Most AI business cases are built on a token calculator. Input tokens times price, output tokens times price, multiply by expected volume, present to the board. The number is always encouragingly small, and it is almost never the real cost.

The lines that are missing

• Retries and repair. Real traffic produces malformed outputs, timeouts and refusals. Budget 15–30% more calls than the happy path suggests.
• Retrieval. Embedding, storing and re-embedding your corpus is a recurring cost that grows with your content, not with your usage.
• Context growth. Conversations get longer, prompts accumulate instructions, and per-call cost creeps up quietly quarter over quarter.
• Human fallback. Every escalated case costs support time. At 20% escalation on a task a human handles in six minutes, the human line usually dwarfs the token line.
• Evaluation. Running your regression set on every change is compute you pay for and engineering time you schedule.

Latency is a cost, not a feature request

A nine-second response is not a slower version of a two-second response — it is a different product with a different abandonment rate. If your agent sits in a checkout or a support chat, model latency converts directly into conversion loss, and that belongs in the same spreadsheet as the API bill.

The break-even question

For an automation use case the maths is simple and worth doing before anything is built: what does the task cost today in fully loaded human minutes, what fraction can the agent handle unaided, and what does the remaining fraction cost after escalation overhead. If automating 60% of a task saves less than it costs to run and maintain, the honest answer is to fix the workflow instead.

An agent that handles 60% of a task perfectly and hands over the rest cleanly beats one that attempts 100% and is wrong 15% of the time. The second design pays for its own mistakes twice.

What we build into the estimate

When we scope an AI feature, the cost model ships with the proposal: cost per successful outcome at three volume tiers, the escalation rate we expect, and the point at which a smaller model or a cached path becomes cheaper than a larger one. It is a boring artefact, and it is the reason these projects survive their first invoice.