One company reportedly hit $12,000 per engineer per month in token spend, org-wide average.
That's the number to hold onto while reading Salesforce's new pitch. Because Salesforce is now offering enterprises a deal that sounds like the antidote to exactly that problem: pay for AI only when it works.
The mechanic, as The Information's Laura Bratton described it, is disarmingly clean. "Rather than just how many messages the AI sends to a customer, it's actually only charging you when AI finishes a task."
Salesforce is going further than task completion. Customers can now pay for Agentforce based on whether the AI "boosts revenue or cuts costs" — a pricing model tied directly to business outcomes, not usage.
Read that as a buyer and it feels like a gift. No more paying for seats nobody logs into, no more usage meters ticking on failed queries. You pay when you win.
But the pitch buries a question it never answers: who decides when it worked, and how would you ever check?
Start with "finishes a task." Finished by whose definition? If an agent closes a support ticket the customer reopens an hour later, did it finish? If it drafts a reply a human rewrites, does that count as a completion or a false one?
Now escalate to "boosts revenue or cuts costs." That's not a completion event you can log. That's a causal claim about a business, and causal claims require attribution — the single hardest thing to establish in any enterprise.
Here is the uncomfortable truth outcome pricing is built on: nobody, including Salesforce, can currently measure agentic ROI with any rigor. Outcome pricing isn't generosity. It's a confession.
Here is the uncomfortable truth outcome pricing is built on: nobody, including Salesforce, can currently measure agentic ROI with any rigor. It's a confession.
The clearest evidence comes from Bessemer, which spent the first half of 2026 talking to more than 20 CTOs about agentic AI in engineering teams. Their finding demolishes the idea that sophisticated buyers can instrument this.
On average, organizations spend $1,000 to $2,000 per engineer per month on tokens. Many spend less. But the average is a lie the moment you look inside it.
The distribution runs from near-zero light users to heavy parallel-agent users at $8,000 to $10,000 a month. Same job title, same company, a ten-x spread in spend.
And spend tells you nothing about value. As Bessemer put it, teams have super-engineers producing extraordinary output and inefficient "token-maxxing" — and "high spending doesn't tell you which."
The $12,000-per-engineer org sits at the extreme of that curve. Another comparison in the same data, per Bessemer, references around $1,700 per engineer a month paired with public claims of 30% productivity gains. Same input intensity, radically different costs, driven by mundane plumbing — packages, caching, model routing.
If a company's own CTO, staring at its own logs, can't separate the super-engineer from the token-maxxer, consider what a Salesforce customer is being asked to do. They must verify a vendor's claim that its agent boosted revenue — using data the vendor helps produce, against a baseline nobody instrumented before the spend hit.
Bessemer even spells out what real measurement would require: define a peer set of comparable companies, then instrument three numbers — AI-attributed code percentage, effective FTE per human engineer, and token spend per engineer — updated quarterly.
Their verdict on who actually does this: "Almost no one does this systematically yet."
That's the phrase that should worry every buyer signing an outcome-based contract. The three numbers that would let you audit a vendor's ROI claim are numbers you don't have. You're being asked to true-up a bill against a ledger that doesn't exist.
And the spikes are arriving before the ledgers do. Uber reportedly burned through its entire 2026 AI budget by April — with 84% of engineers agentic and more than 70% of code AI-generated.
Salesforce itself is substituting tokens for hiring at roughly $300 million a year. Two of the most technically capable organizations on earth, and Bessemer's dry note lands hardest here: "Neither had systematic measurement before the spike hit."
That's the environment into which Salesforce is introducing pay-for-outcomes. Not a world where ROI is legible and vendors are nobly sharing the upside. A world where the base rate is: the money moved, and nobody knows what it did.
So why would Salesforce do this? Why volunteer to price on the one thing it admits can't be measured?
Because the alternative is worse. Salesforce sells seats. Its entire history is per-seat SaaS, and per-seat SaaS has an existential problem in an agentic world: agents don't sit in seats.
If work shifts from humans to agents, the seat count stops tracking the value delivered. And the moment customers notice seats don't map to value, the renewal conversation becomes a downgrade conversation. That's the collapse Salesforce is most exposed to.
Outcome pricing gets ahead of that reckoning. Instead of letting customers discover that seats are decoupled from value, Salesforce reframes the whole transaction around value — and then positions itself as the party keeping score.
That last part is the tell. In an outcome contract, someone has to adjudicate whether revenue rose because of the agent. Salesforce, which owns the CRM, the pipeline data, and the attribution surface, is far better positioned to game that measurement than any customer is to audit it.
The information asymmetry doesn't disappear under outcome pricing. It inverts in the vendor's favor. Before, the buyer at least knew how many seats it bought. Now the buyer is trusting the vendor's account of causality inside the vendor's own system.
There's a deeper structural irony here, and it runs through Salesforce's cash machine.
Salesforce is one of the purest expressions of the negative-working-capital software model. Customers are invoiced annually and pay upfront; the cash lands on day one and gets recognized as revenue slowly over twelve months.
That model reportedly carries net working capital of roughly negative $16.9 billion — customers, in effect, financing the company interest-free. Deferred revenue reportedly in the $18–24 billion range accompanied free cash flow of about $12.6 billion in fiscal 2026.
The entire engine runs on measurable cash arriving before the service is delivered. It is prepaid, contractual, and legible.
Outcome-based Agentforce pricing is the inverse of that engine. Instead of measurable cash upfront, Salesforce takes unmeasurable outcomes on the back end.
The company whose whole financial identity is "get paid first, prove value later" is now, for its flagship AI product, saying "prove value first, get paid after." That's not a minor tweak to a price sheet. It's a trade of the most certain revenue in software for the least certain.
It only makes sense as a bet. Salesforce is wagering that Agentforce works well enough — and that it controls the scoreboard tightly enough — that it can win a game in which it also keeps score.
If the agents deliver, outcome pricing lets Salesforce capture more than seats ever would, because it's paid on value it helped define. If they don't, the vendor absorbed attribution risk it's uniquely equipped to manage down.
Either way, the buyer's position is the same one Bessemer described: writing checks against value it cannot yet instrument. The pricing changed. The measurement problem didn't.
The seductive line — pay only when it works — assumes "works" is a fact you can observe. In agentic enterprise AI right now, it's a claim someone gets to make. Salesforce just made sure that someone is Salesforce.