The CFO's Dilemma: The Proxy That Ran Out of Runway
Charles Goodhart was a British economist who, in 1975, articulated something anyone who has worked inside a large organization already knew: when a measure becomes a target, it ceases to be a good measure. It’s been confirmed so consistently across economics, medicine, education, and public policy that it functions less like a theory and more like gravity.
The seat was always a proxy. It was never actually measuring value.
That’s the underlying challenge worth understanding about the SaaSpocalypse, which has wiped out over two trillion dollars in software market capitalization since February 2026. Most of the mainstream coverage still frames this as AI threatening to replace human workers, which makes per-seat pricing economically absurd. That is partially a factor in this chaos. But the deeper story is that the seat was never measuring what anyone thought it was measuring, and AI has simply made the gap between the proxy and reality too wide to ignore.
Think about what a seat license actually tracked. Not work completed. Not value delivered. Not outcomes achieved. A login. An interface rendered on a screen for a human who may or may not have been productive, engaged, or even actively using the thing. The seat was a rough stand-in for “a person doing something useful with this software,” close enough to value, for long enough, that the entire financial architecture of SaaS got built on top of it. ARR, NRR, the Rule of 40, two decades of premium revenue multiples. All of it resting on the assumption that counting logins was a reliable signal for value delivery.
It worked, until it met something that delivers value without logging in, in the traditional sense, at all.
When an AI agent resolves a support ticket, qualifies a sales lead, or drafts a procurement contract, it isn’t via a rendered GUI. It doesn’t generate a seat usage metric. It just does the work. The per-seat model registers that output as, at best, one license for the agent itself, regardless of how much human-equivalent work it produced. When Monday.com replaced 100 sales reps with AI agents, those weren’t 100 seats quietly consolidating. They were 100 seats gone, permanently, with no replacement count coming. The measure and the thing it was supposed to measure had finally, visibly, separated.
The vendor response has been to reach for consumption-based pricing: charge per token, per API call, per action taken. Which is logical, except it carries its own proxy problem. Tokens consumed and API calls fired are no more connected to business value than logins were. An AI agent that fails and retries three times burns more tokens than one that succeeds cleanly, generating more revenue for the vendor.
You’ve built a model that rewards your product for not working.
The anxiety this produces is measurable. Zylo’s 2026 SaaS Management Index, covering $75 billion in enterprise software spend, found that 78% of IT leaders experienced unexpected charges from consumption-based AI pricing in the past year, and 61% cut projects because of it. The production gap is even starker: 67% of companies report gains from AI pilots, but only 10% ever scale to production. The pricing model is a significant reason why. When your meter runs whether the agent succeeds or fails, you stop running the agent. The industry traded one bad proxy for another and is now living in the consequences.
What’s beginning to work, messily and without a settled playbook, is outcome-based pricing. It measures what the seat and the token never could: the thing that actually happened. A support ticket resolved. A lead that qualified. A workflow that completed. Intercom charges $0.99 per resolved ticket for its Fin AI agent; zero for failures. Sierra, Bret Taylor’s startup, hit $100 million ARR in under two years running exactly this model. Others are using this emerging standard as the default for their AI agents. Hybrid structures, a predictable subscription floor with outcome meters layered on top, are now how 41% of vendors are structuring deals, up from 27% a year ago.
None of this is clean yet. FASB issued new accounting guidance in September 2025 specifically because AI development doesn’t fit the old linear model that’s governed software cost capitalization since 1998. Valuation frameworks are openly in flux: traditional SaaS trades at 6x revenue post-crash, AI-native companies at 25-30x, and analysts are actively debating which metrics should anchor those multiples when ARR tied to human seats no longer tells the story. New proxies are being tested, things like Credit Consumption Velocity, Agentic Work Units, and Cost per Resolved Request, and it’s not yet clear which of them will avoid Goodhart’s fate.
That last part matters. The history here is not reassuring. Every time the industry finds a metric that approximates value well enough, it builds a financial architecture on it, and then the metric drifts until something forces the reckoning. The seat drifted for twenty years before AI made the gap undeniable. There’s no guarantee the replacement proxy won’t do the same thing.
The SaaSpocalypse is proxy collapse at scale. Two trillion dollars repricing the distance between a measure and the thing it stopped measuring.
The seat was a genuinely useful solution to a hard problem: how do you price something as diffuse as the value of software to an organization? Counting the humans who used it was close enough, for a long time. AI didn’t break that by being disruptive. It broke it by being accurate, by doing real work without generating the signal the model was trained to count.
What software vendors are actually selling now is outcomes. The ones who figure out how to price that cleanly will be fine. The ones still defending the seat are defending a measure that already lost contact with the thing it was measuring.