Nobody Selling AI Agents Can Agree on How to Charge for Them
Salesforce has shipped three different pricing models for Agentforce, its AI agent platform, in roughly eighteen months, and currently runs all three at once: a $2-per-conversation rate from the original 2024 launch, Flex Credits billed at $0.10 per agent action introduced in mid-2025, and $125-or-more-per-user-per-month seat licenses added late in 2025 under a "digital labor" framing, according to reporting from SaaStr, Salesforce Ben, and TechRadar. Agentforce has reportedly reached $540 million in annual recurring revenue, but only around 8% of Salesforce's more than 150,000 customers have adopted it.
That's one company, still undecided on its own product's business model years after launch. Zoom out to competitors in adjacent corners of the same enterprise AI agent market, and the disagreement gets starker rather than converging.
Four companies, four incompatible answers to the same question
Sierra AI, the customer-service agent startup founded by former Salesforce co-CEO Bret Taylor, has gone the opposite direction from Salesforce entirely. It refuses seat-based pricing on principle and charges only for resolved outcomes, custom-quoted per client with no public rate card, reportedly landing around $150,000-plus in typical annual contracts, according to the company's own blog. Glean, which sells enterprise search and knowledge agents, uses a hybrid model instead: roughly $45 to $50 per user per month as a base subscription with AI features layered on as an add-on, a structure that helped it cross $200 million in annual recurring revenue at a $7.2 billion valuation. Harvey, the legal AI company, charges per seat but at wildly different rates depending on client size, $100 to $200 per user per month for large law firms versus $1,000 to $2,000 per user per month for mid-market firms, and has scaled from roughly $100 million to somewhere between $190 and $300 million in annualized revenue within about a year, per CNBC's reporting on its $11 billion valuation round.
Four well-capitalized companies, each targeting a different enterprise vertical, have landed on four structurally different answers: per-conversation-then-per-action-then-per-seat, pure outcome-based, hybrid seat-plus-add-on, and seat-based-but-wildly-variable-by-segment. That's not four companies converging on best practice with minor variations. It's four companies that have not agreed, at a basic level, on what unit of value an AI agent actually sells.
The chaos is the signal, not the noise
In a mature software category, pricing model disagreement this severe this late usually means one of the models is wrong and will lose. In this case, the more likely explanation is that none of these companies has actually solved the underlying measurement problem: how much value does an agent produce relative to the human task it's meant to replace, and how do you charge for that without either overcharging light users or undercharging heavy ones. Per-seat pricing assumes the agent behaves like software, cheap at the margin regardless of use. Per-outcome pricing assumes it behaves like a service, priced on what gets delivered. Per-action pricing tries to split the difference and, per Salesforce's own experience, seems to satisfy nobody enough to have replaced the other two.
Salesforce running three models simultaneously, rather than picking one and killing the others, is the clearest evidence that this isn't a solved problem the industry is still catching up to. It's an open problem major players are actively hedging against, four years into the current agentic AI cycle, in a category still being marketed as production-ready. That gap, between confident go-to-market messaging and unresolved internal pricing logic, describes a category that has proven it can build convincing demos faster than it can prove what a customer should pay for what actually gets delivered.
Sources: SaaStr · Salesforce Ben · TechRadar · Sierra AI · Glean · CNBC · Harvey · getmonetizely