The chief operating officer of a wholesale distributor in the Netherlands is watching a demo. A robot opens a mailbox, picks up a PDF invoice, reads the purchase order number, keys the data into the accounting system and closes the task in forty seconds. The vendor quotes a licence fee that looks trivial next to the salaries of the three people who enter invoices by hand today. The board wants to know one thing: what will this cost over three years, and when does it pay back?
In the demo the robot processed a perfect invoice. In reality some invoices arrive as scans, some carry the order number in a different place, some refer to orders that are not in the system. Within the year the finance department will upgrade its ERP, and the robot’s technical account will need a new password every quarter. None of that is in the licence price, and all of it affects the business case.
This article breaks the cost of process automation into its parts: licences, building the process, maintenance and organisational change. It shows when to choose RPA, when an API integration and when an AI-based solution, and how to calculate the return so that it survives the CFO’s questions. The thesis is simple: the licence is usually the smallest line, and the economics are decided by the share of cases that go through without a human and by the cost of keeping the automation alive.
What exactly are you paying for when you buy process automation?
“Process automation” is a label for three different techniques with different cost structures. It is worth separating them before you compare proposals, because the vendor of each will present its own as the answer to everything.
RPA, robotic process automation. Software mimics a person working in the user interface: clicking, copying, pasting, filling in forms. It needs no changes to the systems being automated, so it can be deployed quickly. You pay for robot and orchestration licences, for building each process and for maintenance, which is high because the robot depends on how the screens look.
API integration and workflow automation. Systems exchange data directly through programming interfaces, and the process rules live in code or in an iPaaS or workflow platform. Delivery usually takes longer and needs developers, but the result is more durable and cheaper to maintain. Our glossary entry on integration of systems covers the basics.
AI-based automation. Machine learning and language models read data from documents, classify messages, extract information from unstructured text and increasingly make decisions within defined rules. You pay for a platform or for model usage (often variable, depending on the volume processed), for preparation and validation on your data and for quality control of the output.
In practice a good project combines these techniques. An invoice arrives by email, an AI model reads it, an API integration writes the data into the finance system, and RPA handles one old system that has no programming interface. Our glossary entry on process automation brings the types together in one place.
When should you choose RPA, an API integration or AI?
Choosing the technique is the first cost decision, because it sets the ratio between build cost and maintenance cost. Four questions are enough to make it.
Do the systems have APIs? If they do, API integration is usually the first choice. A robot clicking through the interface of a system with a documented API is faster to start and more expensive for its whole life. The exception is when API access carries an extra charge that changes the maths, or when the system vendor does not allow its use.
How often does the interface change? RPA works well on stable systems that rarely get updates. On cloud applications updated every few weeks, the robot will need frequent fixes. Ask your IT team how often the screens of the systems you want to automate actually change.
Is the input structured? If the process starts from a form or a file with a fixed format, rules are enough. If it starts from an email written by a customer, a scanned document or PDFs in varied layouts, you need an AI component to read the data. Then the key question becomes the share of cases the model reads correctly without verification.
How many exceptions are there? A process in which 95% of cases run identically automates well. A process in which one case in three needs a human decision will automate only partly, and the cost of handling exceptions can eat the savings. Before ordering automation, count the exceptions in a sample of real cases from last month.
Ask these questions before talking to a tool vendor, not after. A company selling an RPA platform will find an RPA use case in every process, and a company that builds integrations will propose an integration. Your job is to pick the technique that fits the process, not the process that fits the technique. An overview of integration patterns is in our article on software integration patterns and approaches.
What do Power Automate and UiPath licences cost in 2026?
Licence prices are the only line of the business case that can be checked publicly, and only in part. The figures below come from the official pricing pages as checked in November 2026. Prices in euros and pounds can differ from US dollar list prices and large customers negotiate discounts, so treat them as a reference point, not a quote.
Microsoft Power Automate. According to Microsoft’s pricing page, Power Automate Premium costs USD 15 per user per month. Power Automate Process, licensed per bot, costs USD 150 per month, and Power Automate Hosted Process, in which Microsoft also provides the virtual machine the robot runs on, costs USD 215 per month. All of these are paid yearly. The Process Mining add-on, available only with Premium, costs USD 5,000 per tenant per month. Microsoft notes that prices on the page are for marketing purposes and may differ from checkout prices.
UiPath. On UiPath’s pricing page, the Community plan is free, Basic starts at USD 25 per month, and Standard and Enterprise require contacting sales. The company also says an updated Basic offering is coming. For most organisations automating production processes with unattended robots, the real price only emerges in a sales conversation.
When comparing licences, watch three things that move the total more than the list price. The first is the difference between attended robots (running on an employee’s computer and triggered by that person) and unattended robots (running on their own on a server). The latter cost more, but they deliver the savings in batch processes. The second is infrastructure: an unattended robot needs a Windows virtual machine and licences for the applications it works in, unless you buy the hosted variant. The third is technical accounts: each robot signs into systems as a user, and in some systems that means an extra user licence.
If your company already has Microsoft 365, check what is included. Basic Power Automate flows between Microsoft services are often available within existing subscriptions, but premium connectors and RPA robots need paid plans. This is a common source of surprise when a pilot on standard connectors costs nothing and the production rollout does not.
What does it cost to build the first automated process?
There are no public price lists here, because the cost depends on the process. What can be shown is what it consists of, with indicative ranges and clearly stated assumptions. The figures below are our estimate for illustration, not market data.
Building an automation has five stages: process analysis (steps, rules and exceptions), solution design (technique, error handling, logging, credential security), build, testing on real cases, and deployment with documentation. Analysis and testing usually take longer than buyers expect, because that is when the exceptions nobody wrote down appear.
As an estimate, a simple process with one system, structured input and a few exceptions is a few to a dozen or so person-days of work. A medium process with two or three systems and document reading is a dozen to several dozen person-days. A complex process with many systems, decisions and an AI component can take several dozen person-days or more. The number of days depends most on the number of systems, the number of variants and exceptions, input data quality, security requirements and whether the process is documented.
To turn days into euros you need a rate. According to ARDURA’s 2026 benchmark, a senior backend developer from Poland costs about EUR 335–485 per day. At EUR 400 per day, twenty person-days is EUR 8,000 and forty is EUR 16,000. Add the time of your own people: the process owner who answers questions about rules and the staff who test the output. That cost is often left out, and without it the project will not succeed.
Start with one process that has a high number of repetitions, few exceptions and a measurable result. The first process is the most expensive because it includes setting up infrastructure, technical accounts, standards and monitoring. The next ones build on that foundation and cost less.
What does bot maintenance cost, and why do proposals stay quiet about it?
Maintenance is the line that decides whether automation pays in years two and three, and also the one proposals say least about. The reason is simple: a build is easy to sell as a project with a price and a date, while maintenance looks like a cost that will “sort itself out”.
Maintenance covers several kinds of work. Fixes after changes in the automated systems: a new application version, a changed screen, an extra form field. Exception handling: cases the robot did not process must reach a person, and someone must check whether a pattern repeats and the automation should be extended. Technical account management: password changes, certificate renewals, permissions. Monitoring: someone must know the robot failed overnight before a customer does. Changes to the process itself: new business rules, new document types, new trading partners.
What does it cost? We know of no credible public study that gives a universal figure, and we do not believe one exists, because the result depends on how stable the systems are. So instead of quoting a percentage, we suggest calculating it for your own case: ask IT how many updates the systems in scope received last year, estimate how many hours adapting the robot to each one takes, and add fixed time for monitoring and exception handling. For a robot working on the interface of a frequently updated application, that number will be clearly higher than for an integration through a stable API.
In the contract, treat maintenance as a separate service: its scope (what it covers and what counts as new automation), response time for failures, the billing method (a subscription with an hours cap or an hourly rate) and reporting. If a supplier will not commit to maintaining what it built, that tells you how it rates the durability of its own solution.
How do you calculate whether automation will pay back?
The payback calculation is simple in formula and hard in data. Annual benefit is the hours of work actually saved multiplied by the fully loaded hourly cost of the employee, plus the value of errors avoided, plus any benefit from faster turnaround. Annual cost is licences, infrastructure, maintenance and amortised build cost. Payback comes when cumulative benefit exceeds cumulative cost.
The key word in that formula is “actually”. A demo shows the time to process one case, but the saving depends on the share of cases that go end to end without a human. A case the robot started but handed to a person often costs more than one processed manually from scratch, because the person first has to work out where the robot stopped.
An illustrative example, with assumptions stated for the calculation. A team processes 1,000 invoices a month, and manual entry takes on average 6 minutes each, so 100 hours a month. Assume a fully loaded cost of EUR 45 per hour for a finance clerk in Western Europe. If automation processes 70% of invoices without a human, the saving is 70 hours a month, EUR 3,150, or EUR 37,800 a year. On the cost side: a build of 30 person-days at EUR 400 is EUR 12,000; licences for one hosted robot and a few user licences come to around USD 3,500 a year at list price; and assume maintenance of two person-days a month, about EUR 9,600 a year. On these assumptions, the automation pays back within the first year.
Now change one assumption. If 40% of invoices go through without a human rather than 70%, the annual saving drops to EUR 21,600, and after maintenance and licences only around EUR 9,000 a year is left to repay the build. Payback moves well into the second year. If the systems also change often and maintenance takes four person-days a month, the project never pays back. The licence price is the same in every variant and barely affects the result. The same logic applies with higher in-house rates: the build gets more expensive, but the decisive variables remain the straight-through rate and maintenance.
For a market reference: in Deloitte’s Global Intelligent Automation Survey 2022, conducted among 479 executives in 35 countries, 74% of respondents were already implementing RPA, and organisations expected an average cost reduction of 31% over three years. An expectation is not a result, and your business case should rest on your own data. We go deeper into calculating return on software investment in the calculator in the CEO’s head.
How do you lay out a three-year automation budget?
The payback calculation answers whether automation pays. The budget answers a different question: how much money to set aside, and when. The board needs both answers, and the second is often skipped, because vendors present the cost of implementing one process rather than an automation programme spread over time.
Structure a three-year budget in four layers. The first is one-off start-up costs: infrastructure, platform configuration, security standards for technical accounts, monitoring and training for the team that will handle exceptions. You pay these once, with the first process, and they do not recur. The second layer is implementing further processes, each priced separately in person-days, on the assumption that later ones cost less than the first because they reuse the foundation. The third layer is recurring costs: licences, infrastructure and maintenance, which grow with the number of live automations. The fourth is business-side cost: process owners’ time, exception handling and reviewing results.
The key point is that the third layer grows with every new process. The first year of a programme usually looks good, because maintenance covers one or two robots. By year three, with a dozen or more automations, maintenance can become the largest line in the budget. If you do not plan for it from the start, year three brings the question of whether to switch some robots off because keeping them alive costs more than they save.
Plan an annual review of the automation portfolio as well. Processes change, some automations lose their purpose because a system or a way of working has changed, and some are worth rebuilding from RPA to an API integration once the system vendor exposes an interface. A review with clear criteria (savings, maintenance cost, number of failures) lets you retire robots that no longer pay rather than keep them out of habit.
Finally, currency and contracts. Global platform licences are often priced in US dollars and sold as annual contracts paid up front. If you budget in euros or pounds, allow for exchange rate movement and renewal dates. Record in the budget the date by which you must decide whether to renew or reduce licences, because once an annual contract auto-renews, that decision moves out by another year.
Why should the process be fixed before it is automated?
Automating a bad process gives you a faster bad process. It sounds like a platitude, but the data backs it. In the same 2022 Deloitte survey, the most frequently cited barrier was immature and fragmented processes that are difficult to manage with a unified flow, and it had been the top barrier in the past four editions, according to Deloitte Insights.
In practice fragmentation looks like this: the same process done three ways in three branches, rules that live in the heads of experienced staff, exceptions handled “as usual” with no definition of what “as usual” means. A robot needs one unambiguous version of the process. If there is none, the supplier will either invent one (and automate something different from what people do) or build three versions (and triple the maintenance cost).
So the first stage should be a process description: steps, decisions, rules, exceptions and data. BPMN 2.0 lets you do this in a way both business and supplier understand, as we describe in our article on modelling the process in BPMN before ordering software. Often at this stage it turns out that some steps can be removed rather than automated.
Process mining tools help you see how a process really runs, based on system logs. They have a price, such as the Power Automate Process Mining add-on at USD 5,000 per tenant per month, so they pay off at larger scale. For the first process, a few workshops with the people who do the work and an analysis of a sample of real cases are usually enough.
What do AI and agents change in the automation business case?
Language models have widened what can be automated. Processes that start from unstructured text, such as complaints arriving by email or document classification, were too irregular for classic RPA. Today a model can read the content, extract the data and propose an action, and rules and integrations do the rest.
The cost structure changes too. A robot licence is fixed, while model usage cost usually depends on the volume processed, so it grows with volume. Then there is quality control: a model can be wrong in ways that are hard to predict, so you need verification mechanisms, output sampling and, for high-stakes processes, human approval. That cost has to be included from the start.
It also pays to keep some distance from the marketing. In a press release of 25 June 2025, Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls. It also pointed to “agent washing”, the rebranding of existing products such as chatbots and RPA tools as agentic. Gartner estimated that of the thousands of vendors describing themselves as agentic, only about 130 offer genuinely agentic solutions.
The takeaway for a buyer: use AI where the problem is unstructured data, not where rules are enough. A pilot should measure the share of correct results on your data and the cost per processed case, not the impression left by a demo. How to budget the path from pilot to production is covered in our article on AI implementation cost from PoC to production.
Which factors drive the price of automating a process?
The table below collects the factors that most affect build and maintenance cost. Use it to assess your process before talking to a supplier, and to understand why two proposals for “the same” process differ in price.
| Factor | Lowers cost | Raises cost | What it affects most |
|---|---|---|---|
| System access | Documented API | User interface only | Maintenance |
| System stability | Rare updates | Updates every few weeks | Maintenance |
| Input data | Forms and fixed-format files | Scans, emails, varied document layouts | Build and straight-through rate |
| Exceptions | Under a few percent of cases | One case in three needs a decision | Savings and manual handling |
| Process description | Written, one variant | In people’s heads, several variants | Analysis and build |
| Number of systems | One or two | Many, including legacy | Build and testing |
| Security requirements | Non-sensitive data | Personal, financial, regulated data | Design and audit |
| Volume | Steady, predictable | Seasonal peaks | Licences and infrastructure |
| AI component | None, or simple field extraction | Content interpretation, decisions | Variable cost and quality control |
| Business owner | Named and available | None, or rotating | Project duration |
If your process lands in the “raises cost” column on most rows, that does not mean it is not worth automating. It means the business case must allow for a higher build and maintenance cost, and the first step may be tidying up the process or building the missing API.
How do you choose a supplier, and what should you ask before signing?
The automation market splits into partners of specific platforms, who mostly implement one tool, and software engineering firms that build integrations and solutions shaped around the process. Either can be a good choice, but know which one you are talking to, because it shapes the technique proposed.
Before signing, ask eight questions. On what basis did you choose the technique for this process, and what alternatives did you consider? What straight-through rate do you assume, and how will we measure it in the pilot? How do you handle exceptions, and where do cases the robot could not process go? In whose name are the licences, and who owns the code, configuration and documentation? What does maintenance look like: scope, response time, billing? How are technical account credentials stored? What happens when an automated system gets a new version? How will you transfer knowledge to our team if we want to take over maintenance?
Watch the licences in particular. If the supplier buys licences in its own name and resells them as a subscription, changing supplier can mean losing working automation. It is safer when licences are issued to your company and the supplier has administrative access you can revoke. Under GDPR, also confirm where the robot and any AI model process personal data and that a data processing agreement covers it.
Insist on a pilot with measurable criteria. A good pilot covers one process on real data for a few weeks and ends with numbers: straight-through rate, processing time, error count, cost per case. The decision to scale is made on those numbers, not on a presentation.
Who should own automation inside the company?
Automation without a business owner degrades quickly. Robots stop working after system changes, exceptions land in mailboxes nobody reads, and a year later the finance team is back to manual work without telling anyone the robot has stopped. So every automation needs a process owner accountable for the outcome and a technical owner accountable for it running.
With a handful of automations, it is enough that these roles are assigned and written down. With dozens, consider an automation centre of excellence: a small team that maintains standards, chooses processes against shared criteria, keeps a register of robots and their owners and looks after security. It does not need to be large, but it does need a mandate.
Low-code tools such as Power Automate let business staff build simple automations themselves. That is both an opportunity and a risk. An opportunity, because people who know the process can remove small annoyances without waiting for IT. A risk, because without rules you end up with hundreds of undocumented flows running on personal accounts and company data. We discuss this in our article on low-code and no-code platforms. For automation within the Microsoft ecosystem, see also our piece on Copilot and AI in automating organisational processes.
What are the key takeaways before you order automation?
Business process automation can deliver real savings, but the payback rarely depends on the licence price. It depends on how many cases go through without a human, what it costs to build the process and what it costs to keep it alive while the systems around it change.
Remember six things. First, distinguish RPA, API integration and AI-based automation, and choose the technique that fits the process, not the vendor’s offer. Second, treat list prices as a reference: Power Automate Premium is USD 15 per user per month and the per-bot plans USD 150 and 215 per month, while larger UiPath plans are quoted individually. Third, estimate build effort in person-days and add your own people’s time. Fourth, calculate maintenance separately, based on how often the automated systems change, and put it in the contract. Fifth, build the business case on the straight-through rate and check how the result moves when that rate drops. Sixth, tidy up the process before automating it and name its owner.
If your processes need API integrations, custom components or a combination of techniques, see how we deliver software development and systems integration, or book a call. We start with a discovery phase of one to two weeks that ends with a scope and a price.