Unit Economics of AI SaaS
Last updated:
Why the old SaaS margin rule stops applying
Classic SaaS enjoys gross margins that make other industries jealous, because serving one more customer costs almost nothing. Hosting is shared, support scales slowly, and a customer who logs in ten times a day costs about the same as one who logs in once.
AI products break that. Every summary, extraction or agent step is a metered call to a model provider or a GPU you are paying for. Your most engaged customers, the ones you want most, are also your most expensive. A flat subscription quietly transfers margin from you to them.
This does not make AI SaaS a bad business. It makes cost of goods sold something you manage weekly rather than something you glance at annually.
A worked example
Take an illustrative product that reads inbound supplier invoices for mid-sized distributors. It charges a flat 400 pounds a month per customer. Each invoice costs, all in, around four pence to process once you include extraction, a validation call and occasional retries.
| Customer | Invoices per month | AI cost | Other COGS | Gross margin |
|---|---|---|---|---|
| Small distributor | 1,000 | 40 | 30 | 82% |
| Typical distributor | 3,000 | 120 | 35 | 61% |
| Large distributor | 9,000 | 360 | 45 | -1% |
Averaged across the book, the margin might look acceptable. But the large customer is loss-making, and they are exactly the account sales will be proudest of. Pricing that tracks invoices processed, with a platform fee underneath, fixes this at the root. The numbers are illustrative; the shape is extremely common.
The costs people forget to count
- Retries and failed calls, which you usually pay for
- Validation or second-pass calls that check the first answer
- Embedding generation when documents are uploaded or re-indexed
- Evaluation runs before every release, which can be substantial on a large test set
- Human review time for low-confidence outputs, if your team does it rather than the customer
- Free trials and pilots that consume real inference with no revenue
- Vector storage and search, which grows with every document retained
The human review line is the sneaky one. If your product promises reviewed results and your staff do the reviewing, you are running a service business with software margins on the slide deck and agency margins in the accounts.
Measuring cost per customer properly
You cannot manage a margin you can only see in the monthly provider invoice. Every model call needs to be tagged with tenant, feature and request, and the cost computed at the time of the call. We explain the mechanics in metering AI feature usage per tenant.
Once that data exists, three reports do most of the work: gross margin by customer, cost per unit of value (per invoice, per call reviewed, per report), and cost per feature. The third one often reveals that a small, rarely-used feature is quietly burning a large share of spend.
Levers that actually move the margin
- Route by difficulty. Send simple, high-volume tasks to smaller, cheaper models and reserve large models for the hard cases. A classifier or a confidence check decides.
- Trim the context. Better retrieval that sends five relevant paragraphs instead of an entire document cuts cost and often improves quality.
- Cache. Identical or near-identical requests, fixed system prompts and repeated reference material are all cacheable.
- Batch. Work that does not need an instant answer can often run through cheaper batch processing.
- Do deterministic work in code. Dates, totals, lookups and formatting should never cost tokens.
- Price the driver. If cost follows documents, price documents. If it follows seats, seats are fine.
In our experience the first two levers usually matter most. Switching providers for a slightly lower list price tends to matter least, and often costs more in engineering than it saves.
Pricing that protects the margin
The fix for unpredictable unit costs is almost never 'charge more'. It is aligning what customers pay with what drives your cost. Common structures include a platform fee plus usage, tiers with included volumes and overage, prepaid credits, and outcome-based pricing where you charge per result delivered.
Each has trade-offs. Usage pricing makes budgeting harder for buyers, which procurement teams dislike. Credits confuse people unless one credit maps to something they understand. Outcome pricing is attractive but requires a clear, agreed definition of the outcome. Our SaaS pricing strategy guide goes deeper on the options.
If your best customer is your least profitable one, the problem is the price list, not the customer.
What healthy looks like, honestly
There is no universal target, and anyone quoting a precise industry benchmark for AI SaaS margins is guessing. What we look for is a margin that improves as the product matures, a clear understanding of which customers and features are expensive, and pricing that does not punish engagement.
Early on, lower margins are acceptable if you are learning and costs are trending down. What is not acceptable is not knowing. When SpiderHunts builds an AI product through our SaaS development practice, cost per tenant is on the internal dashboard before the first paying customer signs.
Frequently asked questions
What gross margin should an AI SaaS product target?
Will model prices keep falling and solve the problem?
Should we absorb AI costs during free trials?
Is self-hosting an open model cheaper?
Worried your AI costs will eat the margin?
Send us your pricing and a rough picture of usage. We will model cost per customer with you and show where the biggest savings are likely to be.
Related services
What we build for problems like this one