Blog · Personal ·
Structural vs. Discretionary: The Question That Actually Matters
If you're building an AI product, forget the bubble debate—focus instead on whether your market's demand is rooted in real operational problems or just hype.
Everyone's arguing about whether AI is a bubble again. You've seen the charts, probably. Concentration risk in a handful of names, valuations that assume a decade of flawless execution, the PitchBook-style data points doing the rounds about how much capital is stacked into how few companies. It's a real debate, and honestly, smart people disagree on it, which is usually a sign the answer isn't obvious yet.
I'm not going to resolve it here. Anyone who tells you with total confidence which way this goes is selling you something, probably a newsletter subscription.
But here's the thing. If you're actually building, if you're a studio or a founder trying to ship something real, that debate is mostly noise. Interesting noise. Worth reading. Not the thing you should be organising your decisions around.
Wrong question, right question
The bubble question is fundamentally a market-timing question. It's asking: is capital mispricing risk right now, and when does that correct? That's a fine thing for an allocator to worry about. It's the wrong frame for someone deciding whether to spend the next eighteen months of their life building a product.
The question that actually matters, if you're building, is much narrower and much more useful: is the enterprise demand you're building against structural, or is it discretionary?
Structural demand doesn't care about the macro mood. It exists because something in how a business operates is genuinely broken or genuinely slow, and there's budget attached to fixing it regardless of what happens to public market multiples. Discretionary demand is different. It's the pilot budget. It's the innovation team's experiment line. It's real money, sometimes a lot of it, but it evaporates the second the CFO tightens the belt.
A bubble popping, if that's what happens, mostly wipes out the discretionary layer. It re-prices the hype. What it doesn't touch, generally, is demand rooted in an actual operational problem that existed before anyone said the words "generative AI" out loud. So the smarter question isn't "will valuations correct." It's "which part of the demand I'm chasing survives a correction." That's the whole game.
Three signals we actually use
We've written before about what it means for a market to be ready for an AI-native product, and this is really the same question wearing a different hat. When we're deciding whether to commit studio resources to a market, there are three signals we look for.
First: is there an existing, painful, manual workflow that someone is already paying people to do badly? Not a workflow that AI makes 10% nicer. One that's genuinely bottlenecked by human throughput, where the constraint is capacity, not preference. If the pain predates the technology, the demand tends to be durable, because it isn't tied to the current wave of enthusiasm. It was there before, and it'll be there after.
Second: does the buyer have a budget line that already exists for this problem, even if it's currently funding a worse solution? This one's underrated. A market where you have to convince someone the problem exists is a much harder market than one where the problem is already funded, just funded badly. Displacing an existing line item is a sales motion. Creating a new one from nothing is an evangelism motion, and evangelism is expensive, slow, and vulnerable to exactly the kind of sentiment shift a bubble correction would trigger.
Third: can the product get to a defensible workflow position quickly, meaning it becomes embedded in how the work actually gets done, rather than sitting alongside it as a nice-to-have layer. This is the difference between infrastructure and decoration. Decoration gets cut in a downturn. Infrastructure gets protected, because ripping it out is its own cost.
None of these three signals require you to have a view on whether the Nasdaq is overheated. That's sort of the point.
Where the durable stuff actually is
If I had to put a stake in the ground, and I do, because vague hedging helps nobody, the durable opportunities right now are sitting in unglamorous places. Back-office operations. Compliance-heavy workflows. Anything involving document-heavy processes in regulated industries, where the manual cost has been quietly enormous for years and nobody's bothered fixing it because the tooling wasn't good enough until recently.
These aren't exciting pitches. They don't generate the kind of headline that gets you invited onto a panel. But they satisfy all three signals: the pain is old, the budget already exists in some worse form, and the workflow embedding is achievable because the process is well-defined enough to actually automate properly.
The flashy, frontier-model, "reimagine how humans think" category of product is where I'd expect the bubble, if it is one, to do the most damage. Not because the technology is bad, but because the demand underneath a lot of it is aspirational rather than operational. Aspirational demand is the first thing to go quiet when sentiment turns.
So, anyway. Is AI a bubble? Ask an economist. Whether the enterprise demand under your specific product is structural or discretionary? That one you can actually go and check, this week, by talking to the people who'd be buying it. That's the more useful question. It's just less fun to argue about on a panel.