Per-seat SaaS math is broken: when pricing stops scaling

Per-seat pricing was an honest model in 2010. The cost-to-serve was roughly proportional to active users, and the value delivered was roughly proportional to seats provisioned. Those two assumptions held, and per-seat was the cleanest way to price.

Neither assumption holds in 2026. The cost-to-serve has decoupled from active users (LLM calls, storage, and compute scale on usage, not seats). And the value delivered is now wildly non-proportional to seats: in most enterprise SaaS deployments, 15–25% of seats produce 80% of the value. The remaining 75–85% of seats are paying full price for software they touch once a quarter.

The math has stopped working, and it shows.

How we know it’s broken

Three clean tells, observed across the 30+ Upstream engagements we’ve shipped:

1. The seat-to-active-user ratio. When more than 60% of provisioned seats have zero meaningful usage in the last 30 days, the per-seat licence is no longer a usage charge — it’s a tax on the procurement decision.
2. The CFO conversation. When the renewal price is rising faster than headcount, finance starts asking why. The vendor’s answer is always “you’re getting more value.” The CFO’s gut says otherwise.
3. The category drift. When a single SaaS tool ($1.2K/seat/mo) costs more than the team that uses it ($800K/yr salary load), the unit economics have inverted. The tool was originally priced as overhead on the team; it has become bigger than the team.

Once you see two of these three, the rebuild conversation has started — even if the contract isn’t up for renewal yet.

The Harvey AI case

We rebuilt the Harvey AI workflow for an AmLaw 200 firm in 11 weeks. Same workflow — document review, drafting assistance, citation checking — at <5% of the cost. Total Y1 reclaim: $686K, on a 50-attorney deployment that had been costing $720K/yr at $1.2K/seat/mo.

The architecture was direct: Anthropic API on the firm’s own AWS account, a thin orchestration layer, a UI calibrated to the firm’s specific document templates. No magic. The cost differential is mostly the per-seat margin Harvey was charging on top of underlying API costs.

The IT team’s first question was “what’s the catch?” There wasn’t one. The catch was that the firm had been paying enterprise per-seat pricing for what is, architecturally, a thin wrapper around the same APIs the rebuild uses directly.

When per-seat still makes sense

Per-seat is still the right model in three situations:

1. The vendor’s marginal cost is genuinely seat-driven. Mostly: support-heavy products with human-in-the-loop service.
2. The value is genuinely seat-proportional. Communication tools (Slack, Zoom, email) where every seat is an active node in the network.
3. The total spend is small enough that the rebuild ROI doesn’t pencil. Below $50K/yr per tool, rebuilding usually doesn’t make sense. The wrapper margin is real but not worth the engineering cost to displace.

Outside these cases — and most enterprise AI tools, observability platforms, marketing automation suites, and CRM add-ons fall outside — per-seat is becoming a transition pricing model on its way to obsolescence.

What replaces it

Three patterns we see emerging:

1. Direct API + thin wrapper. Skip the SaaS layer; pay the underlying provider (Anthropic, OpenAI, Stripe, Twilio) on usage, with a small custom UI. This is what the Harvey rebuild looks like.
2. Usage-based SaaS. Vendors who price on actual API calls, not seats. Modern observability tools (OpenObserve, Logfire) are starting to price this way.
3. Open-source self-hosted. For mature categories (analytics, BI, project management), open-source self-hosted is operationally cheaper than enterprise SaaS at scale.

The right pattern depends on the category and your operational appetite. The wrong pattern is to keep paying per-seat on tools where the per-seat model has already stopped making economic sense.

How to start

If you’re paying $200K+/yr for any single SaaS tool, the rebuild conversation should be on the table. The math we use:

Annualised SaaS cost − (rebuild cost + ongoing infra + ongoing maintenance) = annual reclaim

For most enterprise SaaS in the AI / observability / CRM / marketing categories in 2026, the reclaim is 70–95% of the original spend. Payback lands in 6–12 months. After that, the reclaim compounds annually.

We do this analysis as a free 30-minute stack review — no sales pitch, just the math on your specific stack.

Read more: /upstream/ · /upstream/free-stack-review · /upstream/harvey-alternative