AI SaaS pricing
AI SaaS pricing: a ladder that follows the value.
Three rungs — assistive, workflow, outcome — matched to how much of the job the AI actually completes, with the margin math that keeps each one honest.
What is AI SaaS pricing?
AI SaaS pricing is the practice of charging for software whose marginal cost is no longer near zero. A seat price assumes cost scales with people; an AI feature's cost scales with the work it completes, so the pricing mechanic has to match the amount of the job the AI does. The pattern that holds up is a three-rung ladder: assistive features included in seats, workflow features on a platform fee plus metered bands tied to a unit the customer already counts, and outcome-priced agents charged against the result they deliver.
The long-form argument, including the three ways teams get this wrong, is in Pricing AI Features in a Seat-Based World.
Three rungs, three mechanics.
Each rung matches a level of autonomy. Most products should be on more than one at the same time.
| Rung | What the AI does | Pricing mechanic | Buyer’s comparison price |
|---|---|---|---|
| Assistive | Makes a human faster inside an existing workflow — drafting, summarizing, suggesting. | Included in the seat price, with generous but real fair-use ceilings. | Another vendor’s seat. |
| Workflow | Completes a defined unit of work with light supervision — a document, a ticket, a report. | Platform fee plus metered bands, priced on a unit the customer already counts. | The hours that unit used to take. |
| Outcome | Owns a job end to end. This is digital labor, not a feature. | Per resolved case, per qualified lead, per closed book month — with quality-linked terms. | A salary or a BPO contract. |
Never meter on tokens or credits the customer has to learn. If the unit of work is still unstable, use credits to discover it — then retire them into the ladder.
A worked example on the workflow rung.
Illustrative numbers, chosen to show the method rather than to describe any customer. Substitute your own and the shape of the answer holds.
- Pick the unit the customer counts. Say a support ticket triaged and routed. Not tokens, not “AI actions” — the thing already on their dashboard.
- Cost the unit fully loaded. $0.30 model and retrieval on the happy path, plus retries and failures. Assume 20% of tickets take a second pass: $0.36 per completed ticket. Demo-path cost would have said $0.30 and been wrong by a fifth.
- Price the band, not the unit. A $2,000/month platform fee including 5,000 tickets, then $0.90 per ticket beyond it. The customer can predict the bill; you can predict the load.
- Check the margin at the band, not blended. 5,000 tickets cost $1,800 of the $2,000 base — a 10% gross margin at the floor, which is a trap unless the overage does the work. At 12,000 tickets: revenue $2,000 + (7,000 × $0.90) = $8,300 against $4,320 of cost, a 48% gross margin. The floor is a land price; the band is the business.
- Model the trajectory both ways. Model prices fall, but usage mix shifts toward heavier tickets as trust grows. If the average ticket drifts from one pass to 1.5, unit cost rises to $0.45 and that 48% becomes 35%. Re-price on the cost curve you can see, not the one you hope for.
Three numbers, reviewed monthly, per AI feature: fully loaded cost per completed task, gross margin per rung and segment, and the cost trajectory. Blended margin is how the heaviest decile of users eats the P&L unnoticed.
Pricing is usually the second problem.
If the feature has not shipped to a real workflow yet, readiness comes first: run the free scorer in the Lab or read how the AI readiness assessment works. When pricing is the live question, advisory runs through ProductExec.
Questions this raises
How should AI features be priced in a seat-based SaaS product?
Match the pricing mechanic to how much of the job the AI completes. Assistive features that make a human faster belong in the seat price. Features that complete a defined unit of work belong on a platform fee plus metered bands. Agents that own a job end to end should be priced against the outcome. One product can carry all three rungs at once.
Should AI features be charged as an add-on or included in the base tier?
Include assistive features: their cost per seat is bounded because a human is the throttle, and withholding them invites a competitor to make them table stakes. Charge separately once the AI completes work without a human in the loop — that is the point where cost scales with usage rather than headcount, and where the buyer's comparison price stops being software.
Are AI credits a good pricing model?
Credits are a transitional currency. They are useful while your unit of work is still unstable, because they buy flexibility — but they cost comprehension, and customers who cannot predict their bill discount the product to compensate. Use credits to learn the real unit of work, then retire them into the ladder. A credit system still running in year three is a pricing decision being avoided.
What margin should an AI SaaS feature hold?
Track three numbers per feature, monthly: fully loaded cost per completed task (inference, retrieval, retries and failure paths — demo-path cost is fiction), gross margin per rung and per segment rather than blended, and the cost trajectory as usage shifts toward heavier tasks. Blended margin is how the top decile of users quietly eats the P&L.
Repricing AI into a seat-based product?
Bring the feature, the unit of work, and your cost per task. An hour is usually enough to know which rung it belongs on.