What goes into the cost of running a workflow?

Three things: the inference fee (usually measured in tokens), the human time spent on oversight and approval, and the maintenance cost. The first item is often the smallest.

How quickly should a deployment pay for itself?

For a first process in an SME, a reasonable threshold is a return within 3–6 months. A longer payback means the process is too infrequent or the per-item saving too small.

Guide

For SMB

A workflow that pays for itself: how to calculate the return on automation before you deploy

You calculate the return with a simple formula: time saved times the rate, minus the cost of running it. Do it on a napkin before you sign off on a deployment.

Return = (time saved × rate) − the cost of running the workflow.
The running cost is mostly inference, measured in tokens, plus human oversight.
Calculate it before you deploy — one table is enough to reject the weak ideas.

Calculate the return before you deploy

A good agentic workflow isn't the one that impresses in a demo, but the one that pays for itself. The difference only shows up in the numbers — and the numbers can be estimated before you deploy, on a napkin, in a few minutes. It's a filter that screens out ideas that cost more than they save.

The formula is simple:

Monthly return = (time saved × hourly rate) − the cost of running the workflow

If the result is positive and grows with volume, you have a candidate. If it's negative, you know that before you spend a single zloty on a deployment.

The savings side

You calculate savings per item, not for "the whole department."

Time per task today — how many minutes one item takes a person, by hand.
Time after deployment — how much is left once an AI assistant drafts it and a person only checks and edits.
Volume — how many items per month.

Time saved is the difference between the two times, multiplied by volume. Important: after deployment, the human's time rarely drops to zero — oversight and approval remain. Counting "down to zero" is the most common mistake, and it inflates the return.

The cost side

On the cost side there are three items, not one.

Inference — the fee for running the model. It's the cost of every run, usually measured in tokens; a token is a fragment of a word. Longer documents and longer answers mean more tokens, and therefore a higher cost per model call.
Human oversight — the time for checking and edits, which you just counted on the savings side.
Maintenance — a small, steady cost of fixes and updates to the prompt templates.

In practice, for an SME the first item — inference — is often the smallest of the three. The biggest real cost is people's time, which is exactly why it's worth counting honestly.

A worked example

Process: preparing a draft proposal. Salesperson's rate: PLN 80/h.

Item	Value
Time per proposal today	40 min
Time after deployment (draft + review)	12 min
Saving per item	28 min
Monthly volume	120 proposals
Time saved per month	56 h
Value of time saved (× PLN 80)	PLN 4,480
Inference cost (per month)	approx. PLN 250
Oversight and maintenance cost	approx. PLN 900
Monthly return	approx. PLN 3,330

With a one-off deployment on the order of PLN 12,000, the break-even point falls before the fourth month. That's the number a decision can be made on.

How not to fool yourself

A few corrections that bring the math back to reality:

Don't count time down to zero. A person still checks the output.
Count only the process you actually have, not the volume that "will probably grow."
Add a ramp-up period. The first few weeks mean more edits and smaller savings.
Check the sensitivity. If the return is positive only in the best-case scenario, it isn't a good first process.

What to do with this

Build this table for two or three processes from your short list. The one with the fastest and most certain return goes to a pilot. The rest wait. Calculating the return before you deploy doesn't require an in-house AI team — it requires one sheet of paper and honest assumptions.

Terms in this guide

Want to ship a first process that pays for itself? Tell us your case.

Tell us your case See how we help

Frequently asked questions

What goes into the cost of running a workflow?: Three things: the inference fee (usually measured in tokens), the human time spent on oversight and approval, and the maintenance cost. The first item is often the smallest.
How quickly should a deployment pay for itself?: For a first process in an SME, a reasonable threshold is a return within 3–6 months. A longer payback means the process is too infrequent or the per-item saving too small.