, practice mode, agendas per learner, users trained as the metric, and it plays fine alongside roleplay tools for the conversational half.)
Questions buyers actually ask
Can one platform do both roleplay and product practice?
The mechanisms differ, avatar conversation versus screen-aware supervision in your product, and most vendors are strong at one. Buy for each habitat and insist both feed one per-rep record.
How long should sessions be?
Short and spaced beats long and rare: ten to twenty minutes, two or three times a week, resuming where the last left off. That cadence fits queues and calendars, and it matches how procedural skill consolidates.
Watch it train someone
Completion is not competence, and the difference shows up in the product rather than the dashboard. Fifteen minutes is enough to tell them apart.