In 30 seconds

You do not need a spreadsheet from finance to sanity-check the running cost of an AI workflow. You need three numbers: how many times it runs, how much each run costs, and how much human time sits around it. This guide gets you a rough figure you can trust to one significant figure, which is all a go or no-go decision needs.

Step 1: Count the volume

How many times will this run in a month? Use a real number from the work you already do: tickets per month, invoices per week, requests per day. Volume is the single biggest driver, so get it roughly right rather than precisely wrong.

Step 2: Estimate the cost per run

Each run sends text to the model and gets text back, priced by tokens. A rough rule of thumb: a page of text is around 700 to 800 tokens. Add up what a single run sends (the instructions, the input, any documents it reads) and what it returns.

For most short SME tasks the model fee per run lands somewhere between a fraction of a cent and a few cents. If a run reads long documents every time, it will be higher. When in doubt, run ten real examples through the tool and read the actual usage.

Step 3: Add the human time

For each run, how many minutes does a person spend reviewing or correcting it? Multiply by volume, then by a loaded hourly rate. Early on this often dwarfs the model fee, and it is the number that actually decides whether the workflow pays off.

Put it together

Monthly cost ≈ (volume × model fee per run)
             + (volume × review minutes × hourly rate ÷ 60)
             + setup cost spread across the year

Sanity-check against the value

Compare the total against the time the workflow saves, valued the same way. If the saving is not comfortably larger than the cost, the workflow is not ready, or not worth it yet. A small pilot is the cheapest way to replace these estimates with real figures.

What this does not solve

This gives you a running cost, not a business case. It also will not catch a workflow that is cheap but wrong. Read it alongside what does an AI agent cost for the drivers you might have missed.

Good enough to move on when

  • You have a monthly figure built from volume, per-run cost, and review time.
  • You have checked it against ten real examples, not just a guess.
  • The saving clearly beats the cost, or you have decided to pilot to find out.