Finding AI Gold With Lean Startup Techniques

Most AI conversations jump too quickly from “AI can do something here” to “let’s build the bot.” That is how organizations end up with expensive pilots, clever demos, and very little business impact. A better starting point is the business constraint: where value is currently stuck, which behavior needs to change, and what would count as evidence that AI helped.

The Lean Startup move is simple but still underused in AI work. Treat the AI capability like a product. Pick the outcome, identify the riskiest assumption, and run the cheapest useful experiment before you build the full workflow, agent, or integration. AI might be part of the solution, but it is not the goal. The goal is a better business result with enough evidence to justify the next bet.

There is clearly an AI gold rush happening. Some companies are making very real money building the hardware, platforms, and tools. Inside most organizations, though, the question is more uncomfortable: where is the gold for us? Where does AI actually improve the business, not just create another demo?

That question matters because a lot of AI work still starts from the solution. Leaders ask which tool to buy, which model to use, which data needs to be connected, or whether they should build an agent. Those are not bad questions, but they are second-order questions. The first question is what outcome needs to change.

This is where product thinking is useful. Good product work does not start by asking “what can we build?” It starts by asking who has a problem, why it matters, what behavior would need to change, and what evidence would show that the solution is worth more investment. AI transformation needs the same discipline.

One useful way to look for AI opportunities is to map your customer factory — how you acquire customers, activate them, retain them, turn them into revenue, and turn the happy ones into referrals. When that loop runs well, the business flywheel turns more easily. When one part of it is stuck, throwing AI at another part may feel productive while doing very little for the business.

This is basic theory of constraints applied to AI. If you do not have enough qualified opportunities, using AI to optimize retention is probably not the first move. Acquire plenty of customers who then leave quickly, and more AI-generated leads just pour water into a leaky bucket. And when customers are happy but referrals are weak, the constraint sits somewhere else entirely.

The point is not that every company has the same bottleneck. The point is that the AI conversation gets much sharper once you name the constraint. “Where can we use AI?” creates a massive list. “Where is the constraint in our customer factory?” creates a much better conversation.

AI can be used almost anywhere, which is exactly why it needs strategy. If everything is an AI opportunity, nothing is a strategic AI bet. You need to decide where AI will play internally for the next stretch of work: sales, marketing, onboarding, support, delivery, product development, decision flow, talent, or some other part of the system.

That choice is not only about where AI is technically feasible. It is about where the business needs movement. In my own business, for example, retention is not the current problem. If customers come in, they tend to stay happy and see value. The more interesting question is acquisition and conversion: how do I turn the people who hear me on podcasts, watch videos, read articles, or visit the site into real conversations?

That focus changes the AI discussion. Instead of asking whether I can use AI for everything, I can ask whether it helps with a specific business outcome: understanding which prospects are worth following up with, preparing for a conversation in a way that matches my style, or turning long-form content into useful starting points for the people trying to solve a real problem.

One of the biggest traps is putting AI in the objective. “Use AI in sales.” “Launch an AI agent.” “Automate onboarding.” Those may be useful activities, but they are not the business outcome. The outcome might be a better close rate, shorter time to first value, fewer handoff delays, faster expert onboarding, better customer retention, or more qualified conversations.

Once the outcome is clear, AI becomes one possible way to get there. Maybe the solution uses AI. Maybe it uses AI plus a process change. Maybe the cheapest useful first move does not use AI at all. That is fine. If AI is the means rather than the goal, you are free to learn your way toward the result instead of defending the tool.

This is why Lean Startup language still matters here. Form a hypothesis. Name what you believe will happen. Decide what you need to learn first. Then run the smallest test that gives you useful evidence. That discipline protects you from spending too much money on an idea before you know whether the problem, behavior change, and workflow fit are real.

When leaders jump to building an AI capability, they often start with feasibility — can we connect the data, build the agent, make the model handle the task? Those are important questions, but not always the riskiest ones. Sometimes the bigger risk is desirability: whether people will want this, trust it, use it, or change their work because of it.

Imagine a fast-growing services business where the constraint is bringing the right people on board and getting them productive quickly. The first hypothesis might be that AI-based recruiting will help. Another hypothesis might be that AI-assisted onboarding will help. Another might be that the business can serve more clients without hiring as many people if AI reduces the expert bottleneck.

Those are different bets, and they carry different risks. Before building a full agent or workflow, ask what would need to be true: whether clients trust AI in this kind of work, whether internal experts can explain their judgment well enough for AI to help, whether the data is actually there, and whether new hires would use the guidance once it existed. The sharpest version of the question is the one that decides everything — which assumption, if it turns out wrong, kills the whole idea?

Not every idea deserves the same discovery effort. When value is high and uncertainty is high, slow down just enough to test before you commit. When the value is just as high but the uncertainty is low, skip the ceremony and go do it, then measure. The low-value quadrants tend to sort themselves out: high uncertainty there is usually a distraction, and the cheap, easy ones are either a quick improvement or parking-lot material.

This is the practical value of a hypothesis prioritization canvas. It helps you avoid the two extremes: analysis paralysis on one side and tool-first action on the other. You do not want to spend months debating. You also do not want to burn leadership attention, money, and trust on ideas that could have been invalidated cheaply.

The phrase I like here is “learn before you burn.” Every hour spent working on the business is expensive. Every AI pilot consumes attention. Every integration creates follow-on work. If conviction is low, buy learning first. Once conviction grows, move from discovery to delivery.

People often jump from idea to MVP. In AI work, that might mean building a working chatbot, connecting an agent to internal systems, or giving a team a new AI workflow. But even an MVP can be too much if the core risk is whether anyone wants the thing or whether the behavior change makes sense.

The old product lesson still applies: sometimes you can learn with something smaller than a working product. You can simulate the experience. You can run a concierge test. You can create a landing page, a short demo, a manually assembled result, or a lightweight prompt that gives people a taste of the future capability.

For AI agents, I often think about this as “be the bot before you build the bot.” If you cannot create value manually with the same inputs and outputs, the agent is unlikely to save you. If people do not want the result when a human helps produce it, they probably will not want it because an agent produced it.

There is another useful product move here. Before building full agents, use lower-cost AI surfaces such as projects, gems, spaces, custom GPTs, and carefully designed prompts. They are not the final answer, but they force you to think through the work. What context does the AI need? What inputs matter? What output is actually useful? What decisions should remain human?

I use this myself for sales coaching. I do not want generic hard-sell advice. I want a coaching assistant that understands my business, my style, the kind of advisory work I do, the problems I care about, and the sales coaches whose thinking resonates with me. That context makes a simple prompt much more useful than a generic “coach me on sales” request.

This is where the discovery work creates sawdust that is actually valuable. The more clearly you describe your strategy, constraints, audiences, examples, and decision criteria, the better your AI work becomes later. The context you create while searching for AI gold becomes part of the asset base that helps you find more of it.

This article is based on a LinkedIn Live / YouTube talk where I walked through how Lean Startup, product discovery, and constraint thinking can help leaders find real AI value.

The real prize is not an AI agent, a better prompt, a cleaner demo, or a longer list of use cases. The prize is business impact. A constraint moves. A customer gets value sooner. A team learns faster. A leader makes a better decision. A costly delay shrinks.

That does not happen because an organization sprinkles AI everywhere. It happens when leaders aim AI at the right constraint, treat the work as a learning problem, and build confidence before they scale the bet. That is how you find AI gold without mistaking every shiny rock for value.

AI is powerful. Product thinking is how you decide where that power is worth aiming.