How to price an AI product: VC consensus on seats versus usage versus outcomes

Use a base subscription for access, with a usage meter wherever inference cost moves. Outcome language belongs in the business case. I would rarely lead with pure outcome pricing at seed because it hands your buyer a measurement fight before the product has earned budget trust inside procurement.

Why the seat price cracked

The seat model came from human software, where a buyer could map value to named employees and forecast the invoice before legal touched the order form. AI breaks that map. A claims copilot may review files overnight, while the adjuster opens the product only to approve the final draft.

The useful VC consensus I trust has shifted because the boardroom sees AI as labor capacity, yet the procurement desk still buys contracts it can explain to finance. That tension matters. If your first segment is vague, pricing becomes a costume, and Enterprise or SMB first? Picking your first segment belongs upstream of any packaging debate.

Where does usage belong?

Usage belongs wherever your cost curve moves faster than your customer's willingness to pay, which is why raw AI margins can turn ugly inside a popular workflow. Keep it visible. OpenAI pricing and Anthropic pricing publish model costs in token units, while your buyer usually thinks in documents processed or calls reviewed.

Do not make tokens the headline unless the buyer already buys compute, because raw meters push your customer into auditing prompts instead of judging business value. The buyer's noun works. A recruiting product can price on screened resumes, while a support product can price on resolved conversations once the system has clean event logs.

Why outcome pricing is harder than it sounds

Outcome pricing attracts founders because the buyer's spreadsheet already contains the prize, like fewer manual reviews or more recovered invoices. Stay sober. The moment you charge on the outcome, sales turns into a debate about the baseline, who caused the lift, seasonal noise, data quality problems, plus every exception hiding in the customer's process.

Pure success fees work only when the buyer already trusts the event source and both sides accept the same ledger. That is rare early. Outcome language still belongs in the proposal, especially when PLG or sales-led? The right GTM motion follows proof has forced the team to define which proof a buyer can observe before expansion.

The package I would actually sell

Most seed AI SaaS products should sell a paid base that maps to access, security review, admin control, and the workflow owner with budget authority inside the account. Add a meter after that. The base gives procurement a familiar subscription, while the meter protects your gross margin when users push more files, calls, tickets, or records through the model.

Stripe documents usage-based billing as metered billing, which fits AI cost exposure better than pretending every account behaves the same. The warning is simple. A meter without a value story feels like a toll booth, so the proposal should name the business process before it names the unit.

The clean seed package is usually base access with included usage, then overage or tier movement when consumption crosses a threshold the buyer understands. Keep the invoice boring. If you need a broader seed-stage pricing frame, How to price your B2B product at seed stage covers the non-AI parts that still decide whether finance says yes.

What should a founder do about it?

Start with the budget owner and the event the buyer already measures, because pricing built around an invisible dashboard metric dies in legal review during negotiation. Write the noun first. A people budget can carry seats or role bundles. Workload swings need usage guardrails, and a completed-event budget can carry outcome pricing after attribution is accepted.

Then model the actual workflow before naming the price page, because prompt volume, retrieval calls, human review time, and support load decide whether the deal funds itself. Use real logs. If those logs are thin, How to define your first ICP, uncomfortably narrow, then iterate is more useful than another investor memo about monetization.

Should you price per seat, by usage, or around an outcome? The answer is practical. Human-led adoption points to a base subscription, variable cost or workload points to usage, and measured events the buyer already trusts can carry outcome pricing. Anything else turns pricing into theater, and theater wastes the scarce sales cycles a seed company needs for learning.