Last week I sat in a pricing review where a founder put two numbers side by side. Salesforce Agentforce: ten cents per action. Meta's Business Agent: two dollars per million tokens. Then, on a slide labeled only "vision," a startup called Ema promising to bill solely when an outcome actually landed.
Three numbers. Three completely different theories of what an AI agent is worth.
I have been auditing token economics since 2017, back when I ran blockchain literacy circles in a university library in Hangzhou and manually broke down whitepapers for people who just wanted to understand the mechanism. That habit — read the actual mechanism, not the pitch — is why those three numbers bothered me. Because the entire agentic commerce debate has been framed as a pricing question. It isn't. It's a measurement problem wearing a pricing costume, and the meter that would settle it does not exist yet.
That gap is where the real story lives — and where the Web3 label on this trend has been stuck to entirely the wrong layer.
Let me set the table. Enterprise software has priced itself on "seats" for two decades: you pay per human who logs in. When AI agents start doing the work those humans used to do, the seat stops making sense. You aren't paying for a person's access; you're paying for autonomous output. So vendors are scrambling to invent new units of account.
Gartner frames the stakes at $234 billion in enterprise software spend exposed to "agentic arbitrage" by 2030 — not new money, but existing money that gets reallocated. McKinsey's numbers explain why the arbitrage is real: an AI-handled support ticket costs roughly $0.62, against $7.40 for a human. Twelve times cheaper. That $6.78 gap is the entire battlefield.
Right now the industry is running five pricing experiments at once. Salesforce offers four models simultaneously — per action ($0.10, or $0.15 for voice), per session, per outcome, and a hybrid. HubSpot charges around $0.50 per session; Zendesk bills $1.50–$2.00 per resolution. Monday.com has moved to "AI credits" layered on top of seats. Meta undercuts everyone at $2 per million tokens. And Ema raised money on the promise of pure outcome-based billing.
Underneath all of it sits the settlement question. Chain of Trust demonstrations, Stripe's settlement activity — these point to a plumbing layer where autonomous entities transfer value. That plumbing is the only place the crypto vocabulary genuinely applies. Everything else is SaaS economics in a new costume.
Here is the technical heart of it, and it's a classic tradeoff: measurement granularity versus value alignment.
Think of a dial. At one end, per-token billing. Meta's $2 per million tokens is the finest granularity possible — every unit of compute is counted, settlement is trivial, and it is cheap and automatable. At the other end, per-outcome billing. It aligns perfectly with what the customer actually wants, a solved problem, but it is nearly impossible to verify objectively.
The middle is where everyone improvises. Per-action is coarse but easy to meter. Per-session and per-resolution sit somewhere between. The problem is that "an outcome was achieved" is not a measurable event the way "a token was generated" is. Someone has to define it. And whoever defines it is both player and referee.
This is the verification gap, and it's the same flaw that has kept soulbound tokens a concept for three years — nobody wants an immutable record if the record's meaning is contested. An SBT that says "this person completed a course" only works if everyone agrees what "completed" means. An agent invoice that says "this result was delivered" has the identical weakness, just with more money on the line.
So the industry is trapped. George Brocklehurst at Gartner put it precisely: vendors are caught between the desire for value-based pricing and the need for measurable units of work. You can have a meter, or you can have meaning. Rarely both.
Now follow the margin. Traditional SaaS ran at 80–90% gross margins. AI businesses are running at 50–60%. That is not a pricing tweak — it's a permanent re-rating of the entire software sector, and no clever billing model fixes it unless it pushes margins back above 70%. Per-outcome billing, dressed up as customer alignment, is at its core a margin-defense play. SAP's CEO called it "a unique opportunity to reset pricing." Read that again. Reset upward. The customer was never the beneficiary of that framing.
And then there is Meta's token price. At $2 per million, it carries a forty-to-fifty-times cost advantage over per-outcome players. If token-level pricing becomes the de facto standard, it doesn't merely compete — it anchors the market clearing price and makes "precise and expensive" lose to "cheap and coarse." Cheap and rough beats accurate and pricey. That is the unspoken conclusion sitting inside the pricing tables, and almost nobody selling outcome-based dreams wants to say it aloud.
I watched a version of this dynamic up close in 2025, when I helped draft a community governance proposal after the ETF approvals landed. Fifteen town halls, developers and institutional investors in the same room, and the recurring tension was identical: whoever controls the definition of a unit controls the economics of everything downstream. When I published my 2026 series on AI agents and on-chain identity, I argued for human-in-the-loop verification — precisely because autonomous billing without human-verifiable ground truth is just a faster way to be wrong at scale.
The consensus story is that we are watching a pricing revolution. I think we are watching something narrower and more honest: a profit pool being reallocated, not expanded.
The $234 billion isn't new spend. It's existing spend being arbitraged. Which means the real question isn't "what is the right price for an AI agent?" It's "who keeps the $6.78?" The vendor wants it through a new unit of account. The customer wants it as savings. The model provider wants it as an upstream toll. Somebody loses — and it is probably the software layer in the middle, squeezed between a powerful upstream and customers whose bargaining power just went up twelvefold.
Here is the part the crypto crowd keeps missing. The Web3 hook in this story was never the pricing model. It's the settlement and verification layer underneath. If agents transact autonomously, they need programmable escrow and independently verifiable results — oracles, verifiable computation, on-chain attestation. That is a genuine, natural fit for decentralized infrastructure. But the analysis everyone is citing never closes that loop. It gestures at "Chain of Trust" and stops.
We don't get to call something Web3 just because money moves between machines. Bridges aren't decentralized because they connect two banks.
So watch the meter, not the marketing. The vendor that wins won't be the one with the cleverest pricing slide — it will be the one that makes "an outcome happened" verifiable by someone other than itself. That's a trust problem before it is ever a billing problem. And trust, in the end, isn't compiled, verified, and shared by decree.
The industry is racing to price autonomy before it has agreed on how to prove it. Code is only as strong as the trust it protects — and right now, the meter protects nothing.