What AI Really Costs a Small Business: Tokens, Seats and ROI

A practical framework for budgeting AI tools and APIs, estimating token costs and proving return on investment.

AI spending falls into two buckets: seats (monthly subscriptions per user) and usage (API calls billed per token). Each has a different cost curve, and each needs a different ROI test.

Seats: predictable, but watch adoption

A per-user subscription is easy to budget. The risk is paying for licences nobody uses. ROI is driven by three numbers: hours saved per person, the share of people who actually use the tool, and the loaded hourly cost of those people.

Example: 4 hours a week × 5 people × 70% adoption × 45 per hour ≈ 2,730 of monthly value against a 300 tool bill.

Usage: scalable, but variable

APIs bill per million tokens, usually with output priced several times higher than input. A token is roughly three-quarters of an English word.

At 800 input and 400 output tokens with prices of 3 and 15 per million, each request costs 0.0084 — about 252 a month at 1,000 requests a day.

Five ways to cut AI costs

1. Trim prompts. Long system prompts are paid for on every call.
2. Cap output length. Output tokens are the expensive ones.
3. Cache repeated context where your provider supports it.
4. Route by difficulty. Send simple classification to a smaller, cheaper model.
5. Batch non-urgent jobs if your provider discounts batch processing.

Proving ROI

Measure before and after: tickets closed per hour, articles drafted per week, time to first response. Saved time only becomes money when it is redeployed to revenue work or avoids a hire — otherwise report it as capacity.

Hidden costs

Include human review time, integration work, data-handling compliance and the cost of errors. A realistic model beats an optimistic one when you present the budget.

Run your numbers in the AI token cost and AI ROI calculators.