Somewhere between $7.40 and $0.62, an entire valuation model died.
That's the spread McKinsey put on a single support ticket this cycle โ $7.40 for a human, $0.62 for an AI agent. Twelve times cheaper. Roughly 92% of the cost, vaporized. And sitting inside that gap is $6.78 that nobody โ not Salesforce, not SAP, not the enterprise buying the ticket โ has agreed on how to split.
I've been mapping settlement pipes since the TerraUSD collapse, when I watched liquidity flee into community-owned DAOs and learned that trust had quietly stopped being algorithmic and become social. This is the same pattern in a different suit. The industry is calling it a pricing debate. It isn't. It's a verification crisis, and everyone is sending the bill to the wrong department.
The framing came from Gartner, which pegged $234 billion in enterprise software spend as exposed to 'agentic arbitrage' by 2030 โ a figure Managing VP George Brocklehurst summarized as agentic AI 'changing software economics.' Read the tense. Not changing. Changed.
The mechanism is simple once you strip the jargon. Traditional SaaS charged per seat โ a license for a human who logged in and did work. When an agent does the work, the seat stops being a unit of anything. Salesforce now sells Agentforce at roughly $0.10 per action and $0.15 per voice interaction. HubSpot charges about $0.50 per session. Zendesk runs $1.50 to $2.00 per resolution. Meta prices its business agent at $2.00 per million tokens. Monday.com is migrating from seats to AI credits. Ema is pitching per-outcome billing to raise money.
Four architectures, one room, zero consensus. And here's the detail that should stop you cold: Salesforce is running all four at once. That is not confidence. That is a hedge โ the strategic equivalent of a company that doesn't know what its own product costs.
One caution before we go further: several of these announcements carry 2026 datelines โ a Gartner forecast, a Monday.com transition, an SAP earnings call. Treat the timelines as scenario, not settled fact, and the structural argument still holds.
The metering problem is older than crypto, but crypto is the only place it has actually been solved โ and almost nobody in the SaaS war is looking there.
Start with the tradeoff nobody wants to say out loud: the finer your billing granularity, the cheaper it is to settle, and the further you drift from real value. Meta's token pricing is the extreme. At $2.00 per million tokens it is precise, automatable, auditable โ and utterly indifferent to whether the agent solved the customer's problem or burned a million tokens failing. Salesforce's $0.10 per action is one step up: it counts effort, not results. Zendesk's per-resolution model gets closer to value but demands that someone define what 'resolved' even means. Ema's per-outcome promise is the holy grail โ bill for the result, not the labor โ and it is also the hardest thing on earth to verify, which is precisely why nobody has shipped it at scale.
The industry is, in the words of one analyst, trapped between wanting value-based pricing and needing measurable units of work. That trap is not a pricing problem. It is an architecture problem: you cannot bill for an outcome you cannot independently confirm. And when both parties to a transaction are autonomous agents โ no human signing off, no ticket closer to argue with โ confirmation stops being a courtesy and becomes infrastructure.
This is where my own failure earns its keep. At NeuralLedger Labs in Austin, we spent four months building a decentralized identity protocol so AI agents could negotiate and execute smart contracts without a human in the loop. It buckled on scalability. But the beta taught me something I've been chewing on ever since: agents can absolutely transact. What they cannot do is agree on whether the transaction succeeded. That gap โ call it the success oracle โ is where the entire agentic economy will bottleneck.
Which is why the most important line in this whole story is the one analysts skipped: settlement activity across Stripe and the 'Chain of Trust' demonstrations has been spiking vertically. That is the tell. Settlement is happening whether or not pricing is solved, because settlement gets paid no matter who wins the metering war. The picks-and-shovels logic holds โ except the shovel is verification, not billing.
Run the four models side by side and the shape of the fight appears. Per-action is cheap to meter and easy to automate, but it rewards volume over value. Per-session is a middle child โ granular enough to bill, vague enough to argue about. Per-token is the finest grain in the room and the least aligned with what a customer actually bought. Per-outcome is the only model that maps to value and the only one that requires a trust layer none of these vendors own. Rank them by metering difficulty and you get a perfect inverse of their value alignment. That inverse is the whole ballgame.
Now the number the pricing debate is designed to bury. AI businesses run gross margins of roughly 50-60%. Traditional SaaS runs 80-90%. That is not a rounding difference. That is a structural repricing of the entire software category, and it has already happened โ it is a financial fact, not a forecast. Any new metering scheme that cannot drag margins back above 70% is just a dignified form of decay.
So follow the $6.78. It has three claimants: the customer, who wants the savings as lower cost; the vendor, who wants it as reprice; and the model provider upstream, who wants it as extraction. The vendor that controls the meter decides the split. That is the real reason Salesforce is running four models simultaneously โ not indecision, but a land grab for the right to define the unit.
And watch the language. SAP's Christian Klein called per-outcome pricing 'a unique opportunity to reset price.' Sit with that verb. Reset. This is not a customer-advocacy story about paying only for what works. It is a margin-defense story โ an attempt to re-anchor billing to value so the 80-90% era can be reconstructed on new terms. Per-outcome is being sold as fairness and bought as margin.
Monday.com's quiet shift from seats to AI credits deserves more attention than it got. It is the first real admission by a mid-cap SaaS that recurring, seat-anchored ARR โ the smoothest revenue a software company can hold โ is being voluntarily traded for consumption revenue, which is spikier, less predictable, and harder to model. On a spreadsheet that reads as financial-quality downgrade dressed as innovation. When a company surrenders the reliability of seats, it is telling you the seats were already leaving.
And beneath all of it sits a standard war, the most underrated force in the entire story. Nobody has agreed on the unit โ action, session, resolution, token, outcome โ and until one wins, every vendor is guessing, every CFO is modeling against a moving target, and every contract is a one-off negotiation. Standards wars are not won by the best architecture. They are won by whoever gets the most counterparties to settle in their unit first. That is a distribution fight masquerading as a pricing fight, and it is why Stripe's settlement spike matters more than Salesforce's four-way bet.
Trace it forward and the second-order effect is obvious. If per-outcome ever becomes the norm, it spawns an entire niche that does not exist yet: on-chain outcome verification โ oracles attesting that a task completed, escrow contracts holding funds until attestation lands, verifiable computation proving the agent did what it claimed. That is the crypto-native land grab hiding inside an enterprise-software headline, and the reports that inspired this piece mention the pipes without ever naming who lays them.
A word on the evidence itself. The $234 billion figure is a Gartner forecast, and forecasts from consultancies get recycled into vendor marketing long after their assumptions rot. The 'pipes already exist' claim carries no citation at all. I have parsed enough S-1 filings to know the difference between a disclosed fact and a hopeful slide, and this story leans heavily on the latter. That does not make it wrong. It makes it a narrative sample, not a research finding โ which, for anyone who trades sentiment, is actually the more useful thing to know.
The crypto-native reader should be uncomfortable right now, because this is the exact spot where Web3 has a genuine, non-decorative role โ and the source material never closes the loop. If outcomes must be verified between autonomous agents, you need escrow, you need oracles, you need verifiable credentials and programmable settlement. That is not a narrative add-on. It is the missing rail under a $234 billion projection. Code breaks. Stories don't. But a story with no verifiable settlement under it is just a press release with a spreadsheet.
Here's the contrarian read, and it cuts against both the bulls and the Web3 tourists.
The crowd is watching Salesforce's four pricing models and calling it strategy. I think the interesting signal is somewhere else entirely: Meta's $2.00 per million tokens, which carries a 40-50x price advantage over the outcome-based crowd. That is not a pricing model. That is a market-clearing anchor, and if it becomes the de facto standard, per-outcome economics get punctured by a price war before they ever mature. The most precise, value-aligned model in the room is also the most exposed to being undercut by the cheapest, crudest one.
The second contrarian note: the entire discourse is mislabeled. This got filed as blockchain and Web3, but strip the settlement veneer and 24 of 25 data points are enterprise SaaS pricing commentary. The genuine crypto surface isn't the pricing architecture โ it's the verification and settlement layer sitting underneath it. Don't buy the chart. Buy the chaos. And the chaos is not on the billing page; it's in the oracle nobody has built yet.
Third: the whole narrative assumes customers will happily pay for outcomes. That assumption is unexamined. If a buyer believes the $6.78 savings belong to them by right, no clever meter survives contact with procurement.
The next time a vendor announces per-outcome pricing, don't ask what it charges. Ask who verifies the outcome โ and who holds the escrow while the two machines argue.
Watch the settlement layer, not the pricing page. Watch who holds escrow, not who publishes a rate card. Because the pricing war will end with a winner, but the verification layer will end with a standard โ and only one of those is worth owning.
So: when autonomous agents settle their own bills, whose signature confirms the work was done โ and what happens to the $6.78 if the answer is nobody's?