Trust and quality notes
- Last updated
- October 12, 2026
Planning for AI can fail even when the original insight was reasonable. Capabilities, costs, providers, and market behavior can change while a product is still being built. Dan Pupius argues that the response is not better prophecy. It is a clearer understanding of which strategic bets can be reversed and which ones lock a company into a path.[1]
His view of humanity after automation is grounded in adaptability. People will still make judgments under uncertainty, but they need strategies that reveal their assumptions and make it possible to change course before outside pressure leaves no choice.
The source's central argument
Pupius draws on experience rebuilding Gmail at Google and leading engineering at Medium. In those cases, teams began with a view of how behavior, technology, and markets were moving, then built their bets on top of it. AI makes that approach harder because the underlying assumptions can shift during execution.[1]
He proposes replacing one broad question, what do we think will happen, with two more useful ones. First, which assumptions are irreversible? Second, which assumptions can still be revised?
His example of an irreversible assumption is building deeply on one model provider. Moving away later may require rewiring the product. His example of a reversible assumption is expecting token costs to remain high. If they fall, a team may be able to change its approach without rebuilding the system.[1]
Pupius has developed a framework around these distinctions. Its purpose is to move discussion away from loose labels such as AI wrapper and toward the concrete futures a business is betting on. The goal is to notice sooner when an assumption is wrong and move, not to predict the future with greater precision.[1]
Every strategy contains a theory of the future
A product decision can look purely technical while carrying a forecast inside it. Choosing a provider may assume that its interface, economics, and capabilities remain suitable. Designing around scarce model use may assume inference remains expensive. Building a specialized feature may assume users continue to value that feature after general models improve.
Pupius's contribution is to make those hidden forecasts discussable. Once an assumption is named, a team can ask what happens if it fails and how much of the product would need to change.
This is particularly useful in AI because change does not arrive only as steady improvement. A new capability can remove a constraint that shaped the workflow. A price change can make an uneconomic approach practical. A provider decision can alter access or terms. The source does not list all of these possibilities, but its reasoning is built around assumptions changing during execution.
Reversibility is about the cost of being wrong
A reversible bet can be updated without destroying the surrounding work. An irreversible bet ties so much of the product to the assumption that changing it requires substantial reconstruction.
This does not mean irreversible decisions are always bad. Some differentiation may require commitment. Pupius's framework is intended to clarify the nature of a bet, not eliminate every dependency. The danger is entering a hard-to-reverse position without recognizing it.
My interpretation is that teams should distinguish technical reversibility from organizational reversibility. A component might be replaceable in code while contracts, training, processes, or customer promises make the switch painful. Pupius does not state this distinction in the source, so it should be treated as an extension of his core idea.
A second interpretation is that reversibility has a time dimension. A choice may be easy to change now and difficult after more data, features, and habits accumulate around it. Recording the assumption at the moment of decision gives future reviewers a clearer signal for when to reconsider.
What is argument and what is interpretation
Pupius explicitly argues that AI's rapid change can undermine strategy while it is being executed. He asks founders to identify irreversible and revisable assumptions and to focus on the specific futures behind their choices. Success means recognizing error early and moving, not forecasting perfectly.[1]
My practical interpretation is to maintain a small assumption register for consequential AI decisions. For each one, write the belief, the evidence available now, the choice it supports, the cost of reversal, and a signal that would trigger review. This is not a framework reproduced from the source. It is a simple operating method derived from his distinction.
Another interpretation is that debates about whether a product is merely a wrapper are too coarse to guide action. Two products that use the same external model can carry very different risks. One may be tightly coupled to a provider's behavior and economics. The other may preserve options through its architecture and workflow. Pupius explicitly rejects the broad label, while this comparison illustrates why that rejection is useful.
A practical reversibility review
Begin with a current plan and underline every statement about what the future will be like. Some assumptions will concern model capability, others cost, user behavior, regulation, or access to data. Do not debate whether each is true yet. First make the list visible.
Next, connect each assumption to a decision. If the belief fails, what must change? A configuration adjustment is different from rebuilding the product's central workflow. Rate the consequence in concrete terms rather than calling the decision flexible or inflexible.
Then reduce unnecessary lock-in. Keep interfaces clear, preserve data in usable forms, and avoid letting one provider's peculiar behavior define more of the product than the value requires. These are my design suggestions, not prescriptions stated by Pupius. In some situations, deep integration may be worth the risk, but that should be a deliberate trade.
Set review signals. A team cannot monitor every possibility continuously. Choose evidence that directly challenges the assumption, such as a capability becoming broadly available or a cost constraint disappearing. When the signal arrives, revisit the decision rather than defending it because work has already been invested.
Finally, keep the framework tied to action. An exhaustive list of uncertainties can become an excuse not to build. Pupius's goal is to help people move sooner when wrong. The point is flexible commitment, not permanent hesitation.
Human judgment shifts from certainty to correction
Pupius offers a sober role for people in an unstable technological period. We still need a view of the future in order to choose. But our value does not depend on making that view infallible. It depends on seeing the assumptions beneath it and designing enough freedom to respond.
After automation, strategy may become less like selecting one destination and more like managing the cost of changing routes. That sentence is my interpretation, but it stays close to the source's emphasis on reversible bets. The organization that learns quickly is helped by architecture and decisions that let learning matter.
For each AI bet, do not ask only whether it looks right today. Ask what must be rebuilt if it is wrong tomorrow, and decide whether that commitment is justified.
If you want to introduce AI without hiding the assumptions underneath it, Agentic Workers can help you design a workflow with clear boundaries and room to change.