Patrick Stokes, Salesforce's chief marketing officer, described the consumption pricing behind Agentforce to a technology conference in June. "One engineer can generate like $100,000 a month bill without much problem," Stokes said2. He delivered the number in a flat tone and moved to the next slide. Salesforce built that pricing on purpose.
Read that figure beside Salesforce's own results and an easy conclusion follows. Agentforce revenue crossed one point five billion dollars in annual recurring revenue for the quarter ended in July, up more than 240 percent from a year earlier, the company reported in August3. Executives read such growth as proof the technology delivers and buyers want more of it. The growth is real. That reading runs backward.
Salesforce delivered 7 billion Agentic Work Units through Agentforce and Slack by the same quarter, more than 3 billion of them in three months alone, a jump of 97 percent from the prior quarter3. An Agentic Work Unit is Salesforce's name for one task an agent completes, whether that means processing a prompt, finishing a reasoning chain, or invoking a tool. Charge by the unit and revenue climbs with usage, apart from headcount. Growth of 240 percent measures a meter spinning faster. Whether the customer's bill shrinks is a separate question.
More than 3 billion of Salesforce's 7 billion Agentic Work Units came in the quarter ended in July
Agentic Work Units delivered through Agentforce and Slack, in billions, one cube per billion
Each cube is 1 billion units of agent work: magenta cubes came in the quarter ended in July, and gray cubes came in all the quarters before it.
Source: Salesforce Newsroom [3]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value | Note |
|---|---|---|
| Quarter ended in July | 3 billion | More than 3 billion, up 97 percent from the prior quarter |
| All earlier quarters | 4 billion | |
| Whole | 7 billion |
George Brocklehurst, a managing vice president at Gartner, put the shift in one sentence built for a boardroom. "You are no longer buying software primarily for people; you are increasingly buying it for agents," Brocklehurst said, sizing the change at up to 234 billion dollars of enterprise software spending by 20304. His sentence names a customer that skips the contract and ignores the invoice. Agents work. Their owners pay whatever the agents burn through.
Gartner has a second finding for anyone hoping the meter slows as models improve. Inference costs per agentic workflow will climb more than fivefold through 2028, the firm forecast in August, even as the price of a single token keeps falling5. Will Sommer, a Gartner senior director analyst, explained the paradox in one line: "Product leaders cannot rely on more efficient token economics to rationalize AI costs," he said. Cheaper tokens buy more elaborate agents, and elaborate agents spend more tokens.
Gartner expects inference cost per agentic workflow to climb past 5 times today's level through 2028
Inference cost per agentic workflow, as a multiple of today's cost, from Gartner's August forecast
The blue column is what one agentic workflow costs to run today, and the magenta column is Gartner's forecast through 2028, which it puts above fivefold.
Source: ARN [5]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value | Note |
|---|---|---|
| Today | 1x | |
| Through 2028 | 5x | More than fivefold, even as the price of a single token keeps falling |
Marc Benioff called the same arithmetic a virtue. "The value isn't in the token, the value is what our platform does with it," the Salesforce chief executive said in September1. His sentence moves the argument away from cost and onto value, a term only the seller gets to define, and leaves the ceiling unspoken.
Enterprise software has run this play before. Salesforce spent the two thousands persuading corporate buyers to trade owned licenses for a rented seat, priced by the month, cheaper up front and simpler to budget. Twenty years later, the same company is asking the same buyers to trade the seat for a meter, priced by the task and harder to forecast. Both swaps promised a lower bill. Pricing power stayed with the seller each time.
Ask who pays for a bill that resists forecasting, who profits from an uncapped meter, and who carries the risk once the invoice lands. Enterprise software vendors profit twice, once from the agent's price and again from the human role the agent replaces. Finance departments carry the forecasting risk, because a consumption model breaks the budget cycles built for fixed licenses. The support staff and junior analysts an agent now covers watch a productivity gain they helped create get billed to someone else's ledger.
Salesforce's own numbers carry the verdict. The company built its second act on charging enterprises less for a fixed seat, and its third act now charges them more for a meter that resists an easy read. Executives will keep publishing growth rates and calling the result triumph, a fair word from behind the podium. From behind the invoice, the number reads differently. Readers who want the next reckoning with enterprise software's machine age before the earnings call spins it can find it every morning on The Signal, Ai Lately's daily email brief.
Sources
- Lawton, George. "Dreamforce 2026 - what Salesforce's AWU experiment means for agentic licensing." diginomica, Sept. 15, 2026. https://diginomica.com/dreamforce-2026-what-salesforces-awu-experiment-means-agentic-licensing
- Lauchlan, Stuart. "Tokenomics - the Salesforce worldview as CMO Patrick Stokes seeks the ideal 'Goldilocks' solution to agentic pricing." diginomica, June 10, 2026. https://diginomica.com/tokenomics-salesforce-worldview-cmo-patrick-stokes-seeks-ideal-goldilocks-solution-agentic-pricing
- Salesforce. "Salesforce Delivers Record Second Quarter Fiscal 2027 Results." Salesforce Newsroom, Aug. 26, 2026. https://www.salesforce.com/news/press-releases/2026/08/26/fy27-q2-earnings/
- Swain, Gyana. "Agentic AI puts $234B in enterprise SaaS spending at risk, Gartner says." CIO, July 2, 2026. https://www.cio.com/article/4192242/agentic-ai-puts-234b-in-enterprise-saas-spending-at-risk-gartner-says.html
- Guan, Lilia. "AI agents aren't cheap, even with falling model costs: Gartner." ARN, Aug. 18, 2026. https://www.arnnet.com.au/article/4210805/ai-agents-arent-cheap-even-with-falling-model-costs-gartner.html
