How much does AI automation cost? A payback worksheet
The build, run and upkeep costs of an AI automation, priced from published pages, plus a two-week time study to measure what the work costs you today.
How much AI automation costs depends on what you pay to build it, run it and look after it, set against what the work costs you today.
In short:
- An AI automation costs money in three places: building it, running it and keeping it working. A build quote only covers the first.
- Build: our fixed price is $8,000 for one workflow or $20,000 for three. The $2,500 workflow diagnostic is credited in full to the build, so after the diagnostic an $8,000 build costs you $5,500 more.
- Run: model tokens, a platform plan and hosting. In our example of 1,500 invoices a month on Claude Sonnet 5.5, the tokens come to about $16.50 a month.
- Upkeep: $1,500 to $5,000 a month if someone else runs it, or a few hours of a developer’s time each month if you run it yourself.
- The other half of the sum is what the work costs you now, and the way to find out is a two-week time study. In our worked example, invoice entry takes 155 hours a month at $34.80 an hour, and an $8,000 build pays back in about 3.5 months. At 200 invoices a month, the same build never pays back if you pay for managed upkeep.
How much does AI automation cost to build, run and keep working?
| Cost | What it covers | When you pay | Where the number comes from |
|---|---|---|---|
| Build | Designing the workflow, connecting your tools, testing on your data | Once | A fixed quote |
| Run | Model tokens, an automation platform plan, hosting | Monthly, rising with volume | The vendors’ pricing pages |
| Upkeep | Fixing breakages, updating models and prompts, small changes | Monthly | A managed plan, or your own staff’s hours |
It helps to price each line separately, because each one moves for its own reasons. Build cost depends on how many tools the workflow touches. Run cost grows with volume. Upkeep depends on how often your tools and the models change underneath you.
Build: what does a fixed price look like?
Our agent builds cost $8,000 for one workflow or $20,000 for three, with 50% paid before we start and 50% on delivery. That covers one agent running in production, the connections to your tools (Odoo, HubSpot or Slack, for example), testing on your real data, and the code and prompts, which you own.
If you start with the two-week workflow diagnostic, its $2,500 is credited in full against the build. So the total for diagnostic plus one workflow is still $8,000.
When you compare quotes from anyone, ours included, check that each one answers the same questions:
- Is the price fixed for a defined workflow, or billed by the hour?
- Are the connections to your existing tools included?
- Is testing on your own data included, before anyone depends on it?
- Who owns the code, the prompts and the data afterwards?
- Are tokens and hosting included, marked up, or billed to you directly?
- What happens after launch, and what does it cost?
Run, in brief: tokens, platform plans and hosting
Model tokens are priced per million on each provider’s page: Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and OpenAI lists gpt-6.1-terra at $1 and $5. If the workflow runs on an automation platform, n8n listed its Starter cloud plan at €20 a month, billed annually, for 2,500 workflow executions when we checked, and Zapier lists its Professional plan from $19.99 a month, billed annually. We bill tokens and hosting to your own accounts at cost with no markup, and our post on what an AI agent costs per month works through the per-run numbers.
Upkeep: who keeps it working, and what does it cost?
An automation that worked on launch day can break later without anyone touching it. The model provider retires the model you built on, a connected tool changes its API, or the documents start arriving in a new format. Someone has to notice, fix it and check that the fix didn’t break something else.
You have two ways to pay for that.
Pay someone to run it. Our managed plans have a 3-month minimum, then 30 days’ notice to cancel:
| Plan | Price | Workflows | Changes included | Tests rerun |
|---|---|---|---|---|
| Basic | $1,500 a month | 1 | Up to 4 hours a month | Monthly |
| Standard | $3,000 a month | Up to 3 | Up to 10 hours a month | Every two weeks |
| Extended | $5,000 a month | Up to 6 | Up to 20 hours a month | Weekly |
Or run it yourself. Then upkeep costs the hours your own developer spends on it. The US Bureau of Labor Statistics’ occupational wage survey puts the median wage for software developers at $65.38 an hour (May 2025). Wages are only part of what an employee costs you. In the BLS employer costs survey, wages made up 70.0% of total compensation in US private industry in June 2026, with benefits making up the rest. Dividing $65.38 by 0.70 gives about $93.40 an hour, so 6 hours a month of upkeep costs about $560. This is cheaper on paper, but it only works if someone on your team understands the system well enough to fix it.
How do you measure what a workflow costs today?
You can’t work out payback without knowing what the work costs now, and an estimate made from memory can’t be checked. A two-week time study gives you numbers that can.
- Define the workflow’s start and end. For example: an invoice arrives in the shared inbox, and it ends when the invoice is posted in the accounting system.
- Give everyone who does the work a task log. For each item they record the start time, the end time and the outcome: done, sent back, or passed to someone else. A shared spreadsheet is enough.
- Run it for two weeks, so the log catches quiet days and busy ones. If the work peaks at month end, make sure the window includes it.
- Use sampling if logging every item gets in the way. Several times a day, at random times, note what each person is doing. The share of observations spent on the workflow, multiplied by hours worked, estimates the time it takes.
- Record errors and rework separately: every item that came back, what was wrong with it, who caught it and how long the fix took. Rework is the part people forget when they estimate from memory.
- Take the volume from the system: the number of invoices the accounting system shows as posted in the period.
Then fill in the worksheet. Copy this table into a spreadsheet and replace the example inputs with your own:
| Line | What to enter | Example input |
|---|---|---|
| A | Items per month, from the system | 1,500 |
| B | Median minutes per item, from the task log | 5 |
| C | Share of items needing rework | 6% |
| D | Minutes per rework | 20 |
| E | Hours per month today: (A × B + A × C × D) ÷ 60 | 155 |
| F | Hourly cost including benefits | $34.80 |
| G | Cost per month today: E × F | $5,394 |
The payback worksheet, worked through with example inputs
The example is invoice entry in accounts payable. Every input below is an example. Swap in your own.
For the hourly cost, the BLS wage survey puts the median wage for bookkeeping, accounting and auditing clerks at $24.36 an hour (May 2025). Divided by 0.70, the wage share from the employer costs survey, that becomes $34.80 an hour. In the UK, the ONS Annual Survey of Hours and Earnings puts median hourly earnings for full-time employees at £20.81 (April 2026, excluding overtime). For your own sums, use the ASHE occupation tables for the role that does the work.
After automation, we assume a person spends 1 minute checking each invoice, and 15% of invoices still need 5 minutes of handling as exceptions. We also assume the review catches what used to be rework.
| Hours a month | Cost a month | |
|---|---|---|
| Today | 155 | $5,394 |
| After automation: 25 hours of review, 18.75 of exceptions | 43.75 | $1,522.50 |
| Saved | 111.25 | $3,871.50 |
Now take off the monthly costs. Run is an example input of $100 a month. At 3,000 input tokens and 500 output tokens per invoice, Claude Sonnet 5.5 accounts for $16.50 of that, and the rest covers a platform plan and retries.
| Managed upkeep (Basic) | In-house upkeep (6 hours) | |
|---|---|---|
| Saved each month | $3,871.50 | $3,871.50 |
| Run | −$100 | −$100 |
| Upkeep | −$1,500 | −$560.40 |
| Net gain each month | $2,271.50 | $3,211.10 |
| Build | $8,000 | $8,000 |
| Payback | 3.5 months | 2.5 months |
| First-year net after the build | $19,258 | $30,533 |
Hours saved only turn into money when the time goes to other work or saves you a hire. If the same people just have quieter afternoons, the payback exists only on paper.
When does automation not pay?
Low volume is the first thing to check. Run the same example at 200 invoices a month and the time saved drops to about 14.8 hours, or $516 a month. With managed upkeep at $1,500, the gain each month is negative and the build never pays back. Run it yourself with 2 hours of upkeep and about $30 of run cost, and payback stretches to about 27 months.
Too many exceptions is the second reason. Go back to 1,500 invoices, but suppose 60% of them still need a person. Hours after automation rise to 100, the saving falls to $1,914 a month, and with managed upkeep payback takes about 25 months.
Watch for these signs too:
- The rules change every month, so you’d be paying to rebuild the automation each time they do.
- Every item needs a person’s approval anyway, for legal or money reasons, and the approval is most of the time spent.
- The process itself is broken. Fix it first, because an automation repeats the steps it’s given, including the bad ones.
- Nobody can say where the workflow starts and ends, which means it isn’t ready to be measured, let alone automated.
Getting it measured for you
If you’d rather not run the time study yourself, our workflow diagnostic does it in two weeks for $2,500. We map 3 to 5 workflows with the people who run them, measure what each one costs, and pick one to start with. You get a written plan with a go or no-go and a fixed price for the build. If the answer is no-go, you keep the plan and the workflow map.
To see whether a workflow is worth measuring at all, start with the free two-minute workflow check.