Seats Were Always A Proxy

When Snowflake leaned into consumption pricing, the market hammered it. Every legacy SaaS metric deteriorated during the transition, and the thesis was declared broken. It turned out to be one of the most insightful pricing decisions in modern enterprise software. Data consumption was never going to decline. Consumption pricing captured every increment of it. Analysts came around eventually, after acknowledging the demand impetus behind the concept that Bob Muglia (Snowflake CEO, 2014-2019) identified and Frank Slootman (Snowflake CEO, 2019-2024) defended. The deeper lesson should have stuck. It hasn’t.  

The same play is running across multiple incumbents now. The market has learned to punish the transition. It still hasn’t learned what’s on the other side.  

The Wrong Unit  

Per-seat pricing was always a proxy for human productivity, not for the work itself. The seat stood in for a worker, and software was priced as a slice of what that worker produced. AI breaks the arrangement. Sometimes the agent does the work directly, with no worker attached. Other times it multiplies what a single worker can do. Either way, the link between seats and output is severed. When an agent processes 50,000 tickets a week, asking ”how many seats?” is the wrong question. Workload is the unit of value. Consumption is the only pricing geometry that captures it.  

The bigger reframe is what AI gets priced against. Enterprise software lived inside the IT budget— a marginal slice of OpEx — because productivity tools belonged there. AI doesn’t. A product that displaces a million dollars of labor at a hundred thousand in cost is being benchmarked against the wrong line. The right line is labor: upward of 60% of OpEx in most service businesses. That’s the prize.  

Any vendor still on seat math is capping itself at headcount when its value scales with workload. The seat ceiling is employee count. The consumption ceiling is business activity. Different by an order of magnitude.  

Rails Are Guardrails  

Here’s where I think the market is missing the harder point. The consumption upside doesn’t go to whoever runs the model. It goes to whoever owns the rails the model runs on.  

By rails, I mean audit trails, identity and permissions, data lineage, compliance scaffolding, integration with systems of record, vendor accountability when something breaks. None of it is glamorous. It is also the entire difference between an AI that drafts a credit memo and an agent authorized to approve the loan.  

No serious enterprise turns AI loose unchecked. The regulatory, reputational, and operational cost of getting it wrong is too high, and no CFO will let an unproven vendor skip the structural integrity work that incumbents have spent ten to twenty years building. The integrity layer is the deployment gate, and the platforms that built it charge for what runs over it. The bear narrative — that AI agents will cut out horizontal SaaS — ignores this entirely. AI doesn’t replace SaaS. It runs through it.  

That’s a compounding moat: capital, time, and continuous refinement, indefinitely. An AI-native vendor rebuilding from scratch is chasing incumbents who are layering AI onto rails they already own. The math rarely closes.  

The Asymmetry  

The market has partially learned from Snowflake. There’s now a playbook for punishing consumption transitions: mark down the multiple, warn about variable revenue, flag the forecasting risk, wait for the rebuild. That playbook is broadly priced in. The trough is understood.  

What isn’t priced in is what comes after — and the mispricing operates on two dimensions at once.  

The first is revenue math. The average investor still applies seat-era metrics — ARR growth, ratable revenue, Rule of 40 — to businesses whose unit economics have changed. Workload-shaped revenue compounds differently. 

The customer’s expansion isn’t capped at hiring; it’s capped at business activity. Net revenue retention isn’t bounded by seat upsell; it’s bounded by workload. Different growth function. Consensus is using the wrong one.  

The second is multiple structure. A rails-owner is two businesses in one. The rails themselves are a utility — stable, mission-critical, recession-resistant. The AI workload running on top is variable and grows with customer activity. Each piece deserves a different valuation. The market applies one blended multiple. Only rails-owners have this structure, which means only they get mispriced twice over.  

Not a Basket Trade  

Plenty of SaaS companies will announce AI strategies, gesture at new pricing models, and fail to attach rails-level defensibility. They end up as commodity compute providers. The customer pays for compute, but value capture stays with whoever owns the enterprise relationship.  

Selectivity is the trade. The filters I’m watching: defensible rails, real commitment to off-seat pricing, workload-aligned product design, and a customer base whose own business is growing. Miss any of those and the labor cost the AI absorbs never reaches the vendor — it goes to whoever is positioned upstream.  

The cyclical risk is real. Consumption moves with customer activity, and the rails floor moderates that without eliminating it. Anyone betting the math is recession-proof will be surprised in the next downturn. The bet is that structural expansion overwhelms cyclical compression. A bet, not a certainty — but the window where consensus still applies the seat-era toolkit is exactly the asymmetry worth taking.