Jeff Bezos told Amazon shareholders in 2016 that most decisions should be made with around 70 percent of the information you wish you had. Wait for 90 percent, he argued, and you are almost always being too slow. The line gets quoted to death in leadership decks, but product teams ignore the part that actually matters to discovery: he was describing a rule for when to stop gathering information, not a rule for being reckless. The hard skill is knowing which 70 percent is enough and recognizing the moment you have it.
Most discovery I see does the opposite. It keeps running long after the answer has stopped moving. A team books the eleventh interview, commissions another survey, pulls one more cohort, and the findings come back looking like the findings from last week. The research is no longer reducing uncertainty. It is producing the comfortable feeling of being thorough while the real decision sits untouched, waiting for someone to make it.
Discovery is scoped to a decision, not to a topic
The cleanest way to think about how much product discovery is enough comes from framing it as risk reduction rather than understanding. Marty Cagan’s teams at SVPG treat discovery as the work of killing four specific risks before anything gets built: whether customers will value it, whether they can use it, whether engineering can build it, and whether the business can support it. His product risk taxonomy exists precisely so a team can name the risk it is chasing instead of “learning about users” in the abstract. Discovery has a job, and the job is attached to a decision.
That framing carries a quiet test with it. Before you run anything, you should be able to say what decision this research informs and what finding would change that decision. If you cannot name a result that would make you build something different, or build nothing at all, you are not doing discovery. You are collecting reassurance. The interviews will happen, the deck will get made, and the roadmap will come out exactly as it would have with no research at all, because no realistic answer was ever going to move it.
This is where a lot of well-intentioned “customer-centric” teams quietly go wrong. They treat discovery as a standing virtue, something you can always do more of, like exercise. But research has a cost, and the cost is not just calendar time. It is the decision deferred, the competitor who shipped while you interviewed, the team’s momentum bleeding out in another round of synthesis. Douglas Hubbard’s work on measurement makes the economic version of this point: the value of information is what it is worth to reduce the uncertainty around a decision, and once that uncertainty no longer changes what you would do, additional measurement is worth nothing. You are paying for data that cannot pay you back.
Two signals that you are past the point
There are two independent readings that tell you discovery has done its job, and a well-run team watches both.
The first is saturation in the data itself. Qualitative researchers have studied this for decades, and the pattern is consistent: new themes stop appearing surprisingly early. A 2024 secondary analysis in the Journal of Medical Internet Research found that most studies reach near saturation somewhere around the midpoint of data collection, and that for reasonably homogeneous groups, six to eight interviews often carry most of the signal, with well-scoped projects landing between ten and thirty. When your interviews are producing confirmation rather than new information, the curve has flattened. The eleventh conversation is not making the decision safer; it is making you feel safer, which is a different thing and worth less than it feels.
The second signal has nothing to do with the data and everything to do with the decision. How reversible is it? Bezos’s two-way door framing is the right lens here. A decision you can cheaply undo does not deserve the same evidentiary burden as one you cannot. If you can ship a change behind a flag, watch it for a week, and roll it back on a bad reading, the research required to justify trying it is close to nothing, because the experiment itself is the research. Shipping in small, reversible increments is its own form of discovery, which is one reason the big release is the riskiest way to ship. Save the deep, slow, expensive research for the one-way doors: the pricing architecture you will anchor on for years, the platform migration you cannot walk back, the market you are about to move into and cannot quietly exit.
Put the two together and you get a usable rule. Match the depth of research to the reversibility of the decision, and stop when new findings stop arriving. Most teams invert this. They over-research the reversible calls, where the cost of being wrong is a quick rollback, and under-research the irreversible ones, where they are strangely confident. The anxiety that fuels the eleventh interview tends to attach itself to the decisions that scare us, not the ones that actually warrant the certainty.
The reversible decision I studied for six weeks
I did exactly this, expensively, during my years running IT operations at a large telecom. We needed to standardize on an internal ticketing and workflow configuration, and I treated it like a one-way door. We ran weeks of requirements gathering, interviewed team leads across three divisions, built a weighted scoring matrix, and argued about edge cases that affected a handful of tickets a month. Six weeks in, we had a mountain of research and a decision that had, honestly, been obvious by week two. Every additional interview confirmed what the first few had told us.
The thing was reversible. The configuration lived in software we could reconfigure in an afternoon. If we had picked wrong, the cost of switching was a maintenance window and some retraining, not a catastrophe. We had spent six weeks de-risking a two-way door.
The contrast that still stings came the same year. We committed to a vendor platform for a customer-facing capability, a genuine one-way door with multi-year contractual lock-in, and we made that call fast, on a fraction of the diligence, because it felt exciting and the pressure was to move. That one we should have researched into the ground. We had the discovery budget. We had just spent it in the wrong place. Later, doing fractional operations work for smaller companies, I watched the same misallocation play out in miniature constantly: teams agonizing over a reversible copy change while waving through an integration they would be married to for five years.
The lesson was not “do less research.” It was “spend the research where being wrong is permanent.” Thoroughness is not a virtue you apply evenly. It is a budget, and the budget should follow the cost of the mistake.
Name the decision and the kill criteria before you start
The practice that fixes this is almost embarrassingly simple, and it happens before a single interview gets scheduled. Write down three things.
Write the decision. Not the topic. Not “understand how users manage approvals.” The decision: “we will or will not build an approvals inbox this quarter.” A decision has an owner, a deadline, and at least two possible outcomes. A topic has none of those, which is exactly why topics can absorb infinite research without ever resolving.
Write what a changed mind looks like. State, in advance, the finding that would flip the decision. “If fewer than a third of the teams we talk to have ever built a manual workaround for this, we don’t build it.” Naming the kill criterion up front does two things. It protects you from the confirmation reflex, the one that reads every ambiguous interview as a yes, which is the same reflex that makes a validated problem look like validated demand when it is not. And it tells you exactly when you are done: the moment the evidence clears or fails that bar, the research is over, whether that takes four conversations or fourteen.
Write the reversibility. One sentence on what it costs to undo this if you are wrong. That sentence sets your budget. A cheap reversal earns a light touch and a bias to just try it. An expensive, permanent one earns the slow, careful, over-prepared process that the reversible decisions keep stealing.
None of this requires a new framework or a tool. It requires the discipline to treat research as a cost you are spending against a specific risk, and to stop when the risk is retired. The best discovery I have ever seen was not the most thorough. It was the most decisive: enough signal to act, matched to how much the decision could hurt, and then the nerve to close the notebook and make the call while the finding was still worth something. The teams that cannot do that are not more rigorous than everyone else. They are just afraid to decide, and they have found a respectable-looking place to hide.