Why Flat-Fee Subscriptions Don't Work for AI, and What to Charge Instead

The flat-fee subscription model is dying for AI products, and the replacement is already visible: charge per verified result. A lead-gen workflow charges per qualified lead. A document pipeline charges per document processed correctly. The customer pays when the thing works, which happens to be the exact moment they stop doubting that it works. This came out of trying to price our own AI products, and two forces pushed me off flat fees.

Problem 1: subscription fatigue

People are tired of signing up for yet another product that charges a flat fee they cannot evaluate. What does it save in money? In time? A subscription answers neither; it just recurs. This predates AI, but AI product flood made it worse.

Problem 2: nobody trusts AI products right now

The buyers fall into camps, and both have been burned.

The first camp got lost in the sheer volume of vibe-coded, ad-pushed AI products. They signed up for two, three, four of them, paid, and got nothing. Vaporware: shiny landing page, great-looking demo, and then a tool that either fails to deliver on the promise or is too hard for an average person to operate.

The second camp burned themselves. Leadership had no AI strategy and no idea how to use the tools, declared "we need to use ChatGPT," and sent people off with no training or handholding. The results were predictably horrible, and the lesson people took was that AI does not work. Stack the widely cited statistic that around 90% of AI adoption projects fail on top, and trust sits at an all-time low.

Selling a flat-fee promise into that market means asking people to pay first to find out whether you are lying.

How pay-per-result pricing works

Charge for the outcome, only when it is delivered. Per lead, per correctly processed document, per insight that produced a win. The "will it even work?" objection disappears because not working costs the customer nothing. Subscription fatigue disappears because the cost is variable and self-justifying. And incentives finally point the same direction: the vendor optimizes for the customer's results, because that is the only thing that generates revenue, and carries the risk alongside them.

This is value-based pricing, an old idea; the older version wrapped the value estimate back into a subscription premium. Pricing the results themselves is the honest version of it.

The honest constraint: it does not fit everything. Mostly it fits vertically integrated solutions and workflows, where a countable, verifiable unit of result exists. Where the defensible layer is your own workflow and data, a result is usually countable.

The margin mechanic nobody prices in

There is a structural bonus. Your customer buys the end product, a good lead, and does not care which model produced it. The frontier labs cannot capture token-cost declines, because their users always demand the newest, most expensive model. You can: if Haiku can produce the same verified result that Opus produced last quarter, your cost drops and the price stays put, because the deliverable did not change. Every model generation quietly widens your margin, so long as the customer does not demand a better end result, which for a warm lead they rarely do.

Use the bare minimum model that gets the job done. The customer only ever sees the result, and in this pricing model, the token bill is your problem and your opportunity.