"We'll learn by shipping" is the most expensive sentence in product. It sounds pragmatic: bias to action, real data beats opinions, all true in isolation. But run the numbers on shipping as a research method and it stops looking pragmatic.
Three months of a small team building: R600k–R1.2m in salary or agency fees, plus the opportunity cost of whatever they didn't build. What you learn at the end: whether one specific version of one idea works. That is a sample size of one, confounded by execution quality, marketing, and timing. If it fails, you don't even know which part was wrong.
Twelve customer interviews: two weeks, a fraction of one salary. What you learn: whether the problem is real, who has it worst, what they do about it today, and what they'd abandon their current workaround for. If the idea dies, you know why, and the corpse usually points at a better idea standing next to it.
Why teams skip it anyway
Because most interviews are bad, and bad interviews genuinely are a waste of time. The failure modes are predictable:
Pitching instead of listening. If you describe your idea in the first half of the conversation, everything after is contaminated. People are polite. They will tell you your baby is beautiful.
Asking about the future. "Would you use a tool that…" produces fiction. People are terrible predictors of their own behaviour and excellent generators of hypothetical enthusiasm.
Collecting quotes instead of behaviour. A quote that agrees with you feels like evidence. It isn't. What someone currently does about a problem is evidence, especially what they pay, in money or time.
The interview that produces decisions
The structure I use is boringly consistent. Ask about the last time they dealt with the problem: the specific instance, not the general opinion. Walk the timeline: what triggered it, what they tried, where it hurt, what it cost them. Then the two questions that do most of the work: "what have you already tried to fix this?" (if the answer is "nothing", the pain isn't real, whatever they claim) and "what happened the last time this went wrong?" (severity in their words, not yours).
No pitching. No feature lists. The prototype comes later, in a separate session, once you know which pain you're testing against.
Patterns, not tallies
Twelve interviews isn't a survey, and you're not counting votes. You're looking for the moment three different people independently describe the same workaround: the spreadsheet they all secretly maintain, the WhatsApp group that patches the process. That convergence is the signal. One person's strong opinion is noise; three people's identical behaviour is a product.
In a recent sprint, that pattern killed the feature we'd been hired to scope and surfaced an adjacent workflow nobody had mentioned in a single planning meeting. The client built the adjacent thing. It converted eleven of twelve pilot customers.
Two weeks. That's what discovery costs. It's not a phase that slows the build down. It's the reason the build points somewhere.