Stop Building Unicorns. Start Killing Subscriptions.

Spend a week in any of the communities orbiting AI-native app builders — Lovable, Bolt, v0, whichever this quarter's darling is — and one grievance shows up on rotation. The credits are a scam. The AI keeps breaking things. The bug-fixing loops burn through a month's subscription in an afternoon. The pricing is punitive. The platform hates you personally.

It is very tempting to nod along, because everyone likes a villain and pricing pages make excellent villains. But if you sit with the complaints long enough you notice something inconvenient: the people running out of credits are, with striking regularity, the people trying to build the same thing. A speculative, sprawling, unvalidated moonshot. 'Uber for dog groomers.' 'Airbnb but for allotments.' 'A social network but ethical this time.' The AI is not the problem. The scope is the problem. The pricing page is just the invoice for the scope.

The silver bullet trap

The seduction is real. Generative tooling has dropped the barrier to shipping software so far that it feels physically possible to prompt a multi-sided marketplace into existence over a long weekend. Two coffees, a decent Wi-Fi connection, and — surely — a Series A by Tuesday.

Software engineering, even when the actual typing is done by a model, punishes a lack of structure with unusual cruelty. Here is the lifecycle of the unicorn-hunter, compressed:

• →Vague opening prompt. 'Build me an app like X but for Y.' The model, being obliging, generates a sprawling scaffold that mostly compiles.
• →The refactor loop. The user realises they forgot half the business logic. Or changed their mind. Or watched a YouTube video. They prompt a major structural shift.
• →Cascading breakage. Multi-file refactors are the most expensive thing an AI builder does. Data structures were never locked down, so the bugs are load-bearing.
• →Credit exhaustion. Monthly allowance evaporates on fixes for problems the user shipped themselves.
• →The forum post. It is always the platform's fault. It is never the spec's fault, because there was no spec.

This is speculative development. Every credit spent is a bet on a hypothetical market that the builder has not spoken to, priced, or validated. The overwhelming majority of unvalidated MVPs die. When yours does, the credits died with it. That is not a pricing problem. That is a decision-making problem wearing a pricing problem's coat.

The pivot: build to stop spending, not to start earning

There is a much less glamorous, much more profitable way to use these tools. Point them inward. Instead of building software to make money from a market that has not agreed to exist, build software to stop giving money to a market that already has your card on file.

Every freelancer, agency, and small business is currently paying monthly rent on a stack of SaaS products they use roughly the way most people use a gym membership. The tool was picked for one feature. The bill is for the full enterprise suite. The renewal happens automatically because cancelling requires a phone call to Utah.

The risk profile for building a replacement collapses to almost nothing, because you are not gambling on product-market fit. You already are the market. You already validated the need — every month, by direct debit, for years. All you have to do is build the ten percent of the tool you actually use, and stop paying for the other ninety.

Three subscriptions begging to be strangled

• →Invoicing and accounting suites. The £30/month platform you use to generate five PDF invoices and log a dozen expenses. You are subsidising payroll modules, multi-currency reporting, and international tax compliance you will never touch. A single-page dashboard, a free-tier database, and a PDF library will replace it in a weekend.
• →Image compression and 'digital asset management'. The £15/month API that queues your PNGs and hands them back as WebPs. The browser's Canvas API does this for free, client-side, in the tab the user already has open. You are paying a monthly fee to avoid learning that this exists.
• →'Enterprise' CRMs and client portals. HubSpot is magnificent if you have a sales team. If what you actually need is a login, a milestone list, and a file table, you are paying for pipeline forecasting nobody has ever opened. Auth, a table, a timeline. That is the entire brief.

“The most valuable app you will ever build with AI is not the one that promises to make you rich. It is the one that quietly stops you from being fleeced.”

The maths nobody wants to do

Take a boringly typical solo-operator stack. Invoicing at £30 a month. Image tooling at £15. A client portal at £45. Ninety pounds a month, £1,080 a year, in perpetuity, forever, for software you do not own, cannot modify, and will lose access to the moment you miss a payment.

Redirect one focused build week into replacing all three with narrow, single-purpose tools running on free-tier hosting. Running costs afterwards: pennies, or nothing. That £1,080 is not saved once. It is saved every year, compounding, and it never has to be renegotiated with a customer success representative who will not stop calling.

You will not tweet about this. There is no launch. Nobody on LinkedIn will call you a founder. You will simply have more money and fewer logins. It is the least sexy strategy in modern software, and it is the only one that reliably works.

How to do this without setting your own credits on fire

The cost-avoidance play only works if you avoid the same undisciplined habits that hollow out the unicorn-hunters. The point is to build narrow tools cheaply — not to reinvent Salesforce on a hobbyist budget.

• →Write the spec before you open the prompt box. Inputs, outputs, database fields, the three screens. On paper. Treat it as a hard boundary. Feature creep is where the credits go to die.
• →Use Chat mode for the arguments, not the UI. Cheap tokens, structural decisions. Do not let the model start generating components until the schema is agreed.
• →One fat opening prompt, not fifteen skinny ones. These tools build coherent scaffolds in a single pass and mediocre patchwork over many. Front-load the detail.
• →Select and edit for tweaks. 'Fix the styling of the app' is an expensive way to say 'change this button'. Click the button. Change the button.
• →Ship at 'good enough for me'. This is internal tooling for an audience of one. Nobody is grading the animations.

The moral, briefly

Almost every complaint about AI development cost, when you drain the emotion out of it, is a complaint about undisciplined scope wearing a costume. The tools are not too expensive. The dreams strapped to them are.

Treat an AI builder as a slot machine for hypothetical markets and it will happily take your money for as long as you are willing to feed it. Treat it as a scalpel for the recurring line items on your own bank statement and the same subscription starts paying for itself inside a month. Same platform. Same credits. Entirely different outcome.

Stop trying to build the next unicorn. Start killing the subscriptions that are quietly eating this one.