RPA ROI Calculator

What this tool does and who it is for

The calculator turns a process you run by hand into a payback figure: hours removed and errors avoided on one side, build cost, licences and maintenance on the other. It is for whoever signs off the project — an operations or finance lead deciding whether a bot is worth it, or a consultant putting a number in front of a client.

Two neighbouring tools: the automation ROI calculator is the general version for any automation rather than screen-driving robots, and the team automation savings dashboard tracks savings after the fact, once the bot is live.

How to read the output, with a worked example

Say a finance team spends fifteen hours a week keying supplier invoices from PDFs into an accounting package. That is 780 hours a year. At a fully loaded cost of £17.67 an hour — what a £30,000 employee actually costs once employer National Insurance and pension are added, which the savings dashboard above works through — the gross annual saving is £13,783, or £1,149 a month.

That is the easy half, and the only half most calculators ask about.

Licences do not scale one to one with the process. Microsoft lists Power Automate Process, the unattended desktop-flow product, at $150 per bot per month paid yearly, and attended desktop flows at $15 per user per month on Premium — two different products at a tenfold difference. In the footnotes on the same page sits the line that decides your budget: a bot assigned to a machine executes one unattended desktop flow at a time, and concurrent runs need multiple bots.

So time the batch, not just the saving. Say the robot handles one invoice in 45 seconds; measure yours rather than trusting that. At 320 invoices a night that is four hours of run time. If the window between the last person logging off and the first arriving is two hours, one bot cannot finish it. Two concurrent streams means two licences: $3,600 a year, not $1,800. At 480 invoices it steps again, to three bots and $5,400, because licences buy whole bots and do not divide.

I have deliberately not converted those dollars into the pound saving to hand you one net number — take the list prices to your own accounts at today's rate. The shape is the point: the licence line is one line per concurrent run.

Maintenance is the other missing line. Two hours a month keeping selectors working, at the same £17.67, is £424 a year, and that is before the target application first changes its interface. Exceptions come off too: if one invoice in ten falls out for a human, a tenth of the saving never leaves the wage bill, so £13,783 becomes £12,405.

Payback is therefore not the build cost divided by £1,149 a month. It is the build divided by what survives licences, maintenance and exceptions — roughly a fifth less here before a single dollar is converted. A model without those lines does not produce an optimistic payback so much as an unfalsifiable one.

The method

Savings. Labour = hours saved × loaded hourly cost × the share of the process actually automated, which is where the exception rate belongs. Error reduction = cost per error × error rate × volume, counting only errors a robot cannot make, which usually means transcription rather than judgement. Cycle time = the value of finishing sooner, bankable only where it changes something measurable, such as invoices going out earlier and cash arriving earlier.

Costs. Build, once, including process documentation and acceptance testing rather than development alone. Licences, recurring, stepped by concurrent runs. Maintenance, recurring, and the term that compounds as the estate grows.

Payback in months = build ÷ (monthly saving − monthly recurring cost). A negative denominator is a real answer: the process should not be automated at this volume.

What moves the payback figure

Common mistakes

Licensing one bot for a process that needs two. The commonest budgeting error in unattended RPA, and it comes from a vendor footnote rather than anything obscure.

Valuing the hour at gross salary. Roughly 15% low for a UK employee once employer NI and the minimum pension contribution are counted.

Ignoring the exception rate. A robot handling 90% of cases removes 90% of the hours, and the residual often takes longer per item than it used to, because the human has lost the routine.

Counting the same benefit twice. If reduced rework is already inside the labour saving, do not add it again as an error cost.

Treating maintenance as a rounding error. It decides whether twenty robots are an asset or a second job. What is workflow automation covers why API-based automation carries less of it than screen-driving does.

Frequently Asked Questions

How do you calculate RPA ROI?

Add the labour saving (hours removed × fully loaded hourly cost × the share of the process automated), the avoidable error cost, and any bankable cycle-time gain. Subtract the one-off build, the recurring licences and the recurring maintenance. Payback in months is the build cost divided by whatever monthly saving is left once the recurring costs come out.

What is a realistic payback period for RPA?

It depends almost entirely on volume, exception rate and how stable the target application's interface is, which makes a borrowed benchmark close to useless. Figures like 300–500% first-year return circulate widely, but every trail I followed led to a vendor case study or an unsourced blog post, so this page does not repeat them.

How much does an RPA bot licence cost?

Microsoft's public pricing, checked on 21 August 2026, lists Power Automate Premium at $15 per user per month for attended desktop flows, Power Automate Process at $150 per bot per month for unattended runs, and Hosted Process at $215 per bot per month where Microsoft supplies the virtual machine. All three are yearly-commitment prices, and list prices move.

Do I need more than one bot licence?

Yes, if runs have to overlap. Microsoft's pricing page states that a bot assigned to a machine executes one unattended desktop flow at a time and that concurrent runs require multiple bots. So the question is not how many processes you are automating but how many must run simultaneously inside the window you have.

What ongoing costs belong in an RPA ROI calculation?

Licences, the hours spent repairing selectors when applications change, the infrastructure the robots run on if you host it yourself, and the human handling of exceptions. A model with only a build cost and a saving will always show a return, which is a good reason not to trust it.

What to do with the result

This calculator prices a process; it cannot tell you whether the process is automatable at all. That question is about interface stability, whether the application exposes an API you could use instead of driving the screen, and how many exceptions hide in the ten per cent nobody documented. The RPA script generator shows the shape of the automation before you commit to a figure, and API automation for businesses covers when screen-driving is the wrong tool entirely — which, where an API exists, it usually is.

If the payback looks worth having and the build is the part that keeps not happening, that is the work I take on: automation in n8n, Make.com and Python where an API exists, screen-level automation where it does not, with the exception handling and logging that makes the maintenance line predictable rather than a surprise. You can see how I scope projects on Upwork or Fiverr, or email iam.mbilalvirk@gmail.com with the process and the figures this calculator gave you. Bringing the volume, the window and the exception rate turns a quote from a guess into an estimate.