News
8 min read
Last week I pulled every case study we’ve produced over the past two quarters into a single file. I wanted a clean list of what customers actually use us for. I expected a taxonomy. Sales training in one bucket, clinical skills in another, compliance somewhere near the bottom where compliance always goes.
Ninety-two entries later, I had something else.
Here is a sample of what came out. A federal bank examiner rehearsing how to tell a community bank CEO that his institution has a serious problem, while the board chair sitting next to him pushes back hard. A nurse practicing how to redirect a patient with dementia without escalating the situation. A soldier making a cold prospecting call. A retail associate learning to read a customer in a jewelry store on a Saturday in December. A sponsored student telling his academic advisor that his grades have slipped and he needs help before it’s too late. A field technician walking up to a house he has already visited twice for the same broadband fault.
Nothing about those six people is alike. Different countries, different regulators, different budgets, different words for the thing they do all day. But they bought the same product, and as far as I can tell they bought it for exactly the same reason.
Each of them has a conversation coming that they cannot afford to get wrong, and no realistic way to practice it first.
That’s the whole business. It took me building a 92-row spreadsheet to see it clearly.
Why Realistic Practice Has Always Been Too Expensive to Scale
For most of industrial history we solved this with apprenticeship. You sat next to someone who had done the job for twenty years, you watched, you tried, they corrected you, and eventually you were the one being watched. It worked. It also had a hard ceiling, because it consumed the scarcest resource in any organization: the attention of your best people.
That ceiling is where every training budget goes to die. A live role play with a skilled facilitator runs somewhere around $200 a session once you count the trainer, the room, and the time your people are not doing their jobs. At that price you get one shot. Maybe two if the program is well funded. Nobody gets good at anything in two attempts.
So organizations did the rational thing and stopped trying. They replaced practice with information. Slide decks, policy documents, an annual module you click through while eating lunch. Everyone involved understood this was a substitute for the real thing, and everyone kept the arrangement going because the real thing did not scale.
The AI Roleplay Training Cost That Changes Everything
An AI rehearsal session costs under $5.
I want to sit on that for a second, because the interesting thing about a cost curve like that is not the saving. It’s what becomes possible on the other side of it.
A federal healthcare system ran the comparison properly and published it through a peer-reviewed conference, which is rarer than it should be in our industry. In-person simulation cost them $130 per learner. Doing it on our platform cost $11.29. That’s a 91% reduction, and the study found no drop in educational quality. Thirty-nine clinical staff completed the training over thirty days without a single disruption to ward operations, because nobody had to be pulled off the floor at the same time.
The money is nice. The second finding is the one that matters. When practice costs almost nothing and can happen at 6am or on a Sunday, you stop rationing it. And once you stop rationing it, the entire logic of workforce development inverts. You are no longer deciding who is worth training. You are deciding what is worth practicing.
High Stakes Scenarios Traditional Training Could Never Serve
Once repetition is basically free, the highest value scenarios turn out to be the ones organizations had quietly given up on.
Think about what a bank resolution team does in the first thirty minutes after walking into a failed bank on a Friday evening. Securing records, talking to staff who just learned they may not have jobs, beginning customer triage, all under time pressure. It’s one of the most consequential half hours in the whole of financial regulation, and almost nobody in the building has ever done it. You cannot run a drill for that with real depositors. You cannot wait for the next bank failure to build the muscle.
Or take a wealth management firm we work with. Their advisors asked to practice giving financial advice to a client going through a divorce, and having a conversation with a family when a client of twenty years becomes terminally ill. We did not propose those scenarios. Their own learning team built them, because those were the conversations their advisors were actually losing sleep over.
Same pattern in defense, where the proposed scenarios include bystander intervention before an assault happens and recognizing when a colleague is in crisis. Same pattern in education, where a student practices refusing a friend who is offering him a very bad idea.
Low frequency, high consequence, no safe way to fail. That is the sweet spot, and it is the exact category traditional training was structurally incapable of serving.
Does AI Roleplay Training Actually Improve Real World Performance?
Fair question, and I’d rather answer it with data than adjectives.
A military recruiting program ran hundreds of people through three cohorts and logged more than a thousand practice sessions. The team ran a proper regression on it. Platform usage explained 84.7% of the variance in how those recruits performed in their real assessed skills lab. Session quality was the single strongest predictor they measured. Ninety-five percent reported higher confidence in real conversations afterward.
A global health and hygiene company ran a pilot with sales reps across the UK, France and the Middle East. Those learners completed thousands of sessions across several technical medical product use cases, an average of 32.5 practice runs each, voluntarily, on top of their day jobs. Average scores moved from 56.2% on first attempt to 66.8% on best attempt.
My favorite number in the entire file came out of that pilot, though, and it isn’t a percentage. One rep in the Middle East counted his filler words across attempts. He went from 14 to 5. Then he told the debrief that he felt more professional in front of real customers.
That’s the actual product. Everything else is scaffolding.
How Customers Are Building Their Own AI Roleplay Training Capability
I assumed we would be building most of this content. We’re not, and the customers who get the most out of the platform are almost always the ones who took the keys themselves.
One telecom customer started with a single compliance use case. Within a year they were running five departments on it, including something we never would have thought to sell them: virtual simulation of high risk field work like ladder safety, gas detection and bucket truck operations. Their own operations team built it from their own incident data. They now have an internal capability that normally takes an enterprise two or three years to develop.
That reframes what we are. We are not a training vendor with a catalog. We are closer to infrastructure, and the organizations that treat us that way get results the ones waiting for a content library never do.
Who Owns Conversation Performance in Your Organisation?
There’s a question hiding inside all of this that most companies have not asked yet: who owns the conversation?
Somebody owns your uptime. Somebody owns your defect rate, your pipeline conversion, your safety record. Every one of those has a named executive, a dashboard and a budget. Yet the quality of the ten thousand human conversations your organization has this week, the ones that actually determine whether the customer stays, whether the patient calms down, whether the finding sticks, whether the recruit signs, belongs to nobody in particular.
It was unmeasurable, so it went unmanaged. That’s no longer true. Every rehearsal now produces a record of how someone communicated, decided and improved. Capability becomes something you can look at instead of something you assume.
Ninety-two use cases across defense, healthcare, banking, retail, telecom, pharma, education and professional services. One underlying job.
Give people the reps they were never able to get, before the moment that matters.
Kurt Kratchman – connect with me!
