The front-end is free now. The other ten per cent is the product.
A salesperson with no engineering background sent me a finished personal-finance application. It was genuinely good. It also couldn't be used by anyone but him, on any device but his.
The file arrived over WhatsApp on a Sunday. A salesperson had been describing an idea for a cash-flow app for a couple of weeks, and I had been nodding along the way you nod at an idea. Then he sent the thing itself: a single HTML file, sixty-five kilobytes, two external dependencies, six working pages.
I opened it expecting a sketch. It was an application. Manual transaction entry with categories and rules. Charts that recalculated as you typed. A goals tracker, a debt tracker, a monthly summary. Bank statement import: upload a PDF, and it extracted the transactions into structured records. An insights page that read the numbers back to you as a short financial-advisor narrative. Three weeks of evenings, by a man who does not write software.
Whatever your position on where AI coding is heading, that file settles one question. The cost of producing a working front-end has gone to roughly zero for anyone who can describe what they want clearly. The mockup is dead. What used to be a Figma file and a fortnight of pretending is now the actual thing, running, in an afternoon.
What the artifact gets right
Nearly all of it, and that is the surprising part. The layout is coherent. The state management is sane. The rules engine for auto-categorising transactions is the kind of small, clever thing an engineer would have argued about for a day and then built exactly that way. Nothing about the code says amateur; it says a good afternoon's work by any standard.
So I stopped looking for what was wrong with it and started looking for what it couldn't do. That list turned out to be short, precise, and identical to the list for every artifact like it.
The four walls
The AI dies outside the room. The statement parser and the insights narrative both call a model directly from the browser, with no key. That works inside the environment it was built in, and nowhere else. The moment the file is emailed to a second person, the flagship feature stops.
One user, one device. Everything persists to browser storage. Clear the cache and a year of finances is gone. Open it on a phone and you are starting from an empty ledger. There is no account, because there is nowhere for an account to live.
No identity, so no second customer. Without authentication there is no way to have two users, which means there is no way to have a hundred, which means there is no business — only a very good personal tool.
Nothing is metered. Every PDF parsed costs real money. An app that calls a model on demand with no accounting attached is a product with an unknown and unbounded cost of goods, which is to say not yet a product.
Read that list again and notice what is not on it. Not the interface. Not the features. Not the domain logic, the charts, the categorisation rules, or the import flow — the parts that used to be the work.
The gap between a working artifact and a shippable product has collapsed from ninety per cent of the effort to a thin, well-defined layer.
The layer that's left
Persistence that survives a cleared cache. Identity, so there can be a second user. A server-side route to the model, with a key that isn't in the browser and a log of what every call cost. Tenancy, so one customer's data cannot become another's.
That layer is unglamorous, it is the same every time, and it is precisely what Datassier is. Model calls run server-side through a metered route, so the key never reaches the browser and every call is accounted for; each tenant's data sits in isolated tables behind real authentication. Nothing about that has to be invented for the next app. It is the part that was already built.
Which reframes what the platform is for. A data warehouse that unifies your accounting, inventory and CRM is one use of it. The other is this: you build the screen, we make it a product. Someone in your business who understands the problem deeply can now produce a working interface for it in an evening. The reason those don't ship is never the screen. It is the four walls.
The uncomfortable second lesson
There is a version of this note that ends triumphantly, and it would be dishonest. Personal finance apps are one of the most crowded categories in consumer software — a search turns up a dozen "best of 2026" listicles, each with fifteen entries, most of them competent and free. If the front-end was the moat, it was never much of a moat, and it is now no moat at all.
So the thing that was hard to build is now easy, and the thing that was easy to dismiss — distribution, retention, a reason to open it in March that survives into June — is now the entire contest. That is uncomfortable for anyone who enjoyed building software as a craft. It is very good news for anyone who understands a business and has been waiting for permission to build for it.
The screens are free. What you need is somewhere for them to live.