SaaS pricing strategy 2026

Per-seat is dying for AI products. This is not a clickbait take. It is our observation across four quarters of pricing experiments. Here is what we believe, what we tried, what failed, and what stuck.

Why per-seat dies on AI

Classic SaaS per-seat assumes that a user's monthly cost is roughly constant. A logistics admin working 8 hours daily in the panel and an accountant logging in twice a month cost us about the same. It averaged out in cash flow.

With AI it does not. A user sending 50 chat messages a day and editing orders through tool calls can cost us 30x more in model tokens than the twice-a-month one. Per-seat pricing is the same. You feel it on the entire margin.

Concretely: in Q1 2026, the finance app's top-10 users generated 41% of the entire engine bill. Everyone was paying the same monthly fee.

The three pricing experiments that failed

Per-message

Every chat message has a price. Logical, but users immediately started writing long messages ("five things at once") which made the chat experience worse. A few stopped using it entirely because they got anxious about every button click. Reverted after two weeks.

Per-conversation

Fixed amount per conversation. Maybe the worst version. Users would start one 4-hour conversation a week and cram every question into it. The context bloated, model cost grew, and we ended up giving the service to the customer for less than it cost us. Reverted after a month.

Per-tenant flat

"Pay X a month, do anything." We lasted eight weeks. In three of those weeks two companies blew through our monthly model bill. One of them was straight-up loss-making for us. Reverted.

What works: core seat + token bundle + overage

What we run now across three product lines (admin, finance, logistics):

1. Core seat fee — 19 EUR per user per month. Covers hosting, infra, classic CRUD operations, normal monitoring, first-line support.
2. Token bundle — every tenant gets 1.5M Engine tokens per month included. Roughly 7,500 chat messages, 500 medium generations, 50 longer workflow runs.
3. Overage — past the bundle, 0.0024 EUR per 1k tokens. Soft warning at 80%, hard email at 100%. Above 120%, the tenant has to confirm they do not want tool calls blocked.

The overage is what protects the margin. Heavy users pay, light users do not overpay.

The renewal data

Q1 2026: 14 design partners + 9 paying customers. 20 of 21 renewed. The single churn: internal restructuring at their end, not product dissatisfaction. From the heavy-user segment (top 10% token use) we had zero churn. From the light-user segment (bottom 30%) we had one churn — 19 EUR core seat was too much for them given their actual usage. We are now working on a "viewer seat" tier (5 EUR, read-only, no chat agent) for that segment.

What we are not saying

This is not the only model. Some markets (e.g. tooling with very predictable monthly load) still work with flat. Some (e.g. heavy seasonality, small transactions) work with per-message. It is just that for our B2B AI tooling, the hybrid above is what landed.

The decisive factor is not pricing cleverness. It is transparent communication. Every tenant sees real-time on the dashboard how much they have used and what their expected monthly bill is. Nobody gets surprised by an invoice. That matters more than any single pricing twist.

What customers ask for, and what we say no to

Six of the 21 customers have asked for an "unlimited tokens" plan. It simply does not exist. We know per model exactly what a heavy enterprise's 500k-tokens-a-day usage costs us: at least 12-15 EUR per day, or about 450 EUR per month. Offering that inside a flat fee either eats margin or pushes the entire product upmarket. Our answer to those customers: you can prepay bigger token bundles at a discount (10M tokens for 18 EUR per month, 50M tokens for 75 EUR per month), and overage inside a prepaid bundle is 20% cheaper. That is honest, and the heavy user gets the predictable monthly cost they actually wanted.

Another frequent request: "bill us only on conversion events" — for example, when an AI chat conversation ends in an order. Attractive. But the preparation work the chat does (10 messages, 5 tool calls, document lookup) is a real cost regardless of whether the customer buys at the end. If we tie pricing to success, the cost of the unsuccessful conversations gets paid by somebody else. Over time, that eats the margin again.

The pricing model evolves

In Q3 2026 we plan to introduce the viewer seat (5 EUR) tier and a team-bundle offer (10+ users, 15% discount on the core seat). We are also refining the overage discount bands: based on top-5% user historical data, token purchases above 10M will be 5% cheaper. Neither change is dramatic alone, but the impact on total ARR is significant.

The deeper lesson: in the AI era, pricing is not a static document. It is a living system you measure monthly and reform quarterly.