There is a question hiding inside the software industry’s repricing that almost no one is asking correctly. The consensus has converged on a tidy narrative: AI agents will replace human workers, seat counts will decline, and per-seat revenue will erode. The prescription follows naturally: transition to outcome-based pricing. Charge for results, not access. Align the vendor’s incentive with the customer’s value.

This is directionally correct but mechanically incorrect. The pricing transition that the software industry now faces is not a business model upgrade. It is a fundamental inversion of the economic structure that made software the most attractive asset class of the last two decades. And the companies that misunderstand this will not merely lose revenue. They will lose the financial identity that justified their valuations in the first place.

The Machine That Per-Seat Pricing Built

To understand what is breaking, you need to see what per-seat pricing actually accomplished. It was never just a billing mechanism. It was an economic architecture.

Per-seat pricing created three properties Wall Street loved. First, predictability: if you know how many employees a customer has, you can accurately model their spend. Second, expansion mechanics: as companies grew headcount, software revenue grew automatically, without a new sales motion. Third, and most critically, near-zero marginal cost. Once the software was built and hosted, the incremental cost of adding user number 10,001 was effectively nothing. This produced gross margins of 80%+ or more, the signature financial characteristic that sets software apart from every other industry.

The entire valuation framework for SaaS was built on these properties. Discount a decade of predictable, high-margin, automatically expanding cash flows, apply a terminal value, and you have a business worth 10 to 20 times revenue. Private equity firms built lending franchises around this math. Public market investors paid premium multiples for this math. The math was the product.

AI does not merely threaten to reduce seat counts. It threatens to dismantle every pillar of this architecture simultaneously.

The Three-Front Assault

Consider what happens when a company like Salesforce, with roughly 150,000 enterprise customers paying per seat, transitions to an outcome-based pricing model. The problems arrive on three fronts at once.

The revenue front. A customer with 100 sales representatives, each paying $300 per seat per month, generates $360,000 in annual recurring revenue. If AI agents reduce the team to 30, and the vendor transitions to per-outcome pricing, the vendor needs to capture more value per outcome than it loses per seat, while the customer’s entire rationale for adopting AI is to spend less. The customer and the vendor have directly opposed incentives at the exact moment the transition occurs. Salesforce’s own journey illustrates the difficulty. When Agentforce launched in late 2024, it charged $2 per conversation. Customers revolted—the metric was unpredictable and disconnected from value. By May 2025, Salesforce introduced Flex Credits at $0.10 per action. By summer 2025, facing continued pushback, it added per-user licenses at $550 per month for unlimited AI usage. Within eight months, Salesforce had cycled through three fundamentally different pricing architectures and still had not converged on an answer.

The margin front. This is the problem hiding in plain sight. Traditional SaaS had near-zero marginal cost per user, which produced those 80%+ gross margins investors worshipped. Outcome-based pricing reintroduces high variable costs through token consumption. Every AI action Agentforce performs, every resolution Intercom’s Fin delivers, and every workflow ServiceNow’s agents execute consumes inference compute that the vendor pays for. The gross margin profile of “Service-as-a-Software” differs structurally from that of SaaS. Data from Bessemer Venture Partners shows scaling AI companies averaging 25% gross margins. Even mature AI-first companies reach only about 60%, far from the 80%+ that defined the SaaS era. 84% of enterprises report that AI infrastructure costs erode gross margins by 6% or more. This is not a temporary scaling problem. It is a permanent feature of a model in which every unit of value delivered incurs a real, recurring compute cost.

The measurement front. Outcome-based pricing sounds elegant in a conference keynote. But it runs into a brutally practical question: what counts as an outcome? Customer support is the easy case—Intercom can charge $0.99 per resolution because a resolution is a discrete, measurable event. The model has worked: Fin now resolves over a million tickets per week, and Intercom has maintained positive gross margins on these interactions. But customer support is the exception, not the rule. What is the “outcome” of an HR platform? What is the measurable result of a project management tool? What about an ERP system, where the value is diffuse, continuous, and impossible to attribute to any single action? For every Intercom, there are fifty enterprise software companies whose value cannot be decomposed into billable events. These companies are being told to adopt outcome-based pricing by analysts who have never tried to define the outcome.

The Intercom Exception and the ServiceNow Problem

The bifurcation in the pricing transition maps precisely onto the bifurcation in the market itself, and this is the insight most analyses miss.

Intercom represents one pole. Its product delivers a discrete, countable outcome: a customer question answered without human intervention. It can price against that outcome cleanly. Its customers can measure the savings directly: each automated resolution replaces minutes of human agent time. The value chain is short and transparent. This is why per-resolution pricing works. The customer and vendor can both see the same unit of value and agree on its value.

ServiceNow represents the other pole. Its platform is a system of record that orchestrates workflows across IT, HR, customer service, and security. The value is not in any single action but in the cumulative effect of having all of those workflows governed by a unified platform. ServiceNow has responded by introducing consumption-based “Assist Packs” for its AI features, effectively layering usage-based pricing on top of its existing subscription model. It has also positioned itself as an “AI Control Tower”—the governance layer that manages and monitors AI agents from any vendor.

This is a revealing strategic choice. ServiceNow is not trying to become an outcome-based business. It is trying to become the platform on which outcome-based businesses run. It is selling trust, governance, and orchestration. None of these can be priced per resolution, but all command premium subscriptions because the alternative is ungoverned AI agents operating unsupervised across the enterprise.

The market is beginning to price this distinction. ServiceNow crossed $10 billion in annual revenue in 2025, reported 98% retention, and achieved $600 million in AI-specific annual contract value, exceeding its $500 million target. Yet its stock still dropped 11% on earnings because investors are uncertain whether its governance positioning can sustain premium multiples in a world where the application layer is compressing.

This uncertainty is rational. The question is not whether ServiceNow is a good business. It is whether the business's financial characteristics—predictable, high-margin, and automatically expanding—survive the transition.

The Gross Margin Identity Crisis

Here is the structural problem that the software industry has not yet confronted honestly.

For twenty years, the defining characteristic of a software company was its gross margin. 80% gross margins meant that for every dollar of revenue, eighty cents was available for R&D, sales, and profit. This separated software from services, hardware, and every other category. It justifies software multiples over service multiples. This is why investors paid 15 times revenue for a software company and 2 times revenue for an IT services firm.

The transition to AI-delivered outcomes compresses this margin structurally. When Intercom charges $0.99 per resolution, it captures that revenue but also incurs the inference cost of the AI agent that delivered the resolution—the tokens consumed by Claude or GPT and the compute required to reason through the customer’s problem. These are real, recurring costs that scale with usage. Software companies that once had the marginal economics of a media company—produce once, distribute infinitely—will increasingly have the marginal economics of a services company: every unit of output has a unit of input cost.

The implications for valuation are severe. If gross margins compress from 80% to 60%, the market will not simply apply a 25% discount to the revenue multiple. It will reclassify the business. A 60% gross margin company is not a software company by Wall Street’s standards. It is a services-adjacent business, and the market will value it accordingly. This is the trap: the transition meant to save software companies from seat-count erosion may cost them the margin profile that justified their premium valuations.

The counter-argument is that inference costs are falling rapidly. OpenAI’s compute margins improved from 35% in early 2024 to roughly 70% by late 2025. Anthropic projects gross margins reaching 77% by 2028. If inference costs fall fast enough, AI-delivered software could eventually recover traditional software margins. But this argument has a timing problem. Inference costs on older models are falling. Frontier models, required for complex enterprise workflows, are becoming more expensive as agentic workloads increase token consumption by 10 to 100 times per task. Companies betting on margin recovery through inference deflation may find the goalposts move faster than the cost curve.

The Cannibalization Arithmetic

The uncomfortable truth at the center of the pricing transition is a math problem that no earnings call has addressed directly.

Consider a mid-market SaaS company with 5,000 customers averaging $50,000 in annual recurring revenue. That is $250 million in ARR at 82% gross margins, generating about $205 million in gross profit. The company trades at 8 times revenue, a $2 billion valuation. This is a healthy SaaS business.

Now, suppose AI agents reduce the average customer’s headcount needs by 40% over three years. Under per-seat pricing, ARR declines to $150 million. The company must capture $100 million in AI-based revenue just to stay flat. At 60% gross margins on AI revenue, it needs even more top-line revenue to maintain the same gross profit.

Meanwhile, the customer’s entire motivation for deploying AI is to reduce cost. The customer is not looking to pay the same amount through a different mechanism. The customer is looking to pay less. This is the fundamental tension: the vendor’s survival requires capturing more value per outcome than it loses per seat, while the buyer’s incentive is precisely the opposite.

Some vendors will try to resolve this by expanding scope, arguing that AI allows them to absorb labor budgets, not just IT budgets, so the total addressable market grows even as per-customer spend on traditional software shrinks. This is the “Service-as-a-Software” thesis, and it is intellectually coherent. If your AI agent replaces a $60,000-per-year employee, you can charge $2,000 per month for it, and the customer still saves money. The TAM expands from IT budgets to labor budgets, a much larger pool.

But there is a catch. If you are competing for labor budgets, you are no longer competing with other software companies. You are competing with staffing firms, outsourcers, and the customer’s own internal capabilities. And you are doing so with a product that has real marginal cost, because every AI action consumes compute. The comfortable oligopoly of the SaaS era, where a handful of vendors divided up the IT budget with limited price competition, gives way to a market where you are bidding against the fully loaded cost of a human, and your margins depend on the spread between what you charge and what inference costs.

What to Watch

The pricing transition will unfold over the next eighteen months, and the leading indicators are not the ones most analysts track.

Gross margin trends on AI revenue. The most important number in enterprise software right now is the gross margin vendors earn on their AI-specific products. If margins stabilize above 65%, the financial identity of software survives, and multiples can re-expand. If they settle at 50% to 60%, the industry gets reclassified and repriced.

Expansion revenue versus seat compression. The telltale sign of a failing pricing transition is a company that reports stable retention while expansion revenue declines. This means existing customers are staying but spending less, so the annuity is intact but shrinking. Watch for companies that quietly shift from reporting net revenue retention to emphasizing gross retention, a classic leading indicator of pricing pressure.

The hybrid pricing convergence. Salesforce’s journey from per-conversation to per-action to per-seat for AI is not a sign of confusion. It is the industry discovering in real time that pure outcome-based pricing does not work for most enterprise software. The likely equilibrium is a hybrid: a platform subscription for access, governance, and data management, plus a consumption layer for AI-delivered work. The companies that find this equilibrium first win. Those who overcorrect toward pure consumption lose predictability and are punished by the market.

The procurement reclassification. When CIOs start routing AI agent purchases through HR or operations budgets rather than IT budgets, the Service-as-a-Software thesis is being put to the test. Until then, software companies competing for labor budgets are fighting with the wrong buyer.

The Nature of the Transition

What is happening to the software industry is not a pricing change. It is an identity crisis.

For two decades, software companies told a specific story about what they were: high-margin, capital-light businesses that produced intellectual property once and monetized it infinitely. This story was true. It was the basis for the premium that software companies commanded over every other sector. Investors did not pay 15 times revenue for software because the products were good. They paid 15 times revenue because the marginal economics were extraordinary.

AI-delivered outcomes change the marginal economics. Every resolution, every automated workflow, every agent action has a real cost. The software company of the future may generate more revenue than the software company of the past. The TAM expansion into labor budgets is real, but it will generate that revenue with a fundamentally different cost structure.

The market is not wrong to reprice the sector to 4.1 times median revenue. It is pricing in the possibility that the financial characteristics that justified premium multiples are being permanently altered. The companies that survive will be those that find a pricing model that preserves enough margin to remain “software” in the market's eyes while capturing enough of the AI opportunity to grow. That is a narrow path, and most companies will not find it.

The sorting continues. But the variable that matters most is not retention, growth, or even AI capability. It is gross margin. It has always been gross margin. And for the first time in two decades, the software industry cannot take it for granted.