A simulation is AI-native when AI changes what the game contains, how students experience it, and how instructors run it, and when the designer used AI to build it in the first place. Adding a chatbot to a thirty-year-old engine is a bolt-on, and it does not make a simulation AI-native.
AI can enter a simulation in three ways. As a design tool for the game: we use AI to write the engine, calibrate the market, and test the game with thousands of automated players before any student plays. As features in the game: characters who talk and disagree, a rival who reads the player’s strategy, and feedback written for each student from the game that student played. As the subject of the game: the student decides how a firm should put AI to work, and the game shows the result. The ten titles in the Pedagogix SIMS series use AI in all three ways, and I will draw on them for examples. Building them has taught me what defines an AI-native simulation, so let us look at six characteristics.
Built and tested with AI
The first characteristic concerns how we build and test a simulation. DigiStrat, which I designed with Michael Bean seven years ago, took a year to build and calibrate, because we set every parameter by hand and found every calibration problem through a class. We now take a title in the Pedagogix series from a design document to a working engine in weeks. AI does not make simulation design easy. It makes the mechanical work fast, and the mechanical work was never the differentiating part; it was simply tedious. The real skill lies in knowing which decisions teach and which add irrelevant complexity, what the research says a manager in that situation faces, how a market must behave for a committed strategy to beat a hedged one, and where a game will break when pushed to its limits. Three decades of building simulations and years of working with AI models have honed that judgment, and AI supplies the speed.
Testing is where the design discipline comes in, and the instrument that makes it possible is the test harness. Every title runs as a headless engine, which means the game logic runs without a screen, so we can play it thousands of times a minute. We write automated players with named strategies: one commits to a single way of winning, one tries to be good at everything, one copies the rival, and one chooses at random. We run each of them across every market the game can draw. The harness answers questions a human playtest cannot. Does skill beat luck, or does the market a player inherits decide the result? Does a dominant strategy beat every other strategy? Is there an exploit, a sequence of moves that wins without teaching anything? What is the best result achievable in each market, so that we can grade a student against a real ceiling rather than against classmates who drew an easier market?
In DigiStrat we were still finding bugs years after thousands of students had played the game, because a simulation driven by hundreds of variables has trillions of permutations, and human testers can explore only a few dozen paths. What fascinates me is that AI can now test what AI built, without my involvement. Automated players search the state space at a rate no human tester can match. Players written with AI examine the engine written with AI, and I read the summary. This is a first glimmer of auto-recursion, and it accelerates testing and hardening by orders of magnitude. The design judgment stays human, and the search for what the design does goes to machines that run millions of times faster than a human can.
AI does not make simulation design easy. It makes the mechanical work fast.
The calibration numbers in our catalog come from those runs. In DashKart, the founder’s-seat simulation, the strongest committed strategy captures 73 percent of the best result for its market, and the generalist captures 36 percent. We know those figures because thousands of runs produced them. In the Turnaround airline simulation, the strategy we call Creeping Conventional, which adds one defensible extra at a time to a low-cost airline, goes insolvent in 22 of 24 markets. That figure is measured. When we change a parameter, we rerun the harness before the change reaches a classroom, so a simulation stays calibrated as it evolves. Without this kind of testing, a simulation teaches whatever its untested parameters happen to teach, and the designer finds out from the students.
Characters who interact and adapt
The second characteristic concerns what happens inside the game. Older simulations had a results screen and a market report; ours have AI-generated people. In Agency, the managing partner of a 650-lawyer firm can talk with two of four advisers each period. Each adviser knows something the dashboard does not show, each is paid in a way that colors the advice, and each tells the player more as trust grows. The innovation chief, for example, is paid on AI adoption and overstates readiness. Students practice taking counsel from people who have interests, which a results screen cannot teach.
In QUORUM, five colleagues hold private information the product manager needs, and none of them reports to her. She types or speaks in her own words. Colleagues disclose what they know once she has earned their trust, brought evidence, or asked about what matters to them. If she pushes too hard, they stop talking. If she tells two people different stories, one of them exposes her in front of the group. In EYEWALL, the debrief opens with a named resident of a flood-basin district who testifies to what the cabinet’s decisions meant for her street. Characters who interact and adapt make the learning experience personal and visceral.
Personalized and contextual feedback
The third characteristic concerns the ending. A traditional simulation ends with a ranking: the winning team celebrates, the others feel disappointed, and the instructor draws conclusions from a leaderboard. An AI-native simulation ends with a personal debrief. In EYEWALL, each member of the cabinet receives an after-action review written from that cabinet’s six days: which decisions cost lives, which information a chief held and never put on the table, and what the city would have looked like had the cabinet chosen differently. In MarketMoves, the board review names the strategy the student’s decisions built, scores how consistent those decisions were, and shows the hand the rival was playing. In QUORUM, each student learns which influence style they used under pressure and how they can improve their pitch.
The personal debrief changes both learning and grading, because it assesses what the student actually did. The written deliverable we recommend is a memo based on that debrief, which the instructor grades on reasoning.
Traceable rules and behavior
The fourth characteristic is a design goal we set for ourselves: everything that affects a student’s score must trace to the student’s own decisions, and a language model must never decide the grade. This goal separates an AI-native simulation from a chatbot with a scoreboard. An instructor must be able to defend a grade, and a student must be able to see why a colleague stopped talking or why the plant lost money.
The AI that speaks in our games is grounded in rules and numbers. The advisers speak from rules tied to the state of the player’s own game, and every line traces to a figure on the screen. In DashKart, the three executives read the same numbers and want different things. In YieldForge, the copilot computes the value of every machine hour live, and the plant manager and the CFO brief the player from the same numbers and disagree about what they mean. Where a title does use a language model, as in the press feed in EYEWALL or the colleagues in QUORUM, fixed rules still decide what happens; in QUORUM, rules decide what each colleague discloses and how the player is scored. This design lets an instructor explain and defend every evaluation with confidence.
AI as the subject of the game
The fifth characteristic is new, because an AI-native simulation teaches lessons that did not exist a few years ago. In Agency, the student is the newly elected managing partner of a law firm, facing AI that can do a growing share of legal work. Each period, she decides how much work to hand to AI agents, how to price that work when hourly billing passes every efficiency gain to the client, and how to keep training the junior lawyers who will become partners in ten years. The game encodes what the research found: AI helps on some tasks and hurts on others that look similar, only the errors that are caught enter the log, and the uncaught ones surface two to four periods later. The dashboard looks best when the firm is most exposed. Students practice governing a risk before the evidence for it is available, which is the position of most executives adopting AI today.
In MarketMoves, a retailer can fund four AI investments: personalization, markdown optimization, a retail media network, and content automation. Each pays back only when the business has the customer data and the habits to use it, and the score rewards technology the firm uses rather than technology it buys. In EYEWALL, one chief must judge competing storm models and a weather feed that may have been tampered with. In ACUITY, our hospital simulation, a scheduling assistant offers to allocate operating room time to maximize contribution per hour. It fills the rooms with insured orthopedic cases and starves the emergency reserve and the behavioral unit, because nobody gave it a mission constraint. A lecture can state that lesson about AI; a simulation makes a student live through it.
Zero administrative effort
The sixth characteristic is the least glamorous, and it is the one instructors feel first. The instructor posts a class code. Each student plays one sitting before class, 25 to 75 minutes depending on the title, in a browser with nothing to install; the briefing, the practice round, and the advisers are inside the game. When play ends, the instructor opens the cohort analysis: the distribution of results, the strategies the class built, and the two or three decisions that separated the strong runs from the weak ones. A save code lets the instructor replay any student’s game exactly. Each student receives a personal debrief, and the instructor receives a debrief deck built from the title’s lessons. There are no rounds to run, no teams to chase between sessions, no results to compile by hand, and no ranking to explain. The class session is the debrief, and the instructor walks in with the material.
What this changes for the instructor
Taken together, these six characteristics transform how simulations are designed and used. A traditional simulation takes years to build, may take weeks to play, comes with a hundred-page manual, requires a facilitator, and leaves the instructor to spend hours assembling the debrief. An AI-native simulation is an hour of play with characters who respond, a rival who adapts, a debrief written for each student, a scoring model the instructor can defend, and a cohort view that appears when play ends.
The lessons change too. Students in Agency watch a clean dashboard hide the errors their firm is accumulating, and they come away understanding the verification problem in AI adoption in a way no reading provides. Students in MarketMoves fund four AI systems and see that only the ones backed by customer data pay off. Students in ACUITY watch an optimizer do exactly what it was told and starve the emergency department, and they see why an objective function needs a constraint. These are management lessons for the next decade, and we can teach them by simulation because AI is inside the simulation as tool, as feature, and as subject.
That is my definition of AI-native. We built the game with AI, AI characters speak in it, AI writes the debrief, and AI is what the game is about. The instructor decides what to teach and leads the debrief, and the software does the rest.
Mohanbir Sawhney is Co-founder & Chairman, Pedagogix.