Salesforce now sells its AI agent product three ways at once: per user per month, per credit consumed, and in one case nothing at all until the agent resolves the customer's problem.
Every previous shift in software pricing — per seat, then per unit of usage — measured something the vendor could count on its own systems. Outcome-based pricing asks the vendor to prove its software caused the revenue or the saving, and Laura Bratton says that is the part nobody has worked out.
"I think that it's far too complicated to be the main way that we price AI products."
Bratton writes The Information's applied AI newsletter and reported the Salesforce overhaul, including from a person with direct knowledge of the company's sales strategy who told her customers are now asking for contracts that pay only on results.
I listened to the full segment so you can skip it.
Here are the 5 takeaways that matter.
👤 Guest: Laura Bratton, who writes The Information's applied AI newsletter and reported the Salesforce pricing overhaul this segment is built on
🎙️ Host: Akash Pasricha, who anchors TITV, The Information's live weekday show at 10 a.m. Pacific
📰 Published: 31 August 2026 on YouTube (The Information)
🔴 YouTube | ⏱️ 7 min
Key Takeaways
The startup version of outcome pricing bills only when the AI finishes the job
The unit is a resolved support ticket or a retained subscriber, not a message sent
The radical version bills against the customer's own revenue or cost savings
Palantir has priced this way for years and commands high prices doing it
Salesforce is not picking a model, it is selling three and letting the buyer choose
Nobody has solved attribution, which is what the whole model rests on
Every deal needs a negotiation over whether the software or the business produced the gain
Bratton expects hybrids to win rather than outcome pricing
Plenty of buyers still want a seat-based subscription and a bill they can forecast
1. Pay only when the AI works
Pasricha opened by asking what the outcome in outcome-based pricing actually is and how it is measured.
The mainstream version came from startups and prices a completed task. "I think you know typically what we're hearing from startups that pioneered this model is that they're charging only when the AI actually works."
Her examples were a customer service agent that resolves a support issue, or one that keeps a customer from cancelling a subscription
The change is in what gets counted. "So rather than just, you know, how many messages the AI sends to a customer, it's actually only charging you when AI finishes a task."
2. The Palantir version of it
A second, harder version prices against the customer's own financial results. "There's a more radical form of outcome based pricing that companies are beginning to adopt, similarly to Palantir, which is charging for AI only when it helps you generate a certain figure of sales or when it helps you cut costs by a certain amount."
This is not new, and the evidence that it can work is Palantir's pricing power. "I think that if we look at Palantir as an example, Palantir has been able to command very high prices for its software by charging with this model of you know customers paying for the software when it reaches a certain level of helping them generate revenue or reduce costs."
Pasricha pressed on who absorbs the cost — if customers pay less, either the vendor earns less or this is a roundabout way of charging more. Bratton did not claim to know: she said it is unclear who it helps more
3. Salesforce lets you choose
The interesting move, in her reading, is Salesforce conceding that it does not know the answer either. "But I think what's interesting about Salesforce is that they're sort of admitting that pricing in the age of AI is really complicated and they're letting customers pick how they want to pay for AI."
The menu for Agentforce has a seat price and a consumption price side by side. "You can pay for Agent Force per seat, which is $125 per month per user, or you can pay you know, I think it's like 10 cents per credit for the AI agent."
There is also a premium bundle that comes with flex credits a customer can spend on any Salesforce AI product
Its help agent is the pure outcome product: the customer pays only when it works
A fourth option is being pulled out of the company by its buyers. "And in addition to that, a person with direct knowledge of its sales strategy told me that customers are beginning to want custom contracts where they pay, you know, only when the AI generates a certain amount of revenue or helps them cut costs."
4. Attribution is the hard part
Asked how anyone measures the outcome, Bratton named the problem underneath the whole model.
The question is causation, and it is not answerable from the software's own data. She put it as how to attribute cost savings or revenue gains — whether the result came from the software the company bought or from something the business itself did, such as changing strategy
The consequence is that every contract becomes bespoke. She said it takes a complicated negotiation to settle attribution, case by case, and called it the most difficult thing about the model
That cost falls on both sides before a dollar changes hands
5. Hybrids win, not one model
Pasricha asked whether this is the future the way usage-based pricing once was, or whether there are reasons not to adopt it.
Bratton's answer was that it cannot carry the whole market. "I think that it's far too complicated to be the main way that we price AI products."
Plenty of buyers still want seat-based subscriptions, and she expects vendors to build hybrid usage-and-subscription models the way Salesforce largely has
What she expects to spread is the flexibility, not the pricing model. "I think the future that we're going to see is what Salesforce is doing, which is coming up with a flexible way of pricing its products so that you can pick how you want to pay."
The pitch she described is a vendor quoting what each route would likely cost, and offering to negotiate a custom contract on top
Bonus Insights
The customer evidence points both ways, which is why she would not call it. "You know, like I talked to FedEx, a CIO earlier this summer and he was like, I love outcome-based pricing and then you hear from customers that, you know, or customers of Palantir that think that it's too expensive if you price this way."
Pasricha noted that the show had discussed Marc Benioff talking the model up on Salesforce's earnings call the previous week
Bratton's column on the pricing shift ran that weekend in her applied AI newsletter, and the segment is a walkthrough of it
This was the second of five interviews in one episode; the other four are written up separately
Bratton's bottom line is that outcome-based pricing is real, Palantir has proved it can command a premium, and it will still end up as one option on a menu rather than the way AI software is sold, because nobody can cheaply prove which party produced the outcome.
Products, Companies & Tools Mentioned
Salesforce and Agentforce (The overhaul the segment is about: a seat price, a per-credit price, a premium bundle with flex credits, and a help agent billed only on resolution)
Palantir (The long-running example of pricing against a customer's revenue or cost savings, which she says has let it command very high prices)
FedEx (Whose chief information officer told her earlier in the summer that he loves outcome-based pricing — one half of the split verdict she reports)
Books & Resources Mentioned
Salesforce Is Overhauling the Way It Charges for AI (Bratton's own column, published that weekend, which this segment walks through)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: