Salesforce Shipped Three Agentforce Prices and Killed None of Them

When Salesforce first put a price on Agentforce, it did something bold for a company built on selling seats: it stopped charging per person and started charging per conversation. Two dollars, every time an AI agent held a back-and-forth with a customer.⁵ It was pitched as the future of software economics - you pay for work done, not for logins purchased. On May 15, 2025, Salesforce announced a second way to pay for the exact same product, priced on a completely different meter.¹ And in the same press release, it teased a third.⁷ None of the earlier ones were retired.

The tidy story is that Salesforce boldly moved from per-conversation pricing to flexible credits. That framing is wrong in the most revealing way. Salesforce didn't move from one model to another. It kept the first, added the second, and pre-announced the third - all while the product was still new.⁶ A company that knows how its market wants to buy picks a meter and defends it. A company doing this is not defending a plan. It's placing bets and refusing to close any of them.

The meter changed twice before the market decided how it wanted to buy: a per-conversation charge became a per-action charge became a per-seat charge, and each one measures a different thing

Start with the unit, because the unit is the whole argument. At launch, Salesforce charged $2 per conversation - one price for every use case, whether the agent answered a simple question or ran a ten-step resolution.⁵ That is a clean, legible meter, and it has a problem: a conversation is unpredictable. A buyer approving a budget cannot forecast how many conversations their agents will hold, or how expensive each one becomes when an agent chains together dozens of steps. Flex Credits changed the unit entirely - from the conversation to the action. Each standard action now consumes 20 credits, and with 100,000 credits selling for $500, that works out to about $0.10 per standard action.² Salesforce said so in plain language: it is no longer pricing per conversation, it is pricing per action.³

Notice what moving from conversation to action does. It shrinks the billable unit into something smaller, more granular, and - critically - something a CIO can meter and cap. An action is 'updating a record' or 'answering a question,' a discrete event you can count and budget against.⁴ The shift is a direct answer to the one weakness of per-conversation pricing: unpredictability. But rather than swap one for the other, Salesforce left both running. And in the same announcement that shipped Flex Credits, it previewed 'Agentforce user licenses and add-ons - coming Summer 2025' - the per-seat model it had spent the launch conspicuously moving away from.⁷ The per-user license, once it arrived, was reported at roughly $125 per user per month, though Salesforce itself has not published that figure.⁶

Why three live prices is a symptom, not a strategy: confidence in a pricing model looks like removing options, not accumulating them

Pricing is the single loudest signal a company sends about how well it understands the value it creates. When you know what your product is worth and how buyers consume it, you converge - you find the meter that captures value fairly and you sharpen it. When you don't, you diverge - you offer more ways to pay, because each new option is really a hedge against the possibility that the last one was wrong. Salesforce's own framing gives the game away: the company introduced Flex Credits as an additional option rather than announcing that per-conversation pricing had proven itself and should be replaced.¹ You do not add a second meter to a metric you're confident in. You add it because the first one made buyers nervous, and you weren't willing to bet the entire launch on convincing them otherwise.

The tell is the timing. If per-user licensing was the eventual destination all along, Salesforce could have led with it - it is the model the company has sold for two decades and the one CIOs already trust. Instead it launched on the boldest, least familiar meter, walked it back to a granular consumption model within months, and pre-announced a return to the seat license it had originally left behind.⁷ That is not a sequence a confident plan produces. That is a company iterating its price in public, under the pressure of a market that had not yet told it how it wanted to buy AI labor.

The case that optionality is exactly right for a brand-new market: when nobody knows how to buy a thing, offering every way to buy it can be the smart move

The honest counter is strong, and it deserves saying out loud: nobody - not Salesforce, not its buyers, not its competitors - knows yet how enterprises will want to pay for autonomous AI agents. This is a genuinely new category, and in a new category, refusing to commit to one meter is not weakness; it's humility. Let the small buyer take Flex Credits and pay for what they use. Let the enterprise that hates variable bills take the $125 seat and sleep at night. Let the early adopters keep their $2 conversations. Salesforce is meeting buyers where they are while the market sorts itself out, and it costs almost nothing to run three price lists on the same underlying product.⁶

The counter is right that optionality is defensible - but it is not the same as a strategy, and it is not free. Every additional meter is a tax on the buyer's ability to compare, a burden on the sales team's ability to explain, and an admission that the company cannot yet say what its product is worth in a single sentence. Optionality buys time; it does not buy conviction. The moment the market shows a preference, the discipline will be in killing the two meters that lost - and a company that keeps adding pricing models without retiring any has not yet shown it is willing to subtract one.

Salesforce bet its most-watched AI launch on the idea that software should be priced like labor - by the work, not by the worker. Then, within a year, it hedged that bet twice: once toward a smaller, safer unit, and once back toward the seat license it had left behind. All three prices are still on the table.⁶ That is not the shape of a company executing a vision of how AI should be sold. It is the shape of a company discovering, in real time and in public, that it doesn't yet know - and choosing to keep every door open until the market tells it which one to walk through.