How to Tell If Your AI Subscription Is Actually Worth It

TL;DR

OpenAI's chief financial officer, Sarah Friar, published a framework on July 17, 2026 for judging whether AI spending is actually paying off. She calls it "useful intelligence per dollar." It replaces seat counts and login stats with three practical questions: what did the AI actually finish, what did each finished task cost once you count the cleanup, and is that cost getting better or worse as you rely on it more.

For a five-person business paying for two or three AI subscriptions, that is a genuinely useful lens once you strip out the enterprise language. Below is how to actually run the math, using a use case almost every small business has: an AI tool that drafts invoices.

What is OpenAI's "useful intelligence per dollar" framework?

Friar's framework scores AI spending on four things: whether the AI is completing work that matters, what each successful task costs once you include the cost of fixing mistakes, how often the output can be trusted without a person correcting it, and whether the value produced grows faster than the cost as usage scales.

That last point is what Friar and outlets covering the piece call return on compute. It is really just asking whether a tool gets more valuable per dollar the more you rely on it, or whether the bill and the babysitting grow at the same rate. Friar's own line on it: "Tokens create value when they transform into work people can use" (OpenAI, "A scorecard for the AI age," July 17, 2026). CFO Dive's coverage of the piece is blunt about what it replaces: software's old habit of measuring adoption, meaning seats, active users, and renewals, instead of the work actually done.

Why doesn't "the team uses it every day" tell you anything about ROI?

Adoption tells you people opened the tool. It does not tell you the tool finished anything useful. A team can hit 100% daily usage on an AI tool that still needs someone to rewrite half of what it produces, and that rewriting time is a real cost, even though it never shows up on an invoice.

This is the uncomfortable part of Friar's framework for a small business. The tool with the highest reported usage is sometimes the one quietly costing the most, because "usage" counts a person opening the tool and fixing its work as a win, same as it counts a clean, first-try result.

When I sit down with a client to look at their AI line items, the question I ask is not whether they use a given tool. Everyone says yes to that one. I ask what happens when the tool gets something wrong, and how long that takes to fix. Almost nobody has an answer, because nobody has ever tracked it. That gap is exactly what this framework is built to close.

What does "cost per successful task" actually look like for a 5-person business?

Take an AI tool that drafts invoices for $50 a month and produces 200 draft invoices. On paper, that is $0.25 per invoice. But if a bookkeeper has to fix 60 of those drafts before they go out the door, the honest cost per successful invoice is closer to $1.43, once you add the time spent fixing them.

Here is the full math. Sixty corrections at six minutes each, at a $25-an-hour bookkeeper rate, is $150 in labor. Add that to the $50 subscription and the true monthly cost is $200. Only 140 of the 200 invoices went out without a human touching them, so the real cost per successful task is $200 divided by 140, or $1.43.

Now compare a pricier tool: $120 a month, but only 10 of 200 invoices need a correction. That is $25 in labor, plus the $120 subscription, for $145 total. With 190 successful invoices, the cost per successful task is $145 divided by 190, or $0.76, roughly half the "cheaper" tool's real cost. The tool with the higher sticker price is the better deal, and the monthly invoice alone would never have told you that.

How do you track dependability without a data team?

You do not need software to track dependability, just a tally. For two weeks, log every AI output as one of three things: used as-is, needed a quick edit, or needed a full redo. That gives you a real error rate, which is the same dependability signal Friar's framework asks enterprise finance teams to track, just kept in a spreadsheet instead of a dashboard.

Two weeks is usually enough to see a pattern. If a tool's "needed a full redo" column keeps growing, the subscription price was never the real cost. The correction time was, and now you can put a number on it instead of a feeling.

Does "return on compute" mean anything at this size?

Yes, in a simpler form than Friar's enterprise version. Watch whether your cost per successful task goes down as you use a tool more, or creeps up. A dropping number means the tool is compounding in value: you are getting better at prompting it, or it is genuinely improving. A rising number means more volume is bringing more correction work, and no subscription discount fixes that on its own.

This is also the fastest way to catch a tool that looked great in a free trial and got worse once real, messy client data hit it. Track the number monthly for anything you are paying real money for, even a rough version of the tally above.

Does this only work for invoice drafting?

No. The invoice example is easy to follow because the math is small, but the same three questions apply to almost any AI tool an SMB pays for. A customer-service chatbot has a cost per successful task too: the subscription plus the time a human spends stepping in when it cannot resolve something, divided by the conversations it actually closes on its own. A scheduling assistant has one: the subscription plus the time spent fixing double-bookings, divided by appointments booked correctly the first time.

The pattern holds across all of them. Whatever the tool does, ask what it finished without help, what that cost once corrections are counted, and whether that number is getting better or worse the more you use it. A business running four or five AI subscriptions can run this same tally on each one and quickly see which are actually paying for themselves and which are quietly costing more than they save.

So what does this mean for your business?

It means changing the question you ask about any AI subscription from "do we use it" to "what does it actually finish, and what does that really cost." That single swap is what tells you whether to keep paying for a tool, renegotiate it, or cut it loose, and it applies whether you are spending $30 a month or $3,000.

Run the tally for two weeks on your most-used AI tool. If the cost per successful task is climbing, that is real money leaking out of a line item that looked harmless on a monthly statement. If it is falling, you have found a tool worth building more of your workflow around, which is time freed up for the work that actually grows the business. We built the same discipline into Perpetua's Mission Control view, because an SMB owner should be able to see this without hiring someone to track it, not because the math above requires our product to work.

For more on why "we're using AI" is not the same as "AI is working" for a small business, see our recent look at the gap between agent adoption and agent reliability and our breakdown of the SMB AI skills gap. If you have not yet mapped which of your operations are AI-ready, The AI Operations Blueprint is the place to start.