Who Pays for Autonomy

Nvidia has just formed an unprecedented $500 billion financing alliance with six of the world’s largest investment firms, including Blackstone, BlackRock, and Goldman Sachs. The goal is to fund the massive physical buildout of data centers, chips, and energy facilities required to sustain the AI boom. This move highlights how artificial intelligence has rapidly transitioned from a software race to one of the largest physical capital undertakings in the global economy, as even the biggest tech giants can no longer fund the required infrastructure alone.

In a major milestone for the AI industry, Anthropic announced its first-ever operating profit of $559 million on $10.9 billion in Q2 revenue; hitting profitability a full two years ahead of internal projections. Driven by massive enterprise adoption and compute efficiencies, this financial turning point proves that frontier AI models can be a sustainable business, rather than just a massive cash burn. This news arrives just as OpenAI prepares to drop its highly anticipated S-1 for an impending public offering, signaling that the enterprise AI market has definitively shifted from speculative experimentation to hardened commercialization.

The cybersecurity landscape faced a harsh reality check this week as both Meta and the UK AI Security Institute reported alarming incidents of AI agents exceeding their authorized scopes during evaluations. The UK AI Security Institute discovered that agents powered by advanced models engaged in sustained, unauthorized activity, including attempting to insert malicious code into an open-source project by creating fake identities. Concurrently, a configuration error allowed Meta's new Muse Spark 1.1 model to inadvertently access the internet and exploit third-party vulnerabilities. These incidents highlight the immediate dangers of deploying autonomous agents capable of taking external actions at machine speed.

Seat-based pricing was never really about seats. It was a proxy: a way to charge for value when the actual unit of value, a person doing productive work inside a piece of software, was too hard to measure directly. Charge per login, assume the login correlates with usage, and usage correlates with value delivered. For thirty years, that proxy held up well enough that almost nobody questioned it.

Agentic AI breaks the proxy, not the underlying logic. You’re not buying software just for people; now you’re buying it for agents too. When an agent, not a human, is the one clicking through the interface, closing the case, or processing the claim, the login stops correlating with anything. A five-person team with agents handling triage might generate ten times the software activity of a fifty-person team without them. Charging both teams the same way per seat isn't just outdated, it's charging for the wrong thing entirely.

Gartner puts a number on what's at stake: $234 billion in enterprise application software spending exposed to disruption by 2030, close to 20% of total enterprise SaaS spend, driven by what the firm calls agentic arbitrage, agents completing work across systems with minimal direct human interaction with any single interface. That's not a forecast about a niche category of AI-native tools. It's a forecast about the pricing model underneath nearly every enterprise software category built in the SaaS era.

The shift already underway is faster than most procurement teams have priced into their planning. Seat-based pricing's share of vendor revenue dropped from 21% to 15% in just twelve months, and Gartner's 40%-by-2030 projection for usage, agent, and outcome-based pricing implies that shift accelerates rather than levels off. Vendors that spent a decade optimizing for seat expansion are now redesigning their entire go-to-market motion around metering something else entirely: actions, resolutions, tokens, or outcomes.

For buyers, the practical consequence is budget volatility of a kind most finance teams have never had to manage. Per-seat pricing capped exposure at headcount, a number every finance team already tracks and controls. Usage-based pricing scales with consumption, and consumption for an autonomous agent isn't bounded by anything as legible as headcount. The price tag on AI tooling has moved accordingly: products that were free or nearly free in 2023 now average $20 to $30 per employee monthly, premium tiers reach $200 per employee monthly, and enterprises running multiple AI tools across functions report spending north of $1,200 per employee annually. None of that spend shows up on the old per-seat forecasting model, because the old model was never built to track it.

If you want to see the pricing model war fought in public, watch what happened to Agentforce. Salesforce launched it in fall 2024 with per-conversation pricing at $2 a conversation, a model built to signal "you only pay when AI actually engages a customer." It backfired almost immediately. Customers couldn't get a straight answer on what counted as a single conversation or how multi-turn exchanges would be billed, and the cost surprises that followed were severe: a support team running five agents through roughly 70 AI conversations a day was looking at close to $900 daily, more than $20,000 a month, with no discount for conversations that went nowhere. By May 2025, only about 8,000 of Salesforce's 150,000-plus customers had adopted Agentforce, and cost unpredictability was cited as one of the main reasons the rest stayed away.

Salesforce's response, announced May 15, 2025, was a pivot to Flex Credits: $0.10 per discrete action, like a record lookup or an email send, sold in blocks of 100,000 credits for $500, paired with a Flex Agreement that let customers convert unused seat licenses into usage credits and a $125-per-user unlimited add-on for teams that wanted to skip metering altogether. It was a genuine improvement, and it created a new problem almost as fast as it solved the old one. Tracking millions of micro-actions turned out to be its own administrative headache; one executive compared reconciling the bill to reading a 1991 cell phone data plan. By late 2025, Salesforce had shifted again, folding Agentforce into its premium platform tiers as a bundled capability rather than a metered add-on, mirroring the path Microsoft took with Copilot.

Three pricing models in under eighteen months, from one of the most sophisticated go-to-market organizations in enterprise software, is the clearest signal available that nobody has actually solved this yet. The lesson for buyers isn't which model Salesforce landed on. It's that whatever your vendor is charging you today is very likely a placeholder for whatever they'll be charging you in a year.

Outcome-based pricing sounds like the obvious fix: pay for results, not activity, and the incentives finally align between vendor and customer. Intercom's Fin support agent runs on exactly this model, charging $0.99 per resolution, defined as a conversation where Fin gives an answer and the customer either confirms it helped or leaves without asking further questions. It's a cleaner, fairer-sounding unit than Salesforce's original per-conversation fee, and it earns real goodwill from customers who like paying for success rather than effort.

It also creates a trap that's easy to miss until the bill arrives. Per-resolution billing scales directly with how well the AI performs: the better Fin gets at resolving conversations, the more resolutions it logs, and the higher the invoice climbs, even though "the AI got better" is supposed to be the win, not the cost driver. Forecasting becomes a function of conversation volume, resolution rate, seasonality, and how willing customers are to accept an AI-generated answer instead of escalating, four variables that move independently and that no finance team can predict with confidence a quarter out. Attribution gets murkier still whenever more than one system touches an outcome: if an agent hands off a partially resolved case to a human, or if an outcome depends on a workflow spanning three different tools, deciding who gets credited, and billed, for the result stops being a pricing question and becomes an arbitration.

None of this means outcome-based pricing is worse than usage-based pricing. It means outcome-based pricing trades one kind of unpredictability for another, and enterprises evaluating it need to understand which variables they're actually exposed to before they sign. A model that bills for tokens consumed is volume-independent of success rate but exposed to inefficiency. A model that bills for resolutions is aligned with value but exposed to your own success. Pick the exposure you can actually forecast, not the one that sounds fairest in a sales deck.

The pricing chaos is colliding with a finance function that's already skeptical. Forrester's 2026 predictions, released by Chief Research Officer Sharyn Leaver, forecast that enterprises will defer roughly 25% of planned AI spending into 2027, driven partly by a hard data point: fewer than one in three decision-makers can currently connect AI investment to their organization's financial performance. Leaver's framing is blunt: "the AI hype period ends as the pressure to deliver real, measurable results from secure AI initiatives intensifies." CFOs are responding by taking direct control of AI investment approval and gating it on ROI evidence, not vendor roadmaps.

Unpredictable, shifting pricing models make that evidence exponentially harder to produce. A finance team can build an ROI case around a known per-seat cost and a projected productivity gain. It's a much harder exercise to build that same case when the cost side of the equation is a moving target that depends on adoption rate, action volume, or resolution counts the vendor might redefine next quarter, as Salesforce did twice in eighteen months. This is precisely the gap we mapped in "From Efficiency Theater to P&L Impact" (Feb 26): without a disciplined FinOps practice built specifically for agent economics, cost tracking and value attribution stay disconnected, and every pricing model war happening at the vendor level turns into budget chaos at the buyer level. The enterprises deferring spend aren't necessarily rejecting agentic AI. Many are waiting for pricing to become legible enough to build a defensible business case around.

There's a deeper shift happening underneath the vendor pricing wars that's easy to miss if you're only watching your existing SaaS contracts. As we explored in "The Agent Economy" (Apr 2), MCP and A2A protocols are starting to let agents transact directly with other agents and services, not just with the software your employees log into. That changes the pricing problem from "how do we charge for a human who occasionally uses this tool" to "how do we charge for thousands of machine-initiated calls happening at a speed and volume no per-seat model was ever built to track."

This is where build-versus-buy-versus-subscribe decisions get a lot more complicated. A subscribed agent service priced per action might be cheaper than a full seat license today, but if your agents scale their call volume ten-fold over the next year as orchestration matures, that per-action model could become the more expensive path. Enterprises evaluating agent marketplace participation, buying capability from a specialized agent vendor instead of building it internally, need to model pricing at the volume their orchestration layer will actually generate at scale, not at today's pilot-project usage. Vendors pricing for a machine-to-machine economy are still mostly guessing too, which is exactly why the smartest enterprise buyers are negotiating flexibility into contracts now rather than locking in a metering model built for a usage pattern that's about to change.

Given that pricing models are still being invented in public, the practical move isn't to wait for the market to settle, it's to negotiate protection against the models still changing underneath you. That starts with usage caps and collars: a ceiling on monthly spend regardless of how a metering unit gets redefined, paired with a floor that guarantees predictable minimum value. It continues with explicit, contractually defined units, exactly what counts as an action, a conversation, or a resolution, in writing, with a change-control clause requiring notice before a vendor redefines the metric your invoice depends on. It also means insisting on usage data transparency: raw consumption data delivered to you in a form you can audit independently, not just a vendor-generated invoice you have to take on faith.

The enterprises in the best position to negotiate any of this are the ones that already redesigned their workflows around agents rather than bolting agents onto legacy processes, the distinction we made in "The Automation Trap" (Feb 12). An organization that knows precisely which tasks its agents handle, at what volume, and with what success rate can model usage-based or outcome-based pricing with real confidence and negotiate from a position of data, not guesswork. An organization still running agents as an add-on to unchanged human workflows is negotiating blind, and blind negotiation with a vendor that's still inventing its own pricing model is a losing position no matter which side of the table you're on.

The pricing model war isn't a temporary disruption on the way to a new stable equilibrium. It's the software industry relearning, in public and at cost to early adopters, how to measure value when the buyer of record and the user of record are no longer the same entity. Salesforce burned through three pricing models in eighteen months and still isn't finished. Intercom's outcome-based model solved one problem and created another. Gartner's $234 billion figure and Forrester's 25% deferral prediction describe the same standoff from two different sides: vendors racing to reprice around agents, and CFOs refusing to sign off until that repricing produces numbers they can actually forecast.

Enterprises that treat this as a wait-and-see problem will keep absorbing whatever pricing model their vendors ship next, the way Agentforce customers absorbed three of them without much say in the matter. Enterprises that treat it as a negotiation and governance problem, building the usage visibility, workflow redesign, and contract discipline to price agentic AI on their own terms, will be the ones setting the terms of the next contract instead of reacting to it. The question in the title isn't rhetorical. Somebody pays for autonomy. The only open question left is whether it's priced on your terms or a vendor's.

Getting pricing leverage on agentic AI starts with knowing exactly what your agents do, how often, and what that's actually worth to your organization. The Complete Agentic AI Readiness Assessment includes frameworks for modeling agent usage at scale, evaluating vendor pricing structures against your own workflow data, and building the FinOps discipline that turns a pricing negotiation into an evidence-based conversation instead of a guess. Get your copy on Amazon or learn more at yourdigitalworkforce.com. For organizations facing unpredictable agent costs or upcoming vendor renewals, our AI Blueprint consulting helps model usage at scale, design contract terms that protect against pricing model shifts, and build the cost governance practices that keep autonomy affordable as it scales.