Anyone can build an AI demo now

“Can you just build us something?”

I hear a version of that most weeks. And the honest answer is yes. So can you. A working prototype is an afternoon’s effort now. You describe what you want, the tools write the code, and by dinner there’s something on the screen that looks like it works.

That’s a good thing. I’m not going to talk anyone out of it.

But…

A demo and a working part of your business are different objects. The demo runs on clean examples, one user and a forgiving audience. The business runs on the invoice with the wrong code, the customer who replies to the wrong email, and the person on leave whose spreadsheet everything quietly depends on. Most of the cost lives in that gap, and most AI projects stall there. One large survey found the average company scrapped close to half its AI proofs of concept before they reached production (S&P Global, 2025).

When building was expensive, the build was the risk. You spent months and a lot of money and hoped it worked. Now the build is cheap, so the risk has moved. It sits in two places.

The first is where you point it. If the prototype solves a problem that isn’t costing you much, it doesn’t matter how well it works. You’ve built a fast answer to a question nobody was paying for.

The second is what it costs to keep it running. The price of using these models keeps falling, and yet total spend on running AI keeps climbing, because the work we hand it keeps getting heavier (Gartner, 2026). Then there’s the cost nobody budgets for. Someone has to check the output, handle the exceptions, show the team how it fits their day and own it when it breaks.

So the question we should be asking has changed. We used to ask whether something could be built. Now we need to ask whether it should be built here, and what it will cost to live with.

I saw this play out with a national allied health provider. The brief was AI to support their clinicians, and it was a reasonable brief. Clinicians were stretched, and better tools would help.

When we followed the work instead, the money was somewhere else. Invoices were being rejected over small details, a wrong billing code or a handling slip, and they piled up as aged debt. A lot of it was eventually written off. That’s where AI went, at each point in the chain where a small detail could break. It’s now worth $1.5M a year in aged debt that no longer builds up, and the team who used to chase those invoices is moving to customer service.

None of that came from building faster. It came from building in the right place.

This is why I spend most of my time before anything gets built. I want to know what the problem costs today, in dollars, and who carries it. I want to see where the work waits and where someone handles it twice. And I want a number we both agree on before we start, so we can tell afterwards whether it paid.

I believe knowing where AI will pay is now the scarce skill, and it’s going to matter more as building gets cheaper. The tools will keep getting easier. Pointing them at the right problem won’t.