The first consumer protection AI doesn’t have
You ask a question. The model answers with total confidence. You point out it’s wrong. It agrees enthusiastically. “You’re absolutely right.” Then it generates the corrected answer.
Congratulations. You just paid twice. Once for the mistake, once for the apology.
No other product works this way. Or so I thought, until I actually checked.
The pizza problem
The obvious analogy is the wrong pizza. If a restaurant delivers pepperoni when you ordered margherita, they don’t charge you for the replacement. They fix it. So why does AI charge you for the wrong answer and then charge you again for the right one?
I like this analogy. I’m now going to destroy it, because if I don’t, the comments will.
A wrong pizza is verifiable in four seconds by anyone with a mouth. A wrong AI answer is sometimes discoverable in minutes, sometimes in a production incident three weeks later, and sometimes never. The refund model requires cheap verification. AI answers are expensive to verify. Hold that thought. It’s the whole problem, and it comes back later.
Also, the pizza place doesn’t tell you the wrong pizza was actually a valid interpretation of your ambiguous order. AI does that. Constantly.
The uncomfortable part
Here’s what the pizza framing misses. AI isn’t a pizza. It’s a professional. People use it as an engineer, a paralegal, a tutor, an analyst, a diagnostician. And when you compare AI to the professions it’s imitating, the pricing model stops looking strange and starts looking familiar.
Lawyers bill hourly when they lose. Doctors charge for the wrong diagnosis, then charge again for the right one. Consultants bill you to fix their own bad advice. I spent fourteen years in enterprise consulting before Microsoft. Ask anyone who has survived a large ERP implementation whether the fix came free. I have receipts.
So AI didn’t invent effort-based billing. It inherited the norm from the professions it’s replacing.
The scandal isn’t that AI charges for attempts. Professionals have always charged for attempts. The scandal is that AI is priced like a professional while carrying none of the things that make professional pricing tolerable. No license to lose. No malpractice insurance. No bar association. No board review. Nobody to sue.
We built a profession with all of the billing and none of the accountability.
Plot twist: the market already started answering
Here’s where it gets interesting. “Should you pay for wrong answers” is not a hypothetical. In one corner of the market, vendors have already said no.
Intercom’s Fin charges $0.99 per resolved support conversation. If it doesn’t resolve the ticket, you don’t pay. Zendesk launched per-resolution pricing this year at roughly $1.50 on committed volume and $2.00 pay-as-you-go. HubSpot cut its Customer Agent to $0.50 per resolved conversation in April. Sierra built its entire company on outcome-based pricing. No resolution, no invoice, and escalations to a human are mostly free.
This isn’t fringe. Bessemer’s pricing data shows pure per-seat SaaS pricing fell from 21% to 15% in a single year while hybrid and outcome models surged. Deloitte is publishing revenue recognition guidance for outcome-based agentic AI contracts. When the accountants show up, it’s real.
Outcome-based pricing isn’t a thought experiment. It’s a line item.
The interesting question is why it stopped where it stopped.
Why it stopped: the adjudication problem
Look at where accuracy guarantees exist today. Support tickets. Qualified leads. Booked meetings. Binary outcomes. Timestamped. Logged. Cheap to verify.
Now look at where they don’t exist. Code. Legal analysis. Medical guidance. Root cause analysis. Anywhere the outcome is “was this answer correct.”
A resolved ticket is measurable. A correct answer is contested. And the moment you need someone to decide whether the answer was correct, you need an adjudicator. Adjudication is where this gets ugly.
Zendesk offers a preview. For billing purposes, a ticket can count as resolved after 72 hours of customer silence. Think about what that measures. A customer who got their answer is silent. A customer who gave up and bought from your competitor is also silent. Both count as resolutions. Both generate an invoice.
The moment money attaches to “correct,” the definition of correct gets gamed. Goodhart’s law, now with billing integration.
And it cuts both ways. If “that’s wrong” earns free tokens, “that’s wrong” becomes the most typed phrase in the history of computing. Insurance solved this problem with adjusters and fraud investigators. Who adjudicates an AI answer? Another AI? Now you’re paying a third time, for the referee. Worse, verifying a complex answer can cost more compute than generating it did. The refund costs more than the product.
This is why your token refund isn’t coming. Not because vendors are greedy, although they are not running charities. Because nobody can afford to figure out whether you deserve one.
The actual answer: price by provability
Refunds are the wrong abstraction. They bolt a warranty onto a system that can’t tell you what it’s warranting.
The right abstraction is claims that carry their own evidence.
Accuracy guarantees can’t scale past support tickets because nobody has verification infrastructure that can adjudicate a claim cheaper than generating it. That’s an engineering gap, not an economic law. Close it and the economics flip.
Picture a system that separates its output into tiers. This claim is grounded in verifiable facts, and here they are. This claim is inference from those facts. This claim is a guess. Deterministic validation wrapped around probabilistic reasoning. Claims that cite their receipts.
Once a system can do that, pricing follows naturally. Guaranteed claims cost more. Speculation is cheap. That’s how the rest of the economy already works. A certified inspection costs more than a guy squinting at your roof, and everyone understands why.
I’ve spent the last two years building diagnostic AI where every claim has to survive adversarial review and trace back to evidence before a human sees it. It’s hard. It’s also the only version of this technology I’d attach a guarantee to.
Trust is the product
The AI market’s first phase competed on intelligence. Benchmarks, leaderboards, vibes.
The next phase competes on trust. And trust has a billing model.
The vendor that can say “if we’re wrong, fixing it is on us” and survive saying it will own the enterprise market. Not because the guarantee is generous. Because being able to offer it at all proves the verification layer exists. The guarantee is the demo.
The first AI company that can afford to refund wrong answers won’t need to. That’s the point.
Until then, keep saying “you’re absolutely right.” It’s the most profitable sentence ever written.
Pricing figures as of mid-2026: Intercom Fin per-resolution pricing, Zendesk automated resolution pricing introduced April 2026, HubSpot Customer Agent repricing April 2026, Bessemer Venture Partners 2026 AI Pricing Playbook.
Azure Advice 