Salesforce Booked $800M in AI Revenue Last Quarter. That Money Came From You.

Salesforce’s $800M signal: the meter is changing

Nate opens with the headline that Salesforce’s Agentforce reached an $800 million ARR run rate, up 169% YoY, on 2.4 billion “agentic work units.” His point is that this is not token pricing in disguise — Salesforce is explicitly charging for actions agents complete inside the platform, and that should make every buyer ask what exactly they’re paying for when work moves from a person to a machine.

The old SaaS trick was seats — agents break that math

He frames two decades of SaaS as a simple, hugely profitable conversion of work into seats: 10 people on a CRM meant 10 licenses, easy forecasting, easy multiples. That logic breaks when an agent can read Salesforce, update Jira, trigger ServiceNow, and never “sit” in the software as a human user would — the work still happens, but the old unit of value no longer fits.

Salesforce, Microsoft, and ServiceNow are all adding a second meter

Salesforce is the cleanest case: a rep may still have a seat, but the agent identifying customers, summarizing cases, or triggering workflows also burns flex credits. Microsoft does the same in hybrid form through Microsoft 365 Copilot seats plus Copilot Studio credits for answers, actions, grounding, and “premium reasoning,” while ServiceNow reframes itself as the governed substrate for operational actions like access provisioning and incident escalation. Nate’s throughline is simple: the seat stays, but delegated work gets billed separately.

The pricing jungle is getting harder to forecast fast

He says even modest agent usage at a 100-seat company can become hard to model once runtime credits and premium actions kick in. The anecdote that sticks is his conversation with a developer who burned 8 billion tokens in a month — and Nate’s reaction is basically, that doesn’t even surprise me anymore, which tells you how fast usage is exploding while procurement still lags behind.

SAP shows how agent access can become a contract fight

The sharpest warning comes with SAP’s 2026 API policy, which restricts AI systems that plan, select, or execute API call sequences outside SAP-approved architectures. Nate translates that bluntly: if you want outside agents touching SAP data, your first question is not whether it’s technically possible, but whether your contract even allows it.

What “fair” pricing looks like versus rent-seeking

Nate gives a buyer-side checklist: visible meters, sensible billing units, predictable forecasting, separate treatment for reading vs drafting vs writing vs executing, caps, exportable logs, and no billing failed work like successful work. The bad version is the opposite — vague AI charges, your own data becoming a paid privilege, expiring credits, hidden meters until renewal, and commercial lock-in dressed up as safety language.

Builders need cost-aware agents, not just working demos

He turns directly to developers here: most cost-aware teams still think in tokens, but production agents now need to understand workflow units, contract terms, and which tools are expensive or reversible. His line is memorable because it feels operationally true: an agent that treats every tool call the same regardless of budget is “an incident waiting to happen.”

Don’t wait until renewal leverage disappears

Nate closes with practical negotiation advice: ask up front whether agents acting on behalf of users are covered, whether third-party agents get the same governed path as native ones, whether failed actions count, and whether seat counts should come down if agents absorb real work. Otherwise, he says, you end up in the worst hybrid of all — the old seat bill plus a new, opaque agent-consumption bill that only becomes obvious after the workflows are mission-critical.