The build fee and retainer are only part of what an AI automation costs. Every AI step in a live workflow consumes API tokens with every run, and that ongoing usage cost is often the line item both agencies and clients underestimate most — sometimes not noticing until a monthly bill arrives far higher than expected.
How Token-Based Pricing Actually Works
Most AI model providers charge separately for input tokens (what you send the model — the prompt, the data, any context) and output tokens (what the model generates back), typically priced per million tokens. Output tokens are usually priced several times higher than input tokens, which matters a lot for workflows that generate long responses.
A Simple Way to Estimate Monthly Cost
- Estimate volume: how many times will the AI step run in a typical month?
- Estimate tokens per run: a short classification task might use a few hundred tokens; a document summarization task might use several thousand.
- Apply the provider's per-token rate: multiply input tokens by the input rate and output tokens by the (usually higher) output rate.
- Multiply by volume: total monthly token cost = (cost per run) × (runs per month).
- Add a buffer: volume and prompt length both tend to grow over time as clients trust and expand automation — budget with headroom, not just the current baseline.
Budget vs Premium Model Tiers
Most providers offer a range of models at different price and capability points — a smaller, cheaper model for simple, high-volume tasks like classification, and a larger, more expensive model reserved for tasks requiring deeper reasoning, like drafting nuanced client communication. Using the expensive model for every single step in a workflow is one of the most common, avoidable sources of overspend.
Cost-Control Tactics Worth Building In
- Route by task complexity: use a cheaper, faster model for simple classification or extraction, reserving premium models for genuinely complex reasoning steps.
- Trim prompt length: unnecessary context sent with every call adds up quickly at volume; include only what the model actually needs.
- Use prompt caching where available: many providers offer reduced pricing for repeated, unchanged context across calls, which can meaningfully cut costs for workflows with a large, stable system prompt.
- Set usage alerts: configure spending alerts with the model provider so an unexpected volume spike doesn't turn into a surprise bill.
- Batch where real-time isn't required: many providers offer meaningfully discounted rates for non-urgent, batched processing versus real-time calls.
How to Talk to Clients About This Cost
Be upfront that API usage is typically a separate, variable cost from the build fee and retainer — either passed through directly or built into the retainer with a defined volume cap. Surprising a client with this cost after the fact, even if the total is reasonable, damages trust far more than disclosing it clearly from the proposal stage.
The Bottom Line
Token-based API costs are a real, ongoing part of running AI agents — not a rounding error. Estimating them honestly upfront, routing tasks to the right model tier, and monitoring actual usage against assumptions protects both agency margins and client trust.
Discussion