The general advice is to sell shovels during a gold rush. What makes it a truly timeless idea is the fact that the intuitive thing to do is to go for the easily visible gold, but the smart people can think one step ahead and realize that if you make the shovels, whether or not an individual person gets the gold, the people who make the shovels always make money.
However, with the AI revolution that we have here, the opposite might be true.
With most disruptive technologies, the use cases for the end customer were clear from the start. With AI, we know that more automation would help, but it is not evident exactly how. Because of this, almost everyone is building one layer removed, with building the tools to build the value. It is a more legible and tractable problem.
Shovels are everywhere.
Let us consider the case of personal agents, which are in the vogue right now, the latest one being OpenClaw (née Moltbot, née née Clawdbot). For these, we have gone much beyond what we would have expected as “AGI” pre-2023(unless we move the goalposts to the next pin code). With these amazing tools, everyone is building different kinds of AI tools that claim to solve everyone’s problems under the sun. But at the end of it, when I ask people what problems they are actually solving, it is really equivalent to the simplest Google Assistant / Siri commands. Turns out most people actually don’t have many real personal problems to solve.
Let us be realistic: unless you are Anthropic or OpenAI (or players like Sarvam who are solving problems in more niche, but important areas), you do not have a real moat. What you are creating is a thin wrapper over the AI models, which unlike real pickaxes, become obsolete every few months as models improve. Model providers keep absorbing the tooling layer (RAG → long context, agent frameworks → native features) which makes building shovels an infinite pivot game.
I currently work as an “AI technical consultant”1 in my day job. Prior to the AI boom, this would entail 10% of the time gathering requirements, 60% of the time building it (which once upon a time needed a team and not just agents), and 30% of the time on iteration and improvements.
However, now the hardest problem is squarely the talking to customers, understanding the current business, and figuring out what there is to be built rather than the building itself. The building itself is barely 20%2. 80% is understanding the customer problems and iterating on a solution and getting real user feedback.
The moats are really the distribution and domain knowledge and no longer the integration parts, and engineering capacity is no longer the scarce resource. However, knowing which problems to solve and having a deep domain understanding has never been more important.
Maybe capturing all the context that is flying around in a large company and giving that context to the AI is a problem. However, it is also a problem a million people are trying to solve.
I know that of course you3 will be the exception, and your tooling company is going to become a genuine standard, the way Cursor became the standard for coding, the fastest company to reach a billion dollar business. Oh wait, it is already no longer the standard and people have moved on to Claude code and Codex. Maaaybe they will survive.
But statistically you won’t even be that. You will be one of seventeen-hundred near-identical solutions in a space where the underlying platform keeps eating the ecosystem from below.
While the context-capture problem will certainly be solved in the next couple of years, the foundation models are going to absorb that solution too. Meanwhile, if you can actually solve the real problems facing the company today for their customers, you can generate value immediately.
Building shovels in this particular Gold Rush is not a defensible business. So go mine for gold. The shovels are building themselves.
P.S. - It is appropriate that I am writing this post at a time when gold prices in the real world are also at an all-time high (but fallen from the hype levels) after a long slump.
Comments