Arcana Member Brief — Procurement & AI Spend

Who gets to define the result you're paying for

The next battle in the constant war over AI pricing isn't the rate card, it's who gets to define what results you're actually paying for.

On Sunday, The Information reported that Salesforce is starting to let businesses choose how they pay for Agentforce, including custom contracts tied to revenue the AI helps salespeople close, or cost it cuts through automated customer-service interactions. The full article wasn't in the scout, and no source here has Salesforce confirming this overhaul, so treat it as a reported signal, not a product you can buy today.

OpenAI is also testing the same thing for some large enterprise customers. The Decoder, citing The Information, reports that over the past few months OpenAI has started giving some large customers the option to pay only once its AI actually finishes a task, such as handling customer support requests. An OpenAI spokesperson declined to comment. This is not generally available.

Outcome-based pricing, as this structure is called, is all the rage in AI circles. The argument is that pricing based on number of seats doesn't feel very customer-aligned, as people can have access to AI tools but not use them. Vendors want enterprises to pay when AI actually does real work — and they want that bill to be far bigger than a stack of seat licenses. From a business perspective it's a potential goldmine: you're essentially setting up AI to compete with the budget companies spend on human labor rather than software services.

But the direction is clear enough to plan for. Once you pay for results, the argument moves to the fine print. What was the number before? Over what window? What is left out? Who gets credit when a person, a process change, or another tool also helped?

None of that gets sorted out after signature. Before you take the first pricing call, have finance rebuild one of these numbers out of your own books — last quarter's closed-won, or your cost per resolved service ticket. If that takes your team three weeks, that's three weeks the vendor's version of the number stands unchallenged. And if they won't walk you through the calculation before you sign, the calculation is the price. Sign a results deal with no shared baseline and you've bought a variable invoice you have no standing to dispute.

Also in today's news: Travelers has started sending insurance questions to its own model, Anthropic is changing how much work a weekly seat buys while its own help page still says something else, and OpenAI is threatening to cut Cursor off from its models after the SpaceX acquisition.

Arcana's view

Results-based pricing is worth taking seriously. The word “outcome” on a term sheet, though, commits the vendor to nothing by itself. The test is whether you can pull the number apart, argue with it, cap it, and walk when it stops working. If you can't, the vendor is scoring its own game and sending you the bill.

The week in review—

Three things to bring to your next vendor meeting

01

Travelers built its own model and sends the hard stuff elsewhere Which model

Travelers built TravelersLLM for insurance questions and routes broad reasoning, research, and coding to frontier models. CTO Mojgan Lefebvre says the in-house model is cheaper and answers insurance questions better. Those are company claims, carried by CIO Dive, not an independent ROI study. No dollar figure and no list of frontier vendors is preserved.

02

The seat price stays. The amount of work it buys moves Usage limits

Anthropic says that on September 14, weekly Claude Code limits land 25% above the original baseline for named plans, and about 17% below today's temporary boost. It publishes no actual usage numbers. At the scout cutoff its Help Center still showed an August 31 end date, so September 1–13 is unexplained.

03

OpenAI may stop supplying models to Cursor Supplier risk

OpenAI says it intends to wind down model supply to Cursor after the SpaceX acquisition, and proposes November 12. Cursor says OpenAI is about 5% of user traffic and that talks continue. The date is not final. Whether bring-your-own keys or other workarounds keep access is not verified here.

The decision gate—

Before you sign a deal priced on results

Forward this · five questions before you sign

Five questions before you sign

The mathCan we check the math? Both sides should be able to rebuild the baseline and the claimed result from the same records.

Who gets creditWho gets credit? Write down what is excluded — other people, other tools, process changes, market swings — before any results show up.

Usage limitsWhat happens when usage is throttled? Get the weekly cap, the shorter limits, the overage path, priority, and how much notice you get before it changes.

Which modelWhich model actually does the work? Name the tasks, the quality bar, the fallback, the review step, and the total cost per task.

If they cut you offWhat if the vendor cuts us off? Name a second path you have actually tested, with the same data rules and the same test set.

Chart of the week—

Four things that moved in the same week

Signal	Observed evidence	Supports	Does not prove
The math	The Information says Salesforce is negotiating some Agentforce contracts tied to revenue helped or service cost cut.	How you count results can become a price term.	General availability, a standard meter, audit rights, or customer value.
Which model	Travelers says insurance questions go to TravelersLLM while broad tasks go to frontier models.	Splitting work by task can cut cost and reduce lock-in.	A dollar saving, independent ROI, or that it would work at your company.
Usage limits	Anthropic says September 14 weekly limits will be 25% above the original baseline and about 17% below today's boost.	How much you can use is its own term, separate from price.	Actual usage numbers, or how every Enterprise plan is handled.
Getting cut off	OpenAI says it intends to end model supply to Cursor, with November 12 proposed.	A change of owner can cut off a model you depend on.	A final shutoff date, or that any workaround keeps access.

Operator's edge—

Write the rules so finance can reproduce them

One page should show what you pay, how much you can use, which model runs what, the evidence behind the numbers, who you depend on, and how you get out.

Draft one outcome calculation before you negotiate the rate.
Write down the usage limits, not just the seat price.
Run the same set of tasks through both model paths, scored one way.
Rehearse losing one model you get through somebody else.

Total workflow cost
÷
Accepted tasks

+ a tested exit path

The watchlist—

What still needs an answer

The mathWho can actually get these Agentforce contracts, how is a result counted, and what audit or clawback terms sit behind them?
Usage limitsWill Anthropic fix the August 31 Help Center date against the September 14 announcement, and say how consumption-based Enterprise seats are treated?
Getting cut offWill OpenAI and Cursor confirm or move the proposed November 12 date, and will either one document a supported alternative?
IP riskSony Music Publishing and Warner Chappell allege copyrighted compositions were taken to train Claude. These are allegations, not findings. Damages and Anthropic's response are not established here.
Policy holdA one-sentence teaser on remote chip access has no agency text, no covered products, no countries, no timing, and no enforcement detail. Held until there is primary evidence.

One question—

Which of the five would you refuse to leave blank?

The math, who gets credit, usage limits, which model, or what happens if they cut you off? Hit reply — every response gets read.

Reply to Ian →

Sources & evidence boundaries—

How to read the evidence. This review was refreshed from the supplied local arcana-scout.v2 JSON only. Paywalled teaser claims remain teaser claims; vendor statements remain vendor-owned evidence; allegations remain allegations; proposed dates remain proposed; and held leads stay outside the promoted issue. Research cutoff: 2026-08-30T22:44:50Z. No AgentMail, web, X, or browser call was made for this refresh. The OpenAI outcome-pricing item is from The Decoder citing The Information, supplied by the editor, not from the scout JSON.