Goldman Sachs published a note last Wednesday after meeting with around 40 software companies. The headline finding: companies are “increasingly positioning their AI workflows as selling a unit of labor or a unit of productivity” rather than per-seat access. Salesforce has “agentic work units.” Workday sells credits tied to “units of work.” Sam Altman wants to meter intelligence like electricity.
Clean narrative. Tidy transition. There’s just one problem: the companies actually doing this can’t agree on how.
Salesforce is the most instructive case. In eighteen months, Agentforce has shipped three distinct pricing models — all running simultaneously on the same product. SaaStr documented the timeline: $2 per conversation at launch (October 2024), Flex Credits at $0.10 per action (May 2025), then the Agentic Enterprise License Agreement at $125+ per user per month (late 2025). That last one is, unmistakably, a seat license — wrapped in “digital workforce” language, but structured so CFOs get a number they can budget.
The $2/conversation model collapsed in production because nobody could define what a “conversation” was when a single query triggered eight backend processes. Five thousand deals in the first two quarters; only three thousand paid. The consumption model that followed was more granular but still too unpredictable for enterprise procurement. So Salesforce went back to seats — the very model the industry consensus says is dying.
Agentforce hit $540M ARR by Q3 FY2026, growing 330% year-over-year with 18,500 total deals. But only 8% of Salesforce’s 150,000+ customer base has adopted it. The multi-model approach is working better than any single model did alone, which says something uncomfortable about the state of the transition: the market hasn’t converged on a unit of value.
The data supports this at industry scale. The PricingSaaS 500 Index tracked more than 1,800 pricing changes across the top 500 B2B and AI companies in 2025 — 3.6 changes per company per year. Credit-based models grew 126% year-over-year, from 35 to 79 companies. Seat-based pricing as the primary model dropped from 21% to 15%. But the winner wasn’t consumption or outcome pricing — it was hybrid, surging from 27% to 41%.
Metronome cataloged 50+ AI pricing models and found the same pattern: “single-track pricing models are becoming the minority; hybrid is the norm.” Credits are everywhere, but the word “credits” masks three distinct functions — compute proxies, abstracted value bundles, and access gates. Cursor’s credits map to inference costs. Clay’s credits map to data enrichment actions. Perplexity’s credits gate premium searches. Same word, completely different economics.
Even the terminology is incoherent. Linear launched its Agent into public beta on April 16 — free during beta, with a signal that “high-volume compute capabilities” like automations and code intelligence may move to usage-based pricing at GA. Figma started charging for AI credits in March 2026. HubSpot cut its AI resolution price to $0.50. Every company is experimenting with a different slice of the same problem — and none of them are converging on the same answer.
This matters for the hypothesis. The seat-based model is clearly broken: the data on that is overwhelming, and even Goldman Sachs is telling investors the transition is real. But the hypothesis implies a destination — outcome-based pricing, commoditized software, collapsed margins. What the pricing chaos actually shows is that we’re in a messy middle where the old model doesn’t work, the new model doesn’t exist yet, and vendors are running three experiments simultaneously because nobody knows which one enterprise procurement will actually buy.
The Salesforce AELA is the sharpest evidence. The company that coined “agentic work units” — Goldman’s own example of the labor-pricing shift — looked at pure consumption pricing, watched it fail in enterprise sales cycles, and wrapped AI agents back into per-user licenses. Forrester noted the irony: Salesforce is willing to take short-term losses on these agreements, betting that usage will eventually justify the economics. That’s a bet, not a business model.
Hypothesis update: The transition away from seat-based pricing is verified — 3.6 pricing changes per company per year is not experimentation, it’s structural instability. But the hypothesis needs a timing qualifier: the destination (outcome/consumption pricing) may be correct while the transition takes years longer than the market’s valuation collapse implies. Software companies are being repriced as if the transition is complete. The pricing data says it’s barely started.
What to watch: Whether the AELA model — unlimited AI usage for a fixed per-user fee — becomes the industry standard, which would mean the seat model survived by absorbing AI rather than being replaced by it. That would partially falsify the hypothesis: software doesn’t lose its value if the seat simply becomes more valuable.
Sources: Goldman Sachs via Business Insider, SaaStr, Metronome Pricing Index, Linear changelog, Forrester, PricingSaaS 500 Index