How to Calculate the ROI of an Automation
By CodexierPublished 6 min read
The return on an automation is easy to overstate: count the hours a task takes, multiply by a salary, call it savings. The honest version also counts errors, delays, build cost and the maintenance every automation needs. This worksheet walks through it in five steps, with a rule for when the answer is no.
Measure the task as it is today
Start with a stopwatch, not a spreadsheet. For two weeks, the people who do the task log each run: when it started, when it ended, and whether anything went wrong. Include the hidden parts: looking up a customer in Fortnox, waiting for a colleague to confirm, re-typing the same address into a second system. The two-week window matters because most back-office tasks have a monthly rhythm, and a single day tells you nothing about the peak.
- Runs per week, including the busiest week of the month
- Minutes per run, split into 'doing' and 'waiting'
- How often a run needs a correction afterwards
- Who does it, and what that person's hour costs fully loaded
Hours, errors and delays in kronor
Hours are the easy part: minutes per run × runs per year × the fully loaded hourly cost (salary, employer contributions, tools, a share of the office). Use the cost of the person who actually does the work, not an average. If the task is done by the owner, use the value of the sales or delivery hour it displaces, which is usually higher.
| Cost line | How to put a number on it | Often forgotten because |
|---|---|---|
| Staff hours | Minutes per run × runs per year × hourly cost | It is the only line most calculations include |
| Errors | Errors per year × time to fix × hourly cost, plus any refund or credit note | Nobody logs the corrections |
| Delays | Days a quote or invoice waits × what a day of delay costs you | The cost lands on cash flow, not the wage bill |
| Lost work | Leads or orders that never got followed up because the task queued | Hard to see; ask the sales side |
Be conservative on the last two lines. If you cannot defend the number to your accountant, halve it.
Errors deserve more attention than hours. A mistyped VAT number on an invoice costs a credit note, a re-issued invoice and a delayed payment. A supplier invoice booked to the wrong account costs an hour at year-end. These are real kronor, and they are the costs automation removes most reliably, because software does not get tired on the twentieth invoice.
Build cost plus running cost
The build price is the number on the quote; ours for a workflow automation system is fixed at 19,990 kr and published on the pricing page. The running cost is what most ROI claims leave out, and it has three parts: the tool subscriptions (an automation platform, an AI model, a connector), the API or per-call fees that scale with volume, and the maintenance hours when a connected system changes its interface or a supplier changes an invoice format.
Subscriptions
Fixed per month regardless of use. Cheap at volume, expensive if the automation runs twice a week.
Per-call fees
AI model calls, SMS, document processing. Grow with volume, so estimate from the busiest month, not the average.
Maintenance
Budget a few hours a quarter. Every automation touches at least two systems, and one of them will change.
Payback period made simple
Take the yearly saving from step two, subtract the yearly running cost from step three, and divide by twelve to get the net monthly gain. Divide the build cost by that number and you have the payback in months. It is deliberately simple: no discount rates, no three-year projections. A small-company automation either pays back within a year or it does not, and the simple version shows which.
- Net monthly gain = (yearly saving − yearly running cost) ÷ 12
- Payback months = build cost ÷ net monthly gain
- If the net monthly gain is negative, stop here: the automation costs more to run than it saves.
- Rerun the sheet with volume halved. If payback is still under eighteen months, the case is robust.
The fourth step is the one that separates a real case from an optimistic one. Volumes fall, a customer leaves, the process is reorganised. An automation that only pays back at today's peak volume is a bet, not an investment.
Results that should make you stop
We sell automation, and still: some worksheets should end the conversation. If the task takes under an hour a week in total, the build cost will never come back through hours; only a large error cost changes that. If the process is redesigned every few months, maintenance will eat the saving. If the input arrives as free-text emails with no structure, the reliability will disappoint and someone will end up checking every run, which is the manual task with extra steps.
When the sheet says yes but you are unsure which task to start with, our checklist for judging an AI audit explains how a ranked list should be built. And when you would rather have someone fill in the worksheet with you, that is exactly what the intro call is for.
- Workflow Automation SystemFixed-price build of a multi-step automation across your systems, with the running costs stated up front.
- PricingPublished prices and delivery windows for every package.
- Book a free 15-minute callBring your two-week measurements; we run the worksheet together and tell you whether it pays back.
Frequently asked questions
Should I count the owner's time at salary cost?
No. An owner's hour spent on admin displaces an hour of sales, delivery or decisions, which is worth more than a salary line. Use the value of the work that does not get done. That is also why owner-run tasks often have the strongest automation case even at low volume.
How do I estimate maintenance before the automation exists?
Count the systems it touches. Each connected system can change its API, its login method or its data format, and each change costs a fix. A few hours per quarter per automation is a reasonable planning figure for a two-system flow; more if one of the systems is a supplier's portal without an official API.
What payback period is good enough?
Under a year is clearly worth building. One to two years is worth it if the process is stable and the volume is unlikely to fall. Over two years is usually a sign that the task is too small or too irregular to automate yet, and the money is better spent on the next task down the list.
Does the calculation change for AI-based automation?
The structure is the same, but two lines grow: per-call fees, because model calls are billed per use, and error handling, because a model can be confidently wrong. Add a review step to the time estimate unless the output is low-stakes, and measure the error rate in the pilot before you trust the saving.
Not sure the numbers add up?
Fifteen minutes with an engineer: you describe the task and your measurements, we run the worksheet with you and say plainly whether an automation would pay back, and how fast.
Book a free 15-minute call