Automation ROI calculator: a formula you can check

An automation ROI calculator is a formula in a form. Most online ones ask for hours saved and an hourly rate, then multiply. That is a fine start, but it leaves out the cost of running the automation, which for an agent that makes data calls is not zero.

This post gives you the formula, the inputs that people get wrong, and a way to fill in the run cost from real prices. There is no hosted calculator on the looot site. You get the arithmetic, and you can ask an agent to do it for your numbers.

What is automation ROI?

Automation ROI compares what an automation saves with what it costs. Savings are the work hours that no longer need a person, valued at what those hours cost you. Costs are the building time, the tools and the per-run charges.

The simplest form is:

Net gain = (hours saved x hourly cost) - (build cost + run cost)

ROI is then the net gain divided by total cost, shown as a percentage if you like. The formula is easy. The hard part is getting honest numbers for each input.

What inputs does the calculator need?

Run cost is the one that changes with how you build. A task done by a script you own has near-zero run cost. A task done by an agent that calls data APIs has a per-call cost. For a list of three hundred companies it can be small or not, depending on the endpoints.

How do I find the run cost for an agent?

Count the calls and read the prices. For each item the agent handles, list the endpoints it will call. Multiply the number of items by the calls per item, and then by the price of each call. What 1,000 lookups cost and AI agent cost, model tokens plus data calls show the arithmetic with live prices.

The live table shows today's prices for three calls that often appear in a list-building automation. Use the row that matches your job.

Model token cost is a second part of run cost. Read your model vendor's price page for that, and add a margin.

A worked example with assumptions

Take a task where someone looks up a work email for each new lead and enters it in a sheet. Suppose it takes four minutes per lead by hand, and you handle 300 leads in a batch. That is 20 hours.

Assume the person's loaded cost is a number you know, call it H per hour. Hand cost is 20 x H. The agent version needs about one hour to build and check, so build cost is 1 x H. Run cost is 300 calls times the price of the email endpoint you read in the table, plus a margin for items that need a second provider.

Net gain is 20 x H minus 1 x H minus the run cost. Because the run cost for 300 calls is usually small next to 19 hours of work, the result is positive unless the endpoint is expensive or the failure share is high. Put your own H and your own price in, and the number is yours. I used assumptions on purpose, because a made-up dollar figure would look like a measurement.

What makes the result wrong?

Four things tilt the answer.

Failure share. If 30 percent of leads return no email and a person still handles them, the savings drop by that share. Measure the miss rate on a small sample.

Review time. Checking the agent's output takes time, and you should count it as hours spent, not as free.

Repeat use. A one-off task pays back once. A task you repeat pays back each time. But the repeat only counts if you actually trigger it, and since scheduled runs are not available yet, each repeat needs someone to start it.

Price drift. Provider prices change. Read the price at the time of each run, not once.

How do I measure hours saved honestly?

Time the task by hand on ten items, not one. The first item takes longer because you are working out the steps. Take the average of the last eight. Then ask what share of items needs a person even after automation, and subtract that share.

Be careful with batch effects. A task done in one sitting is faster per item than the same task spread across a week, because there is no switching cost. If your real work is spread out, time it spread out. Otherwise you will count hours that never existed.

Also count the review. If an agent produces a table and you spend twenty minutes checking it, those twenty minutes belong in the cost side. Skip the check and you will find the errors later, at a worse time.

When is automation not worth it?

When the task is rare, short or already cheap. A job you do twice a year and that takes an hour will not repay a day of building. When the task changes each time, the automation needs rework each time. And when a mistake is expensive, the review cost can eat the savings.

A fair rule of thumb is to automate when you do the task often, the steps stay the same and a wrong result is cheap to catch. If two of those three are false, do it by hand and revisit later. For more on a first job to automate, see the data enrichment use case.

Try it: a prompt for your agent

This prompt asks the agent to compute the ROI for a task you describe. It needs your own numbers, so it asks for them.

The agent will stop after step 2 and wait for your numbers.

What it cannot do

• It does not forecast savings you have not measured. Time one real task first.
• It does not include model token cost unless you add it from your model vendor's price page.
• It does not host a calculator page. The formula runs in the agent's reply.
• It does not rerun the check on a timer. Recomputing ROI over time needs scheduled runs, not available yet.

Last checked 2026-10-03 against the live catalog and the looot skill file. Author: the looot team.