Stop Building Courses. Start Building Rehearsals.

AI practice for frontline skills, and what L&D should ask before buying it.

OPINION PIECE

Damian Hehir, Co-founder, Capytech

10/7/20264 min read

Over the last few months I've written here about AI being table stakes in learning, what an AI-Native LMS actually does for L&D, and the lock-in risk that comes with AI authoring. All three were about content and platforms. This one is about what learners actually do once the content exists.

Because for a lot of frontline skills, the content was never really the problem.

The gap between knowing and doing

Most customer service and compliance modules end the same way. A scenario on screen, an angry customer in a speech bubble, four options. Pick the right one. Most learners pick correctly.

Then a real customer raises their voice at the counter on a busy afternoon, and very little of that module comes back.

A completion rate tells you someone attended. A multiple-choice question tells you they can recognise the right answer when it's written out in front of them. Neither tells you whether they can handle the conversation. And for customer-facing roles, the conversation is the whole job.

What has changed?

Roleplay has always worked. Anyone who has run a good facilitated session knows that. The problem was never the method, it was the scale.

A facilitator can roleplay a difficult conversation with each person once, maybe twice, usually with the rest of the room watching. Plenty of people hate that, so they play it safe and learn very little. And once the session ends, there's no more practice until the next one.

AI changes the economics of this. You script the setup: who the character is, what their problem is, how they're feeling, what a good outcome looks like. Then the conversation runs dynamically. The AI customer responds to what the learner actually said, not to option B. At the end, the learner gets feedback against the criteria you defined, not just a pass or fail.

Every learner can run it as many times as they need, in private, at 10pm if that's when they have the time.

I'm not neutral here. We built this into our own CapyCoach Scenario Builder, and we've also put it inside a virtual 3D space: you're the shop manager, every customer walking in is AI, each one arrives with a problem, and an AI instructor grades how you handled it. But the principle matters more than any one product. Practice used to be the expensive part of learning. It no longer has to be.

Where it works, and where it doesn't yet

It works best where the gap between knowing and doing is the problem:

· Complaint handling and customer service

· Retail floor and hospitality conversations

· Sales conversations and objection handling

· Challenging unsafe behaviour on site

· New managers having their first difficult conversations about performance or attendance

In this region, it also matters that learners can practise in the language they'll actually use with customers, Arabic or English.

Where I'd be more careful:

· High-stakes judgement calls where a mistake has legal or safety consequences. Use AI as rehearsal, but keep a human sign-off before anyone is cleared.

· Skills that are physical. You can't roleplay your way to a safe forklift manoeuvre.

· Cultural nuance. AI characters can drift towards a fairly generic customer. Test the scenarios with your own people before you roll them out, and adjust the characters until they sound like the customers your staff actually meet.

What to ask before you buy

A few questions worth putting to any vendor, including us:

1. Who writes the scoring criteria? You should define what good looks like, and the feedback should explain itself. If the tool gives a learner 6 out of 10 and can't tell them why, it's a toy.

2. How does it handle Arabic? Test it yourself, including a learner switching between Arabic and English mid-conversation, which is how a lot of people here actually talk.

3. Does the character stay in role? Try to break it. Learners will.

4. Where do the transcripts go, and who can read them? People practise far more honestly when they know it isn't going straight into their HR file. Decide that policy before launch, not after. And ask the same data question I raised in my last piece: is your content and learner data used to train anyone's model?

5. Does it sit inside your LMS and report there, or is it yet another login and another dashboard nobody checks?

How I'd start

Pick one high-volume frontline skill where you already know there's a gap. Complaint handling is usually a good candidate.

Write three scenarios with your best performers, not with a vendor. They know what the difficult customers actually say.

Run it for six to eight weeks with one team, and measure something that happens on the job: complaint escalations, mystery shopper scores, a manager's observation checklist. Not completions.

If it moves, scale it. If it doesn't, you've learnt something useful for very little cost.

Where this leaves you

AI has made content cheap to produce. That's what most of the conversation has been about. The more interesting part is that it has made practice cheap too.

The organisations that get the most out of this won't be the ones with the most courses. They'll be the ones whose people have rehearsed the hard conversations before they have them for real.

Damian Hehir is Co-founder of Capytech, an L&D solutions company headquartered in Dubai. Capytech holds the Dubai AI Seal and works with public and private sector clients across the GCC on e-learning, AI-native LMS deployments and bilingual content development.

Damian's profile: linkedin.com/in/damianhehir