Here is the data point that should stop every enterprise software investor cold: AI-native business models are running at 50-60% gross margins. Traditional SaaS runs at 80-90%. That is not a rounding error. That is a structural rewrite of the most profitable business model the software industry has ever produced, and almost nobody is pricing it correctly.
The framing you will see in most coverage is about pricing philosophy — per seat versus per action versus per outcome. That framing is comfortable because it sounds like a product decision. It is not. It is a margin survival question dressed up as a UX debate, and the arithmetic underneath it is brutal.
Let me lay out the cost structure before we talk about anything else. McKinsey's service-operations data puts the cost of an AI-handled work ticket at roughly $0.62. The same ticket handled by a human costs $7.40. That is a 12x delta. Ninety-one point six percent of the cost of that unit of work simply evaporates when you move it from a person to an agent.
The immediate reaction from most people is: great, the customer saves money. And they do. But that $6.78 per ticket is not free money floating in the air. It is a contested pool. Somebody captures it. The question that nobody in this industry has answered — and I mean nobody — is who.
I have spent enough time inside leverage structures to know what happens when a spread like that opens up and no one has defined the settlement mechanism. You get a land grab disguised as product innovation. That is exactly what the per-action, per-resolution, per-outcome, per-token pricing debate actually is.
The four architectures and what each one really costs you
Salesforce came out with per-action pricing on Agentforce — ten cents an action, fifteen cents for voice. HubSpot went per-session at fifty cents. Zendesk landed between $1.50 and $2.00 per resolution. Ema is promising per-outcome. Monday.com is running a hybrid of seats plus AI credits. And Meta is out there pricing its business agent at $2.00 per million tokens.
Look at that table and tell me it is a technology roadmap. It is not. It is four companies in various states of panic, testing which measurement unit the market will tolerate.
Understand what per-action pricing means mechanically. You are counting discrete operations. It is easy to meter, easy to automate, easy to reconcile. The problem is that it is completely decoupled from whether the action accomplished anything. An agent that fires ten actions and solves nothing bills the same as an agent that fires ten actions and closes a case. The vendor gets paid for effort, not for outcome, which is the exact opposite of what the buyer wants.
Per-resolution and per-session are a marginal improvement. You are at least tying the meter to a unit of work that has a terminal state. But the granularity drops and the dispute surface grows. What counts as a resolution? If a ticket gets reopened, does the meter rewind? Who adjudicates that? At scale these arguments become a support cost center of their own.
Per-outcome is where the industry says it wants to go, and it is where the technical problem actually lives. If you charge for results, you have to define the result. If you define the result, you are the party being measured. That is not a pricing model, that is an unfalsifiable claim. Unless the outcome verification is external and cryptographically verifiable, per-outcome pricing degrades into per-action pricing with worse PR.
This is the part I find genuinely interesting, and the part almost every analysis I have read skips right past. The only way per-outcome works as a trustable primitive is if you have a settlement layer where the achievement condition is objectively checkable by both sides. Escrowed value, an oracle or verification function that neither party controls unilaterally, and a payout that executes on a condition rather than on a vendor's assertion. That is a protocol problem, not a billing-UI problem. And the moment you frame it that way, agentic commerce stops being a SaaS story and starts being an infrastructure story — with all the same verification gaps that decentralized finance has been grinding on for years.
The number that breaks the whole narrative
Meta at $2.00 per million tokens is the single most under-discussed fact in this entire conversation.
Run the comparison honestly. If your competitor will do the work at token-level pricing with roughly forty to fifty times the price efficiency of your per-outcome premium tier, your premium tier is not premium. It is exposed. You can talk about accuracy, about enterprise integration, about being the system of record — and some of that holds for a while, because switching costs are real. But price anchors are downward-sticky in the opposite direction of what you want. The cheap option sets the reference point. The expensive option has to justify every dollar above it, every quarter, forever.
This is what I mean when I say cheap and crude beats precise and expensive in a commoditizing market. Token-level metering has terrible value alignment — it prices compute, not consequence. But it has an unbeatable property: it can be measured, automated, and defended in a procurement conversation. Per-outcome marketing sounds better in a keynote and dies in the spreadsheet.
The efficiency delta is also asymmetric in a way the industry has not internalized. That 12x cost reduction from $7.40 to $0.62 does not just help customers. It destroys the vendor's justification for seat-based licensing, because the unit that used to justify a license — a trained human being sitting in a chair — is precisely the unit that is disappearing.
Read that again slowly. The old model charged for headcount. The new model eliminates headcount. The vendor's revenue basis is the thing its own product removes. You cannot hedge that. You can only find a new basis or shrink.
What the gross margin does to the valuation model
The 50-60% versus 80-90% gap is not a transitional artifact of AI companies being young. It is a cost-structure fact. Inference costs money that license renewals did not. Every agent action consumes compute. The marginal cost per customer goes from near zero to meaningfully positive, and that flips the entire economic shape of the business from a high-multiple recurring-revenue story into something that resembles a services business with better software.
Here is what that does to a valuation. Software companies trade at revenue multiples that only make sense because gross margins are in the high eighties and because seats expand over time. Take away seat expansion — because AI reduces the number of seats — and lower the gross margin at the same time, and you lose both multiples simultaneously. The market has a name for that. Investors have been trained on software for two decades and are structurally slow to reprice gross margin compression in names they consider defensive.
From my audit background, I will tell you the pattern I look for when a company starts offering four pricing models at once. That is not optionality, that is a confession. It means the company does not have a unified data model for what an agent produces, so it cannot price it with confidence. Salesforce offering per-action, per-resolution, hybrid, and outcome-based tiers in parallel tells me internal billing telemetry is not consolidated enough to bet on one. That is a technical debt signal expressed as a product strategy.
And SAP's CEO calling outcome-based pricing a unique opportunity to reset prices — that is the tell. Listen to the verb. Reset. Not align, not pass through savings. Reset prices. The intent behind per-outcome pricing from incumbents is margin recovery, not customer alignment. It is an attempt to reconnect price to a defensible value claim, and it is being marketed as customer-friendliness because that is the only framing that survives a procurement review.
The missing user
Across everything I examined here, there is a voice missing that should be the loudest.
Where is the customer saying they want to pay per outcome? Nobody in the public record of this debate is a buyer making a purchasing commitment on those terms. Gartner's $234 billion headline figure for enterprise software spend exposed to agentic arbitrage by 2030 is a useful directional number, but it is a forecast drawn from a consulting model, not a signed contract. And that distinction matters enormously, because it is the difference between a narrative and a market.
There is a structural tension at the heart of this that the industry keeps skipping. The customer's benefit from an agent is that their cost drops from $7.40 to $0.62. The vendor's survival depends on capturing a meaningful portion of that drop. These are in direct opposition. A customer who has just understood that the work now costs twelve times less has a strong argument that the savings belong to them. Any pricing model that tries to capture value close to the old human cost will be met with the obvious counterargument.
The only durable answer is that the vendor must move up the value stack into areas where verification is unambiguous and the alternatives are worse. Deep vertical integration where the agent has access to domain context a generic model does not. That is where margin survives. Horizontal agents that just wrap a language model will be priced at the marginal cost of the underlying tokens, because in a market with a clear substitute, that is what they are worth.
What I would watch, not what I would guess
I trade structure, not narrative. So here is the structure.
First, watch whether any major vendor reports gross margin in a way that isolates their AI-attached revenue from their legacy licensing revenue. Until they do, the 50-60% number is an industry aggregate and the actual dispersion is unknown, and dispersion is where the opportunities are. A vendor at 72% is in a wildly different position from a vendor at 48%.
Second, watch the settlement layer. Payment infrastructure processing agent transaction volume is the single cleanest read on whether agentic commerce is real. If agent-to-agent payment volume is accelerating while enterprise software revenue is flat, that confirms the thesis that value is moving from the software layer toward the infrastructure and settlement layers. The people selling shovels get paid regardless of which pricing architecture wins, because settlement happens in every one of them. I would rather own the pipe than bet on the meter.
Third, watch the token price. Every dollar that drops from a million tokens pulls the market clearing price down with it, and drags the premium per-outcome tier closer to obsolescence. The anchor is set by the cheapest credible substitute, not by the most articulate one.
What I am not going to do is pretend that I can tell you which pricing model wins. I cannot, and neither can Gartner, and neither can a CEO who has a book of business built on the old model. Trust is a variable I solve for, never assume. So I will watch the filings, I will watch the settlement volume, and I will watch which vendor blinks on price first when a large customer refuses the outcome tier.
The market does not owe anyone an exit, only a price. The enterprise software industry has spent twenty years being paid for seats. Those seats are the thing the new product eliminates. Until somebody on a stage explains, with verifiable numbers, who captures the $6.78 and how the settlement is enforced, every pricing architecture on offer is speculation with a spreadsheet attached — and the spreadsheet is starting to show red.
Security is not a feature; it is the foundation. In this case the security question is not about exploits. It is about whether the pricing mechanism can be trusted by both parties. A meter that only one side can read is not a meter. It is an assertion. And the entire agentic commerce narrative is currently built on assertions about value that have not yet survived contact with a customer who is doing the math on the other side of the table.
