This is the worked example we use most in finance conversations, because every number in it can be checked with a calculator. It describes a typical customer support SaaS with an AI assistant and a flat price, before and after per-tenant governance.
The business
The company has 500 customers on a $99 monthly plan, so 500 x $99 = $49,500 MRR. The assistant runs on GPT-4o, listed at $2.50 per 1M input tokens and $10.00 per 1M output tokens. Most customers use it lightly; we will assume $12 of model cost per light customer per month.
25 customers are different. They route most of their support queue through the assistant, with long ticket histories in every prompt. Each sends about 120M input tokens and receives 10M output tokens a month: 120 x $2.50 + 10 x $10.00 = $400.
Before: one margin for everyone
Total model cost is 475 x $12 + 25 x $400 = $15,700. Against $49,500 of revenue that is a model-cost gross margin of 68.3%, which looks acceptable on an account-level dashboard.
Per tenant, it is not. Each power user pays $99 and costs $400, a margin of -$301 each. Together the 25 power users lose $7,525 a month, which the other 475 customers quietly subsidize.
After: a cap with a fallback
The team sets a per-tenant model budget of $35 a month on the $99 plan. Capped power users now cost $35 each, so the saving is 25 x ($400 - $35) = $9,125 a month. Total model cost drops to $6,575 and the model-cost gross margin rises to 86.7%.
A cap alone would mean those customers run out of assistant mid-month. Instead, the policy falls back to GPT-4o mini for routine ticket triage once a tenant passes 80% of the budget. The same 120M input and 10M output tokens cost 120 x $0.15 + 10 x $0.60 = $24.00 on GPT-4o mini, so every dollar left in the budget after the switch buys about 16.7 times as much assistant usage as it would on GPT-4o. Most heavy customers keep working for the rest of the month, and the hard stop at 100% becomes rare.
| Before | After | |
|---|---|---|
| Monthly revenue | $49,500 | $49,500 |
| Model cost, light customers | $5,700 | $5,700 |
| Model cost, power users | $10,000 | $875 |
| Model-cost gross margin | 68.3% | 86.7% |
The better option: sell the usage
Capping protects margin. Selling the usage grows revenue. Those 25 customers are telling you the assistant is worth more to them than $99. With per-tenant metering in place, the team can offer an add-on with a larger allowance, or an overage priced at cost plus margin and synced to the billing system. Both options need the same thing the cap needed: exact cost per tenant, in time to act on it.
Run it with your numbers
- Replace the plan price, customer count and model with your own.
- Pull per-tenant cost for a full month and count the tenants whose cost exceeds their price.
- Set the cap at the cost level your margin target allows, and decide which features may fall back.
- Use the runaway cost calculator to add loop incidents on top of the steady-state numbers.
What the example leaves out
The example covers steady-state usage only. It does not include runaway loops, which arrive as spikes rather than as a monthly pattern, and it does not include infrastructure, support or payment costs, which sit outside model COGS. It also assumes list prices with no prompt caching. A support assistant with a long, stable system prompt usually caches well, which lowers every number in the before column and makes the per-tenant spread even more important to measure, because cache behavior differs by tenant.
Ananya Iyer
Head of FinOps Research, PALM SOUQ
Chartered accountant and former SaaS finance lead at a Chennai-based CRM company. Maintains the LLM Price Index and the margin models behind the runaway cost calculator.