The demo is not the product

You have probably seen it by now: someone on your team, or a founder friend, builds a working demo of a real product idea in an afternoon. It has screens, it has data, it responds. Five years ago that demo was a $40,000 conversation with an agency. Today it is a prompt.

So it’s a fair question, and buyers ask us some version of it every month: if the demo is free, why does custom software still cost tens of thousands of dollars?

What actually collapsed

The honest answer starts by agreeing. The cost of producing plausible software collapsed. Generating screens, scaffolding a database, wiring a form to an API: this is exactly the work AI tools are good at, and any studio not passing that speed on to clients is charging for effort that no longer exists. Our own quotes are lower and our timelines shorter than they would have been in 2022, for exactly this reason.

The money followed. Menlo Ventures measured enterprise AI spending roughly tripling in a year, from about $11.5 billion to $37 billion, with more than half of it going to applications rather than raw models (Menlo Ventures, 2025). Businesses are not experimenting anymore; they are buying software with AI in it.

What did not collapse

A production system is a demo plus everything the demo got to skip. Real data, always dirtier than the sample. Permissions, because in a real business not everyone may see everything. A migration from wherever the current process lives today. Then the edge cases, which are not edges at all: the customer with two accounts, the order cancelled mid-payment, the accent in a name that breaks a search index. Someone has to know the nightly sync failed before the client does, and someone has to answer for it afterwards. That last part is the one no model provides.

Security is the measurable version of this gap. Veracode’s research on AI-generated code found that roughly 45% of generated samples introduced known security flaws of the kind on the OWASP top-ten list (Veracode, 2025). The code ran. It demoed well. It was also, about half the time, unsafe to put in front of the public with real data behind it.

None of this work is glamorous, and almost none of it shows up in a demo. It’s also most of what you’re buying when you buy software that touches your revenue, your data, and your customers.

How to buy in this market

The practical conclusion is not “distrust AI-built software”. We build with AI daily, and the speed is real. The conclusion is that the demo has stopped being evidence. It used to be that a working prototype proved a team could build; now it proves they can prompt. So the questions that separate a product from a demo have to be asked directly:

1. Is there a written spec, and who wrote it? Not a proposal deck: a document with scope, data, integrations, risks, and acceptance criteria. If nobody can produce one, nobody has decided what “done” means.
2. Who owns the repository and the infrastructure? The answer should be you, from day one. Anything else is a hostage situation you’re paying to enter.
3. What ships with the launch? Error monitoring, usage analytics, and the number the system is supposed to move should be part of the delivery, not a future phase.
4. Who answers when it breaks at 9 p.m.? A name, not a ticket queue. If the builder’s plan for production incidents is “there won’t be any”, ask what happens the first time they are wrong.

A seller with good answers to those four questions is worth paying real money, whether or not it is us. A seller without them is selling you a demo with a delivery date.

We wrote our own answers into how AetherWorks sells: a paid Discovery that produces the spec before any build is quoted, fixed prices from that spec, instrumentation in every delivery, and Operations after launch so that someone is watching. That structure exists because those four questions deserve structural answers.

If you’re weighing a build like this right now, book a Discovery: one to two weeks, a written spec, a fixed quote, and a date. Worst case, you walk away owning the spec.