There is a number that does not appear in most AI productivity reports: the percentage of AI-accelerated products that customers actually use. I call this gap discovery debt.
Rich Mironov recently made the argument plainly: AI can ship 100 times more code, but customer attention, budgets, and buying behaviour do not scale. You can have a 10x engineering team and still have a one-product market. Your customers haven’t multiplied, their tolerance for switching hasn’t increased, and their willingness to pay is not proportional to your throughput.
The productivity is real. The value may not be.
Most of the chatter about AI in product development focuses on what engineers can now do faster. Write tests. Generate components. Ship features in hours that used to take weeks. What the conversation mostly avoids is what happens downstream of that speed: the validation work that does not have an AI equivalent.
Does the feature solve a problem users actually have? Is there a way to find that out without building it first? Are users adopting your solution, changing their behaviour, and generating the kind of value that shows up in revenue? None of those questions get answered by shipping code faster. They get answered by talking to customers, running experiments, collecting, observing, and analysing, killing ideas before they become sunk costs.
Itamar Gilad calls the pattern code-first discovery: cheap engineering throughput is reviving the oldest bad habit in product development. Instead of validating with a no-code test — a survey, a fake door, a five-minute conversation — teams build a prototype first, because building is now cheaper than the administrative overhead of a proper discovery process. The prototype becomes the experiment. The experiment, if it performs, becomes the product. Discovery is not happening. It is being skipped.
This is not a new concern: waterfall encoded it as a methodology. Agile was largely a correction. What AI has done is make the issue much easier to fall into, and harder to notice — because you have something that looks like a product, shipped fast, by a small team, with confidence.
Let’s call it discovery debt.
Technical debt is code that works today but will cost you tomorrow, because shortcuts compound into systems that resist change. Discovery debt is the same dynamic, one level up: decisions about what to build, made without enough validated knowledge of why anyone would use it, by whom, and at what cost.
It is the kind of debt that accumulates quietly. Teams skip user interviews because the prototype is ready. Roadmaps are built on assumptions baked into the first sprint rather than validated before it. Products ship, metrics are flat, and nobody is quite sure whether the cause is the solution or the problem itself — because nobody confirmed the problem existed at scale before building for it.
Codifying before you automate is the same principle in a different domain: skip the foundational step and whatever you build next is on sand.
The hard part is that it does not feel like debt: it feels like momentum!
You have a demo. You have users who signed up for early access. You have data — incomplete, maybe misinterpreted, but data. The team is shipping. Velocity is high. Every leading indicator is positive, and the lagging indicators have not caught up yet. By the time they do — flat retention, poor activation, customers churning after the trial — the debt is large and the team has moved on.
The real PM test is not whether you can ship. Anyone can ship. It is whether you can distinguish between features that are technically impressive and features that solve a problem the market has and is willing to pay for.
AI did not change that test. It just made it easier to fail at higher speed.
Build fast and validate intentionally. The mistake is treating these as the same thing.
Discovery is continuous: you ship something, you learn something, you update your understanding of the customer and what they actually need. AI can accelerate the build side of that loop. It cannot run the learn side. It can create the illusion the loop is closed when it isn’t: a model that generates usage data from synthetic users, or predicts retention from patterns in a different product and context is not the same as real users making real decisions with their time and money.
We are generating more code, more features, more products than at any point in the industry’s history. And probably more discovery debt too. Because the cost of building has fallen faster than the cost of being wrong.
Is your understanding of the problem keeping pace with your shipping speed? If it doesn’t, you are not building products faster. You are accumulating debt fast.
Leave a Reply