Seat pricing was built for human attention. One login, one inbox, one predictable cost of goods. AI products violate that assumption the moment a single power user — or their agent farm — burns more inference than a whole team used to cost you in servers. Founders feel this first as a finance surprise: revenue looks healthy on seats while gross margin quietly collapses on a handful of tenants. The answer is not a clever blog post about “value-based pricing.” It is packaging that admits AI is a variable cost — without turning every invoice into a fight with procurement.
Why seats break under AI consumption
Classic SaaS seats amortize fixed software cost across humans. Your COGS scaled roughly with support and storage. With LLMs, COGS scale with tokens, tool loops, retrieval, and retries. Two customers on the same plan can produce wildly different bills because their workflows differ — not because one “uses more features” on the settings page. That mismatch creates three failure modes. First, you underprice power users and subsidize them with quiet accounts. Second, you overprice cautious buyers and lose the deal to a usage-native competitor. Third, sales invents one-off contracts that engineering cannot meter, so finance cannot forecast and customer success cannot explain the bill. None of this requires citing invented industry percentages. If you can open a cost dashboard and see tenant-level variance that seats ignore, you already have the evidence.
Hybrid seat + usage: the practical middle
Most B2B buyers still want a predictable base. Keep a seat or workspace fee for access, collaboration, and support — then meter the expensive AI surface: generations, agent runs, indexed documents, or evaluated outcomes. The seat covers the product shell; usage covers the variable brain. Design the meter to match the customer’s mental model. “Tokens” confuse buyers. “Agent tasks completed,” “documents processed,” or “approved workflows” map better to budgets. Internally you still track tokens; externally you sell a unit a CFO can defend in a quarterly review. Include soft and hard caps per tenant. Soft caps warn. Hard caps pause AI features with a clear upgrade path. Cap design is a product decision: silent overages destroy trust; surprise shutdowns destroy workflows. Prefer graceful degradation — smaller models, queued jobs, or human-only modes — before an abrupt stop.
Outcome pricing: attractive and dangerous
Outcome pricing — pay for closed tickets, qualified leads, resolved claims — aligns incentives when you can measure the outcome cleanly and attribute it fairly. It fails when attribution is fuzzy, when customers game the definition, or when your model quality becomes a contractual SLA you cannot fully control. Use outcomes for a narrow, auditable slice of the product — not the whole platform. Keep a base fee so you are not financing R&D on hope. Write definitions of success into the contract with dispute paths. If you cannot instrument the outcome with the same rigor as billing, you are not ready for outcome packaging.
Margin traps founders miss
Retries and tool loops hide inside “one generation.” A chatty agent can multiply cost without multiplying perceived value. Caching, prompt compression, and routing cheaper models for mechanical steps are product work, not infra trivia left to a lone platform engineer. Also watch free seats that still call paid models, sandbox tenants left on production keys, and support staff reproducing customer issues on your dime. Meter internal usage too. If engineering cannot see per-tenant cost next to revenue, pricing debates stay theatrical and discounts stay unconstrained.
Founder checklist
Before you rewrite the price page: measure p50 and p95 AI COGS per tenant for a meaningful window; pick one customer-facing meter tied to a real workflow; keep a seat or workspace base for predictability; ship soft and hard caps with an upgrade UX; decide what is included versus billable inside agent loops; align sales compensation so discounts cannot erase the meter; put finance on the same dashboard as product. You do not need invented benchmarks to know the direction of travel. If AI is material to your COGS, packaging that ignores consumption is a deferred crisis. Treat pricing as architecture: the meter, the caps, and the domain events that bill must be designed with the same care as the model call itself.
How to introduce the meter without killing the pipeline
Announce packaging changes as a migration, not a surprise. Grandfather existing contracts for a defined window. Give champions a usage dashboard before the first bill that includes overages. Train sales on a single narrative: the base seat buys collaboration and support; AI capacity is a consumable with transparent units. Offer annual commits with included usage bundles for buyers who hate volatility — but still meter so you see truth. Bundles without metering recreate the seat problem under a new name. The goal is shared visibility: customer and vendor both see consumption tied to value-producing actions. When a prospect demands unlimited AI on a flat fee, treat that as an enterprise risk conversation. Either price the risk explicitly with a high floor, or walk away. Unlimited AI on a flat fee is how startups fund somebody else’s agent farm.
Packaging experiments that stay ethical
Run packaging experiments on new logos first. Measure expansion revenue, churn reasons, and support tickets about billing clarity. Avoid dark patterns: hidden meters, unexplained spikes, or features that silently switch from included to billable without in-product notice. Partner with finance weekly during the first two billing cycles of a new meter. If invoices need manual fixes, the meter design is wrong. Automate from day one — including credit notes for your own incidents — so trust compounds instead of eroding.
