# How Much Does It Cost to Automate One Task With AI?

*Question — 2026-09-23 — by Mahmoud Zalt*

Bottom-up math for a single task. The three costs nobody separates, how to work out your per-run number, and when one task is not worth automating.

**Short answer.** **Divide your monthly plan fee by the number of times the task runs, then add the setup hours you spend once and the review minutes you spend every time.** That is the whole calculation. On a Sistava entry plan at $25 a month with 20,000 credits included, a task running 200 times a month has roughly 100 credits of headroom per run. If your task fits inside that, the plan fee divided by 200 is your true per-run cost. If it does not, you need a bigger allowance, and it is better to find that out in week one.

You do not want to buy a whole AI strategy. You have one specific job that eats your Tuesday, and you want to know what handing it over would cost. Fair question, and almost nobody answers it at that resolution.

The reason is that pricing pages sell capacity by the month and you are trying to buy work by the unit. Those two things convert into each other easily once you have your own volume number, and not at all before.

**Sistava** prices by plan and meters the work in credits, so the conversion is arithmetic rather than guesswork. You hire an employee for the job, connect the tools it needs, and each run draws down your monthly allowance by an amount that reflects what it actually did.

## At a Glance

- **3** Costs in every automated task: setup, per run, review
- **20,000** Credits a month on the entry plan
- **~100** Credits per run at 200 runs a month
- **$25** Entry plan, per month

## How much does it cost to automate one task?

Your plan fee divided by the number of runs, plus a one-off setup cost measured in your own hours, plus a small recurring review cost measured in minutes. There is no fourth number, and anyone quoting you a single per-task price without knowing your volume is guessing at the biggest variable.

The division matters more than the plan fee. The same subscription is expensive per unit at ten runs a month and very cheap per unit at a thousand. That is why a task that repeats daily is almost always worth pricing and a task that happens quarterly usually is not.

## The three costs nobody separates

Automation cost is not one number, it is three, and they behave completely differently. Setup happens once. Per-run cost happens every time. Review cost happens every time too, and it is the one people forget until it ruins their estimate.

| Cost | When you pay it | How to measure it |
|---|---|---|
| Setup | Once, in your first week | Hours you spend writing the process down and correcting early output |
| Per run | Every time the task runs | Plan fee divided by monthly runs, plus credits consumed per run |
| Review | Every time, at first | Minutes you spend checking output, falling as trust builds |
| Failure | Occasionally | Time to spot and redo a run that went wrong |
| Rework | When the process changes | Hours to re-specify when your business changes shape |

Setup is usually two to five hours for a well-understood task. That sounds small until you notice it is the entire reason a quarterly task fails the test: five hours of setup against four runs a year is a bad trade no matter how good the software is.

## Do the credit math before the dollar math

The question that actually decides whether a plan fits is not what does this cost, it is does this run fit inside my allowance. Take your monthly credits, divide by your monthly runs, and you get a credit budget per run. Then check whether your task lives inside it.

On the entry plan that is 20,000 credits. At 100 runs a month you have 200 credits per run to play with. At 200 runs you have 100. At 500 runs you have 40, which is tight for anything that reads a lot, and that is your signal to look at the next tier rather than to be surprised in week three.

Heavier tasks eat this budget fast. Reading a long document, checking several sources, or producing multi-page output all push a single run well past a hundred credits. Short, structured work with a defined output stays comfortably under.

The practical move is to run the task twenty times, look at what it consumed, and multiply. Twenty real runs beat any estimate, including the one in this article, because they use your documents, your tools, and your definition of done.

## Worked example: screening 195 applications a month

Marcus is an independent recruiter placing mid-level engineers. His single worst task is the first pass on inbound applications: read the CV, check it against the role brief, write a four-line summary, and tag it as yes, maybe, or no. He gets about 45 a week, so roughly 195 a month.

Manually it takes him about twelve minutes each. That is 39 hours a month, which is a full working week spent on the least valuable part of his job. He does not bill hourly, but the time is coming directly out of the client calls that actually earn him fees.

| Line | Marcus's number | Where it came from |
|---|---|---|
| Runs per month | 195 | 45 applications a week, counted over four weeks |
| Manual time per run | 12 minutes | He timed fifteen of them |
| Manual time per month | 39 hours | 195 multiplied by 12 minutes |
| Credit budget per run | About 102 | 20,000 monthly credits divided by 195 runs |
| Setup cost | 4 hours, once | Writing the role brief rules and fixing the first batch |
| Review time per run | About 90 seconds | Reading the summary and confirming the tag |
| Review time per month | About 5 hours | 195 runs multiplied by 90 seconds |
| Net hours recovered | About 34 per month | 39 hours minus 5 hours of review |

His per-run cost in money is the plan fee divided by 195, which is a small number by any reading. His per-run cost in credits came out around 60 in practice, comfortably inside the 102 he had available, so the entry allowance held.

The four hours of setup paid back in the first week. That is what a good candidate for automation looks like: high frequency, clear rules, a checkable output, and a manual version that was quietly costing a full week a month.

## When one task is not worth automating

Plenty of tasks fail this test, and knowing which ones saves you more than any saving you would have made. The pattern is consistent: low frequency, unclear rules, or an output nobody can check quickly.

- **It runs less than weekly.** Setup hours cannot amortise across four runs a year. Do it by hand and stop thinking about it.
- **You cannot write the rules down.** If your answer to how do you decide is it depends, the task is judgment work wearing an administrative label.
- **Checking it takes as long as doing it.** If review costs ten minutes on a twelve-minute task, you saved two minutes and added a dependency.
- **Being wrong is expensive.** Filings, contracts, anything with a legal consequence. Keep an accountable person on it.
- **The process is about to change.** Specifying a process you are already rethinking means paying the setup cost twice.

There is also a category where the answer is simply hire a person. When the task is really a relationship, a decision under uncertainty, or something that needs someone accountable in the room, no per-run number makes it the right call. Those tasks belong to people and they are usually the ones worth doing.

## Price your own task in thirty minutes

### Six lines and you have your number

1. **Count last month's runs** — Go to the inbox, the helpdesk, or the folder and count. Do not estimate. Estimates run about thirty percent optimistic in both directions.
2. **Time yourself on five real ones** — Average them. This is your manual cost per run and it is the number the whole model rests on.
3. **Divide your allowance by your runs** — Monthly credits divided by monthly runs gives you the credit budget each run has to fit inside.
4. **Run it twenty times for real** — Compare actual consumption against that budget. Twenty runs tells you more than any calculator.
5. **Time your review honestly** — Check how long you actually spend confirming output. Subtract it. The number that is left is your real saving.
6. **Compare against your setup hours** — If the monthly saving clears your setup cost inside a month, it was a good candidate. If it takes six months, look for a better task first.

## FAQ

### How much does it cost to automate one task with AI?

Your plan fee divided by the number of monthly runs, plus one-off setup hours and recurring review minutes. On a Sistava entry plan at $25 a month with 20,000 credits, a task running 200 times a month has about 100 credits of headroom per run. Anyone quoting a fixed per-task price without knowing your volume is guessing at the largest variable in the calculation.

### How many times does a task need to run to be worth it?

Weekly is the usual floor. Setup for a well-understood task takes two to five of your hours, and that only amortises if the task repeats often. A daily task pays back setup in days. A quarterly task never does, which is why the frequency question is worth answering before the cost question.

### How do I know how many credits my task will use?

Run it twenty times and look. Consumption depends on how much the task reads, how much it writes, how many steps it takes, and how many tools it calls, and all four are specific to your documents and your definition of done. Twenty real runs on your own material beat any published estimate, including the ones in this article.

### Should I count my own time as a cost?

Yes, on both sides. Count the hours the manual version costs you, and count the setup and review hours the automated version costs you. Models that count only the invoice make automation look free, and models that count only the saving make it look magical. The honest number is the difference between the two.

### What if the task fails sometimes?

Assume it will and price it in. Budget time to spot a bad run and redo it, especially in the first month while you are still tuning instructions. A task where a failure is easy to notice and cheap to fix is a good candidate. A task where a quiet failure costs you a customer or a fine is not, whatever the arithmetic says.

### Can I automate half a task?

Often that is the best version. Let the employee do the reading, sorting, and drafting, and keep the final decision and the send button. You get most of the time back and you keep the judgment where it belongs. Marcus does exactly this: the summaries and tags are drafted, and he confirms every one in about ninety seconds.

One task is the right unit to think in. It is small enough to measure honestly, big enough to matter to your week, and specific enough that you will know within a month whether it worked.

Pick the job that costs you the most hours and the least judgment, count last month, and run it twenty times. Whatever the meter says after that is your answer, and it will be a better one than any range you can look up.

**Tags:** automation-cost, cost-per-task, credits, ai-employee-cost, small-business