The phase model is comfortable. First we discover, then we define, then we build, then we launch. Each stage has a gate, each gate has a document, and at the end of the sequence there is a product.
The model is also wrong about how uncertainty behaves.
Uncertainty does not finish on schedule
In the phase model, discovery ends when delivery begins. The implicit claim is that the team has learned what it needs to know, and the remaining work is execution.
In practice, certainty tends to peak at the moment a plan is approved and decline as soon as real constraints arrive. The build reveals that the data model cannot support the workflow. The first release reveals that people use the feature differently from the prototype. A competitor changes the context. A partner changes the terms.
None of that shows up in the discovery phase, because the discovery phase is over. The team is now in delivery, where the questions are supposed to be closed. So the learning continues, but it happens in the least useful place: after commitments have been made, under pressure, without a budget for examining it.
The budget model
A more useful frame is to give discovery a standing budget rather than a stage. A percentage of every cycle, spent on reducing the uncertainty that matters most right now.
The budget is not necessarily large. For a team with a mature product and a clear problem, it might be a few hours a week: a handful of customer conversations, a quick instrumented look at behaviour, a prototype tested with five people. For a team entering a new market, it might be a third of the cycle.
The point is that the spend is continuous and owned, not a one-time event that either happened or didn’t.
Spend it on the expensive unknowns
Not all uncertainty is equal. The useful question is: which assumption would be most expensive to be wrong about?
A team might be uncertain about whether users will understand a new navigation model, and also uncertain about whether the pricing change will hold up with existing customers. The first is cheap to test and cheap to be wrong about. The second is expensive on both counts. The budget should follow the cost of being wrong, not the interest level of the question.
A simple ledger helps:
- Assumption: what we believe about users, behaviour or the market.
- Confidence: how much evidence we actually have, stated honestly.
- Cheapest test: the smallest thing that would move our confidence.
- Decision it informs: what we will do differently depending on the answer.
- Owner: who is responsible for running the test and reporting back.
If a line has no decision attached, it is not discovery. It is curiosity, which is valuable but should be funded separately.
Discovery is not a department
The other failure of the phase model is organisational. Discovery becomes the responsibility of researchers or designers, who produce findings that are handed to a product manager, who turns them into requirements, who hands those to engineers. Each handoff loses context.
Continuous discovery works better when the people making the decisions are close to the evidence. That does not mean every engineer must run interviews. It means the team should regularly see the raw material: a clip of a conversation, a session recording, a support thread, a usage pattern. Summaries are useful; they are not a substitute for contact.
The most durable habit I have seen is a short, recurring slot where the team looks at something real together (a customer call, a failing workflow, a piece of feedback with the context attached) and asks what it changes. Not a report. A contact.
The anti-patterns
Discovery theatre. Interviews conducted to confirm a decision already made. The output is a set of quotes that support the roadmap. This is worse than no discovery, because it creates the appearance of evidence.
Research without a decision. A study that is interesting, well-run and connected to nothing. It will be read once and forgotten.
The infinite discovery loop. The opposite failure: using uncertainty as a reason to never commit. Discovery is meant to make decisions cheaper, not to postpone them indefinitely.
The handoff. Research happens somewhere else, gets summarised, and arrives after the questions have changed.
The question to keep asking
The phase model asks, “Did we do discovery?” The budget model asks a better question: “Are we still learning about the thing we are building?”
That question does not end at launch. It gets more specific, and the evidence gets better, but it never becomes irrelevant. A team that stops asking it has not finished discovering. It has just stopped paying attention.