The answer lives in this podcast
What is a 'model garden' — and why does Katrina Mulligan consider it a mistake in AI adoption strategy?
A model garden is the practice of making 12 to 20 different AI models available in a single place so employees can pick among them. Mulligan argues it is expensive, creates unnecessary friction for ordinary users, and delivers no material benefit — noting that a typical HR employee does not need 12 choices when typing a query into a chatbot. She predicts organizations will look back on model gardens as a poor use of effort and resources.
A model garden is an internal AI deployment approach where an organization grants employees access to a curated selection of 12 to 20 different AI models within a single interface or platform. The idea is to offer choice. Mulligan's argument is that, for the vast majority of knowledge workers, that choice is noise, not value.
The logic behind model gardens sounds reasonable on paper: give people options, let them find what works best for their task. In practice, as Mulligan explains in 3 Takeaways™, the friction of choosing between models serves specialists far better than it serves the general workforce.
Most employees at any organization — in HR, finance, communications, operations — are asking routine questions and drafting routine documents. For them, the cognitive overhead of evaluating model differences is a barrier, not a feature. Mulligan's point is blunt: the cost of maintaining that range of options is real, the benefit for most users is not.
Choice without context is just friction — especially for non-technical users
The HR example Mulligan uses is telling. Someone typing a policy question into a chatbot does not have the knowledge to evaluate which underlying model will handle it better. They want an answer. Presenting them with a menu of 12 AI engines before they can get one is a design failure disguised as a feature.
This connects to a broader tension in enterprise AI rollouts: the impulse to offer maximum optionality often reflects the preferences of technical teams rather than the reality of end-user workflows. As Mulligan discusses across this episode, successful AI adoption is fundamentally about reducing friction, not multiplying it.
The financial dimension matters too. Running and maintaining access to 12 to 20 models simultaneously is not cheap. Each model requires infrastructure, access management, and ongoing evaluation. When the incremental benefit for the average user is near zero, that cost becomes hard to justify — and Mulligan expects that, in hindsight, many organizations will agree.
A prediction: model gardens will be remembered as an early-adoption mistake
Mulligan does not frame this as a minor inefficiency. She frames it as one of the organizational habits that will age poorly — a pattern that made sense in a moment of rapid experimentation but will look misguided once AI deployment matures.
Her position, laid out in the episode, is that the right question for most organizations is not "which models should we offer?" but "how do we get the most people using AI effectively, right now?" Those are different questions with different answers — and the model garden approach is optimizing for the wrong one.
"I honestly think that I am at least 30% more effective, maybe more than I was a year ago, because of how I've matured my use of these tools. I'm a better leader. I'm a better manager."
Katrina Mulligan — Head of National Security Partnerships, OpenAI.
Mulligan spent years at the apex of U.S. national security structures, holding senior roles at the Department of Defense, the National Security Council, and the Department of Justice. She served as number two at the Pentagon overseeing special operations — one of the most demanding and hierarchical environments in American government. Her move to OpenAI, which she describes as the hardest professional transition of her career, placed her at the opposite extreme: a fast-moving, flat organization where, as she puts it, a month feels like a quarter anywhere else. That dual vantage point — inside both Washington's security apparatus and frontier AI development — gives her an unusually grounded perspective on what AI can and cannot do inside large institutions, and what gets in the way of getting there.
Mulligan's personal productivity gain — at least 30% more effective by her own estimate — underlines the stakes. The goal of AI adoption is not to offer a marketplace of models. It is to get more people to the point where AI genuinely changes how they work. Unnecessary complexity is the enemy of that outcome, a point she makes compellingly in this conversation on 3 Takeaways™.
See also
What is the single biggest predictor of a successful AI transformation in an organization according to OpenAI's implementation experience?
Mulligan says the biggest predictor is the extent to which the C-suite personally uses AI in their own work. A bottom-up approach of simply giving everyone access is not enough without visible leadership adoption driving it from the top.
Beyond saving time, where is the biggest opportunity for AI in healthcare and scientific discovery?
Mulligan cites OpenAI's partnership with a children's hospital, where frontier reasoning models were run against unsolved pediatric cases — patients who had gone years without a diagnosis — and the models surfaced answers that had eluded clinicians entirely.
What was the state of ChatGPT when it first launched and what changed to make modern AI so much more capable?
Mulligan notes that when ChatGPT was first rolled out there were only 200 people at OpenAI, and the initial model produced outputs far less reliable than today. The breakthrough was OpenAI's discovery of scaling laws: multiplying compute and data by 10x each produces predictably better model performance.
Key takeaways
- A model garden makes 12 to 20 AI models available in one place — an approach Mulligan says is costly and adds friction with no meaningful payoff for most employees.
- The typical knowledge worker — an HR employee, a finance analyst — does not benefit from choosing between AI models; they benefit from getting a reliable answer fast.
- Maintaining that range of model options carries real infrastructure and management costs that are difficult to justify when user-level benefit is negligible.
- Mulligan predicts organizations will look back on model gardens as a poor allocation of effort — an early-adoption habit that optimized for optionality rather than outcomes.