BIG PICTURE
The Pizza Doesn't Get Exponentially Better
"Put a data center of PhDs inside Domino's. Pizza delivery does not get exponentially better."
Anish Acharya, a16z partner and serial entrepreneur, grabbed my attention with this statement on a recent episode of Lenny's podcast. It raises a useful question for anyone redesigning work around AI: what actually limits the outcome?
Most AI transformation starts by looking for tasks where added intelligence can make the work faster, cheaper, or better. But many business outcomes are constrained by something other than intelligence. Relationships need to be trusted. Decisions need to be coordinated. Physical goods still need to move through the world.
More intelligence may make some steps faster or cheaper; limits are gaps that prevent intelligence from affecting the outcome. Other limits are ceilings that intelligence fails to raise. Let's look at both.
Gaps - when intelligence is not enough
In The Art of Action, Stephen Bungay explains why even sophisticated military plans break down when they meet reality. He identifies three recurring gaps:
- The knowledge gap between what planners would like to know and what they actually know
- The alignment gap between what leaders want people to do and what people actually do
- The effects gap between what an action is expected to accomplish and what it ultimately produces
These gaps also explain why impressive AI capabilities can generate disappointing organizational results.
Start with the knowledge gap. Chatterji and colleagues' 2026 paper offers a useful distinction between the cost of work and its value. Wages, headcount, and cycle time data already exist in owned systems, but what a task is actually worth usually does not. That requires a judgment about customers, strategy, risk, and what the organization is trying to accomplish.
Leaders may know what a task costs without knowing what improving it is worth. AI can make the task more efficient while leaving that more important question unanswered.
The alignment gap appears when people and systems do not share the same understanding of the goal. Bungay’s answer is to communicate strategic intent to decision-makers at the lowest levels possible. Strategic intent is a clear statement of what must be achieved and why it matters.
An organization’s survivability may depend less on deploying the most powerful model than on having true alignment around a response to the AI era. What is the organization trying to achieve? Why does it matter? Who can make which decisions? Where must human approval remain? Without those answers, intelligence has no stable direction.
Then there is the effects gap. This is how an AI pilot can produce real gains for the person doing the work and nothing discernible on the P&L.
Hollister and colleagues describe tasks as parts of a relational system rather than isolated units of work. Changing how one task gets performed can alter upstream inputs, downstream handoffs, and supervision. Removing work from one employee may save that person time while creating new review work for someone else. Replacing a worker with an agent may eliminate a task while weakening the apprenticeship through which junior employees become experts.
The local gain is real. So are the second-order costs.
This is why Sangeet Paul Choudary argues in Reshuffle that “the real potential of AI lies not in automated tasks but in reconfiguring how systems operate.”
Ceilings - when intelligence is the wrong input
Some work does have extraordinary upside from greater intelligence. Acharya uses drug discovery as an example, where a marginal improvement in identifying a viable drug could produce an enormous outcome. In that setting, paying for peak intelligence is rational.
Other work has bounded upside. A legal contract can be correct, clear, and fit for purpose, but it cannot become infinitely better. A company can close its books accurately and quickly, but it cannot close them 100 times more accurately. Once the work clears the required standard, the economic goal shifts from maximum intelligence to reliable, cost-effective performance.
Acharya extends this argument to AI “loops”: systems of agents that repeatedly perform and improve a set of tasks. A loop may climb toward a local maximum, steadily getting better at the work it's been given, and then plateau. More iterations help it climb the same hill, but they don't reveal which hill it should climb next.
Acharya makes a compelling argument that human intuition is needed to help the system “land at the base of the next hill.” But we know that human intuition isn't automatically better, especially in unfamiliar territory. (Daniel Kahneman and Gary Klein's famous work on skilled intuition explains why.)
We shouldn't ignore human intuition, but a more defensible human contribution is taste: choosing which questions deserve attention, which possibilities are worth exploring, and which outcomes justify the investment.
Taste is a fuzzy word, but I see it as another form of judgment. Judgment about value under uncertainty, followed by a willingness to own the bet that can sometimes send a self-reinforcing signal to others.
Aim intelligence at the constraint
Gaps and ceilings inherent to work outcomes teach us three important lessons about how we can capture more value from AI:
1/ Before adding more intelligence, ask what currently limits the outcome.
AI creates value when leaders point it at a constraint that intelligence can loosen. Is the organization missing knowledge about what matters? Is intent unclear? Are decision rights preventing adaptation? Are local gains being absorbed by coordination costs elsewhere? Or has the work already reached a ceiling beyond which additional intelligence has little economic value?
2/ Describe human contributions more precisely.
People do not earn a permanent role merely by being human, nor can leaders assume that human judgment will beat a model’s. Their distinctive contribution is to define intent, assign decision rights, recognize changes in the environment, choose worthwhile problems, and remain accountable for the consequences.
3/ Divide work according to those contributions.
Early evidence from agentic coding shows what this division can look like in practice. In Anthropic’s analysis of Claude Code sessions, people made about 70% of planning decisions—what to do, which approach to take, and what counted as finished—while Claude made about 80% of execution decisions.
This is just one product and one classification method, so it's not a universal formula, but it illustrates a productive arrangement where people frame and judge the work while AI carries more of the execution. In general, leaders first need to identify what the outcome requires and what constrains it, then assign people and AI based on what each contributes.
Which brings us back to Domino’s. The question was never how smart the kitchen could become. It was what stood between the kitchen and getting hot pizza to your door.