Unit economics
Why AI features can damage SaaS gross margins
Traditional SaaS gross margin is high because serving one more user costs almost nothing. AI changes that: every interaction has a marginal cost in tokens, retrieval and compute. If an AI feature is bundled into an existing flat-rate plan and used heavily by your largest accounts, cost scales with usage while revenue does not, and margin erodes fastest on your best customers. The fix is to know cost per action before launch and to align pricing, limits and architecture with it.
The three ways margin leaks
- Unbounded usage inside a flat plan, where power users consume many times the average.
- Unbounded context, where every call ships far more retrieved text than the answer needs.
- Silent retries and agent loops, where one user outcome quietly costs several model calls.
What to measure per feature
| Metric | Why it matters |
|---|---|
| Cost per successful action | The only number that maps to price |
| Actions per account per month | Reveals the heavy-tail accounts |
| Cost of the 95th percentile account | Margin is set by the tail, not the average |
| Retry and failure rate | Failed work still bills |
Pricing and packaging responses
- Put AI in a higher tier or an add-on rather than the base plan.
- Include a generous but finite allowance, with transparent overage.
- Meter by outcome the customer values, not by tokens they do not understand.
- Reserve the most expensive model tier for the actions that demonstrably need it.
Architecture responses
Route simple work to cheaper models, cache retrieval aggressively, keep context tight, and set hard budgets per tenant. These are engineering decisions with a direct line to gross margin, which is why they belong in the design phase rather than a later cost-cutting exercise.
How Datics helps
We model inference economics alongside the product design, so the AI capability you launch is priced and architected to hold its margin at scale.
Related answers
Cost
How much does production AI cost in SaaS?
The real cost of running AI in a SaaS product: inference, retrieval, storage, egress and engineering. Indicative bands and the variables that move them.
Read the answerCost
How to reduce AI infrastructure cost in production
Concrete levers for cutting production AI cost in a live SaaS product: context discipline, model routing, caching, placement and GPU utilization.
Read the answerYour customers already expect AI from your product.
Let us make it real without rebuilding what already works.