AI cost is the whole path, not the tokens

Model prices are easy to quote and easy to compare, which is why they dominate the conversation. They are also the smallest line in most business cases. A decision that costs a fraction of a penny to infer can still cost pounds once someone has to review it, fix the mistakes it causes and support the system that runs it.

The lines in the budget

• Licences and seats, including premium connectors and any platform capacity.
• Orchestration and hosting: the adapter, queues, monitoring and secrets.
• Data preparation and retrieval: getting the right evidence to the decision, fresh enough to trust.
• Model inference, for every layer that calls a model.
• Review labour: minutes per reviewed item multiplied by how many items are reviewed.
• Error remediation: what a wrong action costs to find and put right.
• Support and change: releases, re-evaluation when a model or question changes, training.

In the pilots this site proposes, review labour and error remediation dominate. The Evidence page records vendor documentation listing inference in the order of cents per hundred thousand items at stated token assumptions; check current pricing before you rely on it. Treat that as a reason to stop arguing about tokens, not as a business case.

A formula worth writing down

For one decision, per item:

cost per item = inference
              + platform share
              + review rate × minutes per review × staff rate
              + error rate × cost per error

Two terms move the result: how many items still need a person, and how expensive the mistakes are. A cheaper model that lowers confidence and sends more items to review can raise the total. A threshold set too low can cut review and raise error cost. You find the right point with held-out labelled cases, not with a price list.

Why chat seats behave differently

A chat licence is a fixed cost per person. Its value depends on each person finding their own uses, so the return is uneven and hard to see. A decision service costs per item, and its value is measured on the same items: how many were handled, how many correctly, how many minutes of review were saved. That makes it easier to fund in stages and easier to stop.

Spend follows difficulty

Put the layers in order of cost. Rules see every item. Jev sees the items rules can't settle. A frontier model sees the smaller set that needs text written. A person sees the smallest share, the uncertain and the consequential. When the shape is right, the expensive layers only work on the cases that deserve them.

When a supplier quotes a price per token, ask for a price per correctly handled item at your review rate. If they can't answer, the business case isn't finished.