Outcome pricing sounds fair, until you ask who decides done

A contract that says you pay only when the AI solves your customer's problem sounds as fair as it gets: if the AI fails, you pay nothing. The industry is arguing right now about whether AI tools should charge per seat or per outcome. But ask what ‘solved’ actually means, and no two vendors give the same answer.

A customer asks, the AI answers, and the customer types nothing more. One vendor counts that chat as a finished job and bills it. Another does not until a second AI has read it and confirmed the problem is really over. What really separates vendors that charge by outcome is not the price per job. It is who decides the job is done.

Productize looks at 7 AI vendors: how each counts a job as done, and who checks it, and closes with 7 questions to ask before you sign. Interests first: nearly every source is a vendor's own sales page or an article by a venture capital firm holding shares in the companies named, so the sources gain if readers believe this. Productize also sells services close to this topic. Every price is dated because prices change fast, and none is here to compare who is cheaper.

Part 1Who decides the work is done?

For the 3 AI customer service products below, the one deciding is the vendor's own system, following rules the vendor wrote. What differs is whether those rules count a customer's silence as done, and whether a checker sits apart from the AI doing the work.

Intercom's pricing page, fetched on 13 September 2026, charges $0.99 per outcome for its AI agent, Fin. An outcome counts in 3 ways: the customer confirms it is resolved, or "They don’t ask for more help after Fin responds", or "Fin completes a workflow (Procedure), including handoffs". Procedure is Intercom's name for a scripted multi-step task. The last two mean a customer who simply closes the chat, or an issue a staff member has to pick up and finish, may still count as a billable outcome.

Zendesk reads the same silence differently. Its help page on resolution tiers, in effect from 18 May 2026, says that if the AI answers and the customer asks nothing more, gives no feedback and does not ask for a human, the issue is only Contained. Once the conversation ends (email after 72 hours of quiet, chat after 2 hours), an AI rereads the whole conversation to check the problem was actually solved. Only then is it Verified. The Contained tier "does not count against your resolution allowance".

A Zendesk announcement the next day says every billed resolution is verified by the AI agent that did the work and also "independently confirmed by a dedicated AI evaluation model". Spam and routine exchanges are left out. But both checkers belong to Zendesk, which sends the bill.

Sierra wrote in a blog post in December 2024: "We provide clear, agreed-upon criteria for each outcome upfront". It does not say who checks whether an outcome meets them, or how. Any kind of work can go this way: if the contract does not say who checks, the one deciding is whoever sends you the invoice.

Anyone who judges their own work can get it wrong without meaning to cheat. When Productize had 11 AI readers go through the 70 sources behind this article, every one reported that its quotes matched the originals, because each checked only its own section. A separate script then compared them character by character and found 36 mismatches, 5.6% of those checked. That script was itself wrong twice, flagging 5 quotes as mismatched when every word matched. The checker has to be separate from the doer, and the checker has to be checked too. How this kind of check works is in the post on having AI read a whole PDF.

Part 2Billable outcomes end where the vendor's work ends

What buyers want most is the end result: sales, closed deals, a drug that gets approved. But in the examples we read, not one vendor bills on an end result like that.

An Intercom blog post from 8 May 2026 tells how it priced Fin for Sales, which screens prospects for the sales team. Intercom considered taking a share of sales, then dropped it, because much else affects a deal before it closes: a salesperson's demo, a prospect's budget getting cut. "Tying Fin’s price to the final outcome holds it accountable for variables entirely outside its control." So the price became $10 per qualified lead, with the customer setting the criteria. The post does not say how the customer checks a lead really met them.

Legal work lands in the same place. General Legal, an AI law firm from the startup accelerator Y Combinator (YC), charges $500 per contract for review and negotiation, per its YC page fetched on 13 September 2026. What gets billed is a finished contract review, not a closed deal. Panacea, an advisor on FDA (US Food and Drug Administration) submissions from the same program, writes: "Clients only pay once we deliver each agreed milestone." Whether the FDA approves is outside the advisor's control anyway. Billable outcomes end where the vendor's work ends, not where your business gets the result.

A cell saying the vendor does not publish this means the fetched source does not say, not a mark against the vendor. A shorter sales page does not mean looser checking. The standard this article sets for itself, that the checker has to be checked, applies to Zendesk too: it does not publish how accurate its evaluation model is.

Part 3Most outcome pricing ends up hybrid

If pure per-outcome pricing is that fair, why have so many vendors moved toward hybrid? One answer comes from buyers. A 30 July 2026 article from the venture capital firm Bessemer Venture Partners says Ada, which sells AI for customer service, started with pure per-resolution pricing, then went hybrid once the large organizations buying it wanted predictable annual budgets. This appears only in the article's opening summary; the detail on Ada is missing from the version we fetched.

Other vendors mix differently. The same Intercom pricing page has a monthly seat fee alongside the per-outcome fee. Zendesk's help page says your starting allowance depends on "Zendesk plan type and agent seats", and whatever you do not use does not carry over.

Salesforce's pricing page, fetched on 13 September 2026, offers per-user monthly pricing, $2 per conversation, and credits, which it announced on 15 May 2025 at $0.10 per action. That announcement calls this "a consumption-based model that aligns cost with business outcomes", but the unit on the bill is the number of actions the AI takes, not the result. So the word outcome in sales material tells you nothing until you see the unit on the bill.

Then what lets a vendor keep its prices high? The Bessemer article concludes that proof decides it. "Durable AI pricing ties revenue to something a customer can independently verify, not to a claim."

Bessemer invests in the companies it uses as examples, so this is an investor's view, but buyers can put it straight to use. If you cannot check the result yourself, what you are paying for is the vendor's claim.

Part 47 questions before you sign

These questions work from both sides. Buyers ask them of the vendor. Sellers priced by outcome write the answers to all 7 into the proposal before the customer asks.

Work out how you are buying. Sign up on the website with a credit card and there is no contract to amend, no salesperson to negotiate with: the terms the vendor publishes before you subscribe are all the rules you get, and what is left is to check whether the reporting drills down to single items, then go through next month's bill line by line. Under a contract, the same questions become clauses: the checker, a right to see per-item evidence, and a right to dispute. Questions 2 and 7 differ most between the two.

If question 2 is answered by the vendor's own system, the per-item evidence in question 7 is how you learn whether the number on the bill comes from real work. You do not have to check every item. Sample by risk: high-value items, ones customers complained about, ones counted from silence. The break-even fits in one line: the cost of inspecting an item has to stay below the chance of a miscount times its value. This article has neither number; both have to be measured on your own bill.

Take the latest AI contract you signed, the last bill you paid, or the proposal you are about to send, and go through it one question at a time. Any question you cannot answer is the spot where someone else is deciding for you that the work is done.

Sources and references

• Intercom, Pricing, fetched 13 Sep 2026 source
• Intercom, Building outcome-based pricing for Fin for Sales, 8 May 2026 source
• Zendesk, About automated resolution tiers and Upgrading from automated resolutions to resolution allowances source 1 source 2
• Zendesk, Zendesk Introduces the Autonomous Service Workforce, 19 May 2026 source
• Sierra, Outcome-based pricing for AI agents, 10 Dec 2024 source
• Y Combinator, company pages for General Legal and Panacea, fetched 13 Sep 2026 source 1 source 2
• Salesforce, Salesforce Introduces New Flexible Agentforce Pricing, 15 May 2025 source · Agentforce pricing page, fetched 13 Sep 2026 source
• Bessemer Venture Partners, Durable monetization: How four AI founders solved the pricing puzzle, 30 Jul 2026 source

Read next: All articles · Before you buy AI, pick the problem and how you will measure it · Y Combinator's playbook is your competitor's plan