What is cost per decision?

It is the total LLM spend to produce one automated decision, including any review or retry overhead you model. It is the unit economics figure you compare against the marginal cost of a human making the same call.

Why does model accuracy matter to cost?

When the model is wrong, you often pay twice — once for the bad decision and again for the human escalation or rework it triggers. Lower accuracy pushes the effective cost per correct decision up, sometimes past the human baseline.

How is break-even calculated?

The tool finds the human cost per decision at which AI and human total costs are equal, and the minimum model accuracy at which AI remains cheaper once escalation costs are included. Below that accuracy, automation can cost more than people.

Does this include infrastructure beyond the API call?

It models the LLM call cost and human comparison directly. For a fuller picture, fold queue, storage, and engineering overhead into your per-call figure before entering it.

Is cheaper always better?

No. A model that is cheap but inaccurate can raise total cost through escalations and errors. Use the accuracy and escalation inputs to find the genuinely cheapest path to a correct decision, not just the cheapest call.

AI Cost-per-Decision Calculator

AI cost-per-decision calculator

Automating decisions with an LLM only pays off if each automated decision is cheaper than the human alternative — after you account for the times the model is wrong. This calculator turns your volume, per-call cost, human baseline, and model accuracy into a clear unit economics picture: cost per decision, monthly spend, savings, and the break-even accuracy below which automation stops paying.

How it works

You enter decisions per day, the cost of one LLM call, the human cost per decision, the model’s accuracy, and what an escalation costs when the model is wrong. The tool computes raw AI cost per decision, then an effective cost that adds the expected escalation cost of incorrect decisions. It compares that against the human baseline to show daily and monthly savings, and solves for the minimum accuracy at which AI stays cheaper than humans.

Tips and notes

Model the failure cost. A wrong fraud flag or bad moderation call usually triggers human rework — include it as the escalation cost.
Use realistic accuracy. Vendor benchmarks rarely match your data; use your own evaluation numbers.
Volume amplifies everything. At high volume, a tenth-of-a-cent difference per call is real money — check the monthly figure.
Re-run per model tier. A pricier, more accurate model can win on effective cost even when its per-call price is higher.