Your AI feature is a cost line, not a feature flag

A classic SaaS feature costs almost nothing once it ships. An AI feature costs money every time someone uses it. That one difference changes how you should think about margin.

The invoice is a lagging indicator

Most founders learn what their AI feature costs when the provider's invoice arrives — weeks after the prompt change, model upgrade or viral week that caused it. By then the margin hit is history.

What changes margin in an AI product

• A longer system prompt is paid on every request.
• A bigger model can multiply cost per request several times over.
• Power users on flat plans can cost far more than they pay.
• Agent loops and retries turn one user action into many API calls.

Measure it like revenue

You already look at revenue daily. Do the same with AI cost — by model and project, next to revenue — and read the result as gross margin. When margin slips, ask what changed: revenue mix, a provider, or a model.

A simple routine

1. Track AI cost per $1 of revenue weekly.
2. Set a margin floor alert a few points below today's margin.
3. Before shipping a prompt or model change, estimate its cost with the LLM cost per user calculator.
4. After shipping, check the next day's margin.