GenAI Simulators

The New Apprenticeship

The New Apprenticeship

The apprenticeship model is one of the oldest learning systems in professional life. A junior person works alongside a senior, watches how decisions get made, handles lower-stakes tasks under supervision, and gradually takes on more complexity. The knowledge that cannot be written down transfers through proximity and repetition. It is slow by design, because judgment does not develop faster than that. AI is breaking this model. Not by replacing the senior, but by removing the rungs on the ladder that let juniors climb. When AI can produce a polished first draft, analyze a dataset, structure a client proposal, the work that used to give junior employees their reps moves to the model. They can submit AI-generated output without developing the judgment to evaluate or improve it. They can produce more while understanding less. And the senior, whose bandwidth was already the bottleneck, has even less reason to slow down and teach. The result is what organizational researchers are starting to call experience starvation: a growing gap between what AI can generate and what the humans working with it actually know. The category filling that gap is GenAI simulators. And the organizations building them now are getting results that traditional training programs have not produced.

The Apprenticeship AI Broke

Traditional apprenticeship worked because it was embedded in real work with calibrated stakes. The junior handled tasks where mistakes were recoverable. They made those mistakes in front of someone who could correct them. Over time, the complexity of what they handled increased, and their judgment developed through accumulated real experience. AI compresses this in a way that looks like efficiency and functions as deprivation. Consider what happens when a junior consultant uses AI to build the first version of a market sizing model. The model looks professional. The client cannot tell the junior couldn’t have built it from scratch. The junior never wrestles with which assumptions matter, what the model is sensitive to, or where it could be dangerously wrong. They have produced output without developing competency. This is not an argument against using AI. It is an argument for designing the practice environments that build judgment alongside the tools that accelerate output. Historically, the apprenticeship created those environments by accident, through the structure of how work got done. AI disrupts that structure. The question is who builds what replaces it.

What the 40 Percent Number Actually Means

Gartner projects that by 2028, 40% of workers will be mentored first by AI rather than humans. That number tends to prompt two reactions. Either it sounds alarming, or it sounds abstract. Both reactions miss the practical implication. The shift is already underway. Every time an organization uses AI to walk a new hire through a difficult scenario, runs a simulated client conversation for sales training, or creates a practice environment for high-stakes decision-making, they are doing AI-first mentorship. Most organizations haven’t named it that. They’ve made a design choice they’re not fully aware of making. The 40% projection is not a warning about a distant future. It is a description of a transition already underway that most leaders haven’t chosen to design deliberately. The organizations getting ahead of it aren’t waiting for a dominant vendor to emerge. They’re building the environments themselves, starting with the roles where the gap between AI-generated output and genuine competency is most costly.

The Category Taking Shape

GenAI simulators are realistic practice environments for high-stakes work, powered by language models. The concept is direct: present the user with a scenario that mirrors real work, have the AI play the other party, provide immediate feedback on what the user does, and let them run it again. What makes a simulator different from using AI as a general-purpose assistant is the design. A simulator has defined personas, calibrated scenarios, and specific feedback criteria. It is not “ask the AI for help with this client situation.” It is “practice this conversation with this type of client until you can read what they need and respond without hesitation.” The category is still forming. No dominant vendor has emerged. Most simulators being built today are created inside organizations, by teams that have identified a high-stakes competency gap and decided to build the practice environment rather than wait for someone else to package it. That is both a limitation and an opening for organizations willing to move.

Where It Is Already Working

Bank of America is using simulators to prepare financial advisors for conversations that carry real weight: clients facing job loss, sudden inheritance, major life transitions. These conversations require advisors to read emotional state accurately, respond with care, and handle complex financial questions without letting the emotional weight derail the practical clarity. Before simulators, preparing advisors for these situations meant role-playing with managers who had limited time and uneven capacity to play a convincingly difficult client. The simulation environment changes the economics. Advisors run through a high-stakes conversation, receive specific feedback on language choices and pacing, and try again. The senior does not need to be in the room for every rep. Hiscox, the specialty insurance company, built simulators for their underwriting certification process. Their results are striking in their specificity: 85% improvement in skill acquisition and 75% reduction in certification failures. Those numbers did not come from redesigning their training philosophy or investing in new L\&D infrastructure. They came from redesigning the practice environment. An 85% skill improvement and 75% fewer failures is not a marginal outcome. It represents a categorically different result than what workshop-based training produces. The mechanism is what makes the difference: practice is the variable, not instruction. People do not develop judgment by being told what to do. They develop it by doing, failing, understanding why, and doing again

What Makes a Simulator Actually Work

Most organizations underestimate the design work required when they first consider building a simulator. The technology is available. The harder problem is knowing what to build. The scenarios have to target the right moments. The highest-stakes situations in a role are not always the most common ones. They are the situations where judgment is the deciding variable, where the difference between a junior and a senior response produces meaningfully different outcomes. Those are the scenarios worth simulating. Practice environments built around low-stakes tasks do not develop judgment. The calibration has to match the actual gap. This requires understanding where junior employees fail specifically, not where they perform poorly on average. What are the precise failure modes that separate novice from competent in this role? A simulator calibrated to a generic version of the job produces generic improvement. One calibrated to the real gaps in judgment produces measurable competency development. The difficulty has to graduate. Effective simulators meet learners where they are and increase challenge as competence builds. Throwing a junior employee into the hardest scenario on day one builds anxiety, not skill. The progression matters because confidence and capability build together. And the feedback has to be immediate and specific. The learning mechanism in simulation is the correction loop: you try, something happens, you understand why, you try again with that understanding. Delayed feedback breaks the loop. Vague feedback makes it useless.

The Facilitation Layer

Here is what the case studies don’t make fully visible: the simulator alone is not sufficient. What Bank of America and Hiscox built were not just practice tools. They built structured learning environments. Someone had to decide when in the development process to use the simulator, how to integrate it into broader work, how to debrief the experience in a way that made learning stick, and how to track whether competency was actually developing over time. That is a facilitation problem. And it is the layer most organizations skip when they decide to build a simulator. AI can generate a realistic scenario. AI can provide immediate feedback on what the user did. What AI cannot do is create the psychological safety to fail and learn, build the structured reflection that connects practice to insight, or calibrate the challenge to the learner’s actual state. Those are human design decisions that require human attention to implement. The pattern Voltage Control has observed across AI adoption broadly holds here specifically. The technology sets the ceiling of what’s possible. The facilitation layer determines whether you reach it. Organizations that build simulators and treat them as self-running tools will see modest results. The ones that design the human layer alongside the technical layer will see results like Hiscox’s numbers.

How to Start

The organizations building effective simulators share a common starting point. They start narrow. One high-stakes role. One specific scenario type. Something small enough to test and iterate before committing to scale. The first version will have design errors that only appear in actual use, and those errors are far easier to fix when the deployment is limited. The scenarios should be designed with senior practitioners, not by L\&D alone. The tacit knowledge that makes a scenario feel real lives in people who have navigated similar situations. A simulator scenario written without that input tends to be technically accurate and experientially hollow. Learners feel the difference immediately. The facilitation structure should be designed before the simulator goes live. Who debriefs the sessions? How often? What are the signals that a learner needs more repetitions before moving to higher-stakes work? How does simulator performance connect to ongoing development conversations? These questions have answers, but they require deliberate design choices before the first session runs. The absence of a dominant vendor in this space is an advantage for organizations willing to move now. The teams building organizational simulator capability today are developing institutional knowledge that later entrants will spend years trying to replicate. They are learning which scenarios matter, how to calibrate difficulty, what feedback actually changes behavior. That knowledge does not transfer by buying a platform.

What This Means for Leaders

Leaders tend to frame AI skills development as a content problem. What should we teach people? Which tools should we cover? How long should the sessions run? These are the wrong questions. The right question is: where does judgment develop in this role, and how do we give people reps against real challenges in conditions where failure is safe? Simulators are one answer. But the insight behind them is transferable to any high-stakes skill development challenge. The goal is not to find a technology. It is to build a practice environment, one that creates the conditions for real judgment to develop alongside the AI tools that are accelerating output. The traditional apprenticeship built those conditions by embedding learning in real work with real stakes under the guidance of someone who knew more. AI has disrupted that structure. The organizations that thrive over the next decade will be the ones that deliberately rebuild it, rather than assuming that AI usage alone produces the judgment that used to develop through experience. The apprenticeship is not gone. It has to be redesigned. GenAI simulators are the clearest model we have for what that redesign looks like. Want to explore what this means for the high-stakes roles in your organization? Reach out to start a conversation about designing practice environments alongside your AI rollout.