Hook
Over the past seven days, three enterprise software vendors quietly repriced their AI agents. Salesforce began billing at $0.10 per action and $0.15 per voice interaction. SAP's chief executive told analysts that outcome-based pricing represented "a unique opportunity to reset prices." Meta undercut the entire field at $2.00 per million tokens. Not one of them published a verification mechanism.
That absence is the story. An industry is migrating from per-seat licensing to per-outcome billing while holding no objective method to confirm that an outcome occurred. The pricing architecture has moved ahead of the measurement layer, and the gap is where the losses will hide.
I have dissected billing and custody contracts before. In 2024 I spent 200 hours inside the custody architecture of three Bitcoin ETF applicants and found a multi-party computation implementation that exposed 0.05% of assets to single-point failure. The flaw was never the cryptography. It was the assumption that a trusted intermediary had closed a loop it had merely papered over. Agentic commerce is repeating that mistake at the scale of an entire software sector.
Context
The per-seat model is dying, and the obituary is financial, not philosophical. Traditional SaaS ran at 80-90% gross margins. The AI-native replacements are running at 50-60%. That 20-to-40 point compression is the only figure in this debate that is already true rather than forecast.
The mechanics are unforgiving. A human support ticket costs $7.40. An AI agent resolves the same ticket for $0.62. That is a 12-fold reduction, and it removes the rationale for charging per human seat. When the labor you were licensing software to disappears, your license disappears with it. Nobody has agreed on who keeps the $6.78 saved per ticket — the customer, the vendor, or the upstream model provider — and that unanswered question is the whole negotiation.
Vendors have answered with a pricing arms race, not a technology breakthrough. Salesforce runs four models simultaneously: per-action, per-session, per-resolution, and hybrid. Monday.com moved to AI Credits plus seats. HubSpot charges $0.50 per session. Zendesk charges $1.50 to $2.00 per resolution. Ema promises per-outcome billing on the strength of a funding round. Meta prices by the token at $2.00 per million.
Gartner estimates $234 billion of enterprise software spend will be exposed to agentic arbitrage by 2030. Read that number carefully. It is not a growth forecast. It describes an existing pool being reallocated — arbitrage, not expansion. And note the provenance: a consulting projection, the type of headline figure that vendors quote in decks and never audit. Forecast inflation is a real phenomenon, and $234 billion is a number built to be cited, not verified.
The incumbents' only real defense is the record system — the ledger of truth that customers cannot easily migrate away from. That lock-in explains SAP's and Salesforce's pricing leverage, and it is eroding precisely as agents abstract the interface away from the record. The labor displacement underneath remains the genuine demand signal. Twelve-fold cost reduction is not a pitch; it is a measured unit economics shift. That is what separates this from the average over-hyped cycle. The demand is real. The business model is not.
Core
Strip the marketing and four billing architectures remain. Every one of them trades metering precision against value alignment. You cannot have both.
Per-action billing — Salesforce's $0.10 per action — is granular and trivial to automate. It is also value-blind: ten trivial actions cost the same as one decisive one. Per-session and per-resolution billing, HubSpot's $0.50 and Zendesk's $2.00, sit in the middle, balancing granularity against billing complexity. Per-token billing, Meta's $2.00 per million, is the finest grain available and the least aligned with value: it meters compute, not results.
Per-outcome billing is the inverse. It aligns perfectly with value and fails on verification. "Outcome" is the hardest quantity in the stack to define objectively, and it is the only one no vendor has shipped at scale. Ema's promise is a fundraising artifact until an auditor, not a marketer, confirms it.
The contradiction the industry will not name is this: the finer the metering, the easier the settlement and the further it drifts from real value; the closer you get to value, the harder it becomes to verify and bill. Token metering is effortless to invoice and prices nothing that matters. Outcome metering prices everything that matters and cannot be invoiced. Per-action is the only architecture that is both billable today and defensible tomorrow, which is precisely why Salesforce is hedging across all four rather than committing to one.
Then there is Meta. At $2.00 per million tokens it holds a 40-to-50x price advantage over the fine-grained competitors. That is not a feature; it is a market-clearing price waiting to be discovered. If token-level pricing becomes the reference, every high-margin outcome-based strategy gets arbitraged toward the compute floor. The narrative that outcome pricing restores software margins runs directly into a vendor willing to sell at cost. This is the unspoken mortal inference: the most precise pricing model does not win. The cheapest billable one does.
The margin damage is structural, not transitional. Salesforce running four models at once is not confidence; it is hedging by a company that does not know which metering standard wins, so it sells all of them. SAP's "reset prices" language is equally transparent — a price increase dressed as value alignment. Monday.com's shift from seats to AI Credits is a quiet downgrade in financial quality: recurring, predictable ARR surrendered for volatile consumption revenue. That is a lower-multiple business wearing the same ticker.
Now the part the "Web3" label obscures. Per-outcome billing cannot function without a verification and settlement layer that neither counterparty controls. If the vendor defines the outcome, the vendor is both player and referee, and outcome billing degenerates into per-action billing with extra steps. Objective, dispute-resistant confirmation that a result was achieved requires escrow, verifiable computation, and an oracle — the exact primitives this sector claims to be about. Chain of Trust and Stripe's settlement activity appear in the pipeline, yet the logic is never closed. That is the argument gap where the actual money is.
Regulation is not absent here; it is merely unbuilt. The Howey test does not apply, because there is no token and no investment contract. But the moment autonomous agents move real value between themselves — the cited "value transfer between autonomous entities" — anti-money-laundering rules, money-transmitter licensing, and sanctions screening attach to the settlement rail, not the pricing page. Regulations are lagging, not absent, and the compliance bill arrives the day the pipe carries value. No vendor in this pricing war has provisioned for it.
This is a metering standard war in its unresolved phase. Standards wars rarely reward the most elegant architecture. They reward the cheapest one that settles. Token pricing may beat outcome pricing for the same reason VHS beat Betamax: it is crude, cheap, and instantly invoiceable.
Contrarian
Here is what the bulls have right, and I will not pretend otherwise. The cost substitution is real and already delivered. $0.62 versus $7.40 is not a projection; it is a measured 91.6% cut in unit cost, and demand for that cut is not speculative.
The genuine beneficiary is the settlement layer. Whatever billing model wins — action, session, token, or outcome — the transaction must still clear. Stripe's settlement volume climbing vertically is the tell. The picks-and-shovels argument holds precisely because it is model-agnostic. This is the one place where the crypto framing is not decorative.
And if outcome-based billing ever scales, it will require on-chain verification. Oracles, verifiable computation, and escrow contracts are the natural infrastructure for a claim neither counterparty can be trusted to adjudicate. That is a real niche, and the pricing debate missed it entirely.
But do not confuse a niche with a thesis. Liquidity vanishes; insolvency remains. The vendors bleeding margin today will not be rescued by a settlement layer that charges them per transaction.
Takeaway
The industry is solving the wrong problem. It is fighting over how to price work it cannot yet verify. Past performance predicts future panic, and the panic here is a 50-60% gross margin dressed as a business model. Watch for the first vendor to publish a settlement audit rather than a pricing page. Until an outcome can be independently confirmed, "per-outcome" is a discount coupon with a legal disclaimer. And the moment these agents custody real value, the regulatory clock starts whether or not anyone reads the terms. Check the source code, not the hype.