RPA (robotic process automation) executes predefined software actions, often through the screen. An AI orchestration platform distributes work among agents, bots, and people, tracks each case, and decides who takes the next step. Major RPA makers now sell both as one package.
This page explains RPA (UiPath, Automation Anywhere, Power Automate Desktop), where it breaks in practice, documented maintenance costs and unproven claims, what orchestration with agents does differently, and how RPA makers are shifting toward agents. You get vendor prices, a comparative study with measurements, one administrative task performed three ways, and a measure for choosing yourself.
The short answer
- RPA follows rules a designer sets in advance. An agent chooses its next action within its tools and permissions. Orchestration monitors the whole chain (UiPath, current documentation).
- UiPath made its coordination layer Maestro generally available on April 30, 2025. On September 29, 2026, it added patterns for human review and approval chains (UiPath, 2025; UiPath, 2026).
- Not every screen change breaks a bot. Power Automate tries the next recognition rule when one fails. Since version 2.66, AI self-healing is in preview (Microsoft, 2025; Microsoft, 2026).
- An unattended Power Automate bot costs €130 per month, €1.560 per year, excluding building and management. One bot runs one flow at a time (Microsoft, pricing).
- In a comparative study, the RPA bot succeeded 10 out of 10 times on an invoice task, the agent 6 out of 10. Building took about 240 minutes for the bot and 15 for the agent (Průcha et al., 2025).
- Public sources contain no reliable, recent, representative percentage for RPA maintenance. The frequently cited 15-25% annually is an unsampled vendor estimate (EasyData).
- The deciding measure is how much unexpected work returns to your team per hundred correctly completed cases.
What is RPA, and what do UiPath, Automation Anywhere, and Power Automate Desktop do?
RPA imitates employee actions: reading fields, clicking buttons, and retyping data between software packages, following a prebuilt plan. UiPath, Automation Anywhere, and Microsoft Power Automate Desktop are the best-known makers. IBM distinguishes attended RPA, assisting an employee during work, from unattended RPA, starting independently (IBM, The Types of RPA).
You can dismiss two misconceptions immediately. First, RPA does more than blindly click. IBM describes AI document recognition within RPA while stating “RPA does not replace API integration” (IBM). A direct integration remains a separate, often better choice. Second, RPA does have process management. UiPath Orchestrator maintains queues with each transaction’s status, processing robot, exceptions, and optionally a unique reference preventing duplicate processing (UiPath, Queues and Transactions).
What characterizes RPA is that the bot decides nothing that has not been specified beforehand. An invoice without a project number triggers a predefined search rule or goes into an investigation queue for an employee. This is its strength for stable work and its limit for variable work. When missing data can be found in email, a file, or an earlier agreement, you need something that can interpret.
What does an AI orchestration platform do differently from RPA?
An AI orchestration platform coordinates: it distributes subtasks among agents, bots, and people, tracks each case, and monitors when approval is required. UiPath describes it as: “Agentic automation is the orchestration of agents, robots, and humans in a unified automation ecosystem” (UiPath, About agents). Its documentation describes robots as deterministic (same input, same action) and agents as probabilistic (they choose). General explanation of this layer is in AI orchestration versus ordinary automation.
Separate three choices marketing mixes together: how software accesses a package, who chooses the next step, and who monitors the whole task. An agent can also operate screens. Microsoft Copilot Studio lets an agent control Windows apps and websites with a virtual mouse and keyboard, without an API (Microsoft, computer use). The Belgian comparison’s claim “An agent is not a screen robot” (translated) is therefore incorrect (Codelines, 29-06-2026). The difference lies in who decides, rather than how clicks happen.
| Purchasing question | RPA execution | Agent execution | Coordination across the chain |
|---|---|---|---|
| Who chooses the action? | Designer defines rules and routes. | Model chooses within its tools and permissions. | Designer or agent assigns subtasks. |
| How does it access software? | Screen, or available activities and connections. | Direct connection, a bot as a tool, or operating the screen itself. | Connects available routes. |
| What happens with variation? | Prebuilt exception, recovery route, or stop. | Interpretation, then another action or escalation to a person. | Tracks case status and who continues. |
| What do you test in recovery? | Does the bot find the intended field again? | Did the agent still choose the intended action? | Was the whole task resumed correctly without duplicates? |
| Where is management needed? | Screen changes, process rules, machines, transactions. | Also behavior, context, model changes, and permissions. | Handover, waiting, exceptions, final review. |
Our organization based on UiPath, Microsoft, and IBM, rather than measured performance. One product often combines columns.
A fourth column belongs here that almost every comparison forgets: ordinary software or a direct connection without AI. When data is structured and rules fixed, that often beats all three. See AI agents versus traditional workflow automation.
Where does RPA break when screens change?
RPA breaks when a bot no longer recognizes its intended element. This can happen on a screen that looks identical to you. Microsoft says changed element identifiers can make a visually identical screen unusable for a flow (Microsoft, Self-healing, 11-07-2026). Conversely, not every movement breaks it: a bot searches element properties, rather than mouse coordinates. “Whenever a selector fails, Power Automate uses the next selector in the defined order” (Microsoft, 12-05-2025). A moved button does not automatically cause failure, as some comparisons claim (SiRo, 25-05-2026).
Since version 2.66, Microsoft offers generative AI self-healing in preview for certain actions and errors only. It requires administrator settings for screenshots, external models, and sometimes cross-region data traffic. Microsoft warns: “Because self-healing uses generative AI, it might occasionally identify an unintended UI element” (Microsoft, 11-07-2026). A recovered bot may fill the wrong field. A green log check mark does not prove an entry is correct.
Distinguish four changes: a changed screen, data structure, business rule, or missing decision information such as an invoice without a project number. Self-healing addresses at most the first. A customer the bot cannot find often means a missing file, rather than a broken button. Safe agent use in accounting software is covered in Can ChatGPT operate your accounting software?.
What do sources say RPA maintenance costs?
RPA maintenance is demonstrably recurring work, but public sources offer no reliable percentage for small and midsize businesses. The best independent material is qualitative. Digital Public Services Wales interviewed 11 organizations: seven local authorities, three public organizations, and one housing association. Interviews lasted 60 to 90 minutes, with research starting in late 2023 (DPS Wales). The housing association said an external party “changed their approach and changed their website, which broke everything”. It subsequently automated only processes within its own control (DPS Wales, Findings).
The study also reports benefits: less retyping, faster handling, and fewer errors. Regular maintenance mainly undermines the business case for unstable processes. Citing only the failure example tells half the story.
Circulating numbers are weaker than they sound. EasyData says, translated: “Budget 15-25% of initial development costs annually for maintenance” (EasyData), without a sample or calculation basis, and counts new features as maintenance. Assumed development costs of €10.000 imply €1.500 to €2.500 annually, or €4.500 to €7.500 over three years. This calculates an estimate, rather than an invoice. Blueprint attributes $5.000 to $15.000 per bot annually to Deloitte without describing the research (Blueprint). Request four separate quote items: repair, monitoring, adapting to new rules, and expansion.
What do RPA and agents cost per month from their makers?
An unattended Power Automate bot costs €130 per month, a UiPath subscription starts at $25 per month, and a Microsoft screen-operating agent charges credits per step. These amounts cannot be compared because billing units differ: user, bot, or step.
| Offer | Price | Billing unit | Source |
|---|---|---|---|
| Power Automate Premium | €13 per month | Per user, attended | Microsoft, 30-09-2026 |
| Power Automate Process | €130 per month (€1.560 per year) | Per bot, unattended | Microsoft, 30-09-2026 |
| Power Automate Hosted Process | €186,30 per month (€2.235,60 per year) | Per bot, including Microsoft VM | Microsoft, 30-09-2026 |
| UiPath Basic | From $25 per month; Standard and Enterprise on request | Subscription, limited scale | UiPath, 30-09-2026 |
| Copilot Studio computer use | 5 credits per step (standard model), 15 (premium model) | Per step; one step may include multiple clicks | Microsoft, 03-07-2026 |
| Automation Anywhere | No public price found | Unknown | Automation Anywhere, 30-09-2026 |
Microsoft prices use annual billing, excluding VAT. Credits are not euros. All amounts are licenses, excluding development, management, and review.
One detail changes the calculation: one allocated Power Automate bot executes one unattended desktop flow at a time (Microsoft). Agents sending more tasks above it increase the queue rather than capacity. Hosted Process costs an extra €675,60 per bot annually for Microsoft’s machine. See AI orchestration platform pricing for the broader comparison.
Is an AI agent faster or more reliable than an RPA bot?
In the only public comparative study we found, RPA execution was faster and more reliable, while the agent was built much faster. In September 2025, Průcha, Matoušková, and Strnad compared UiPath 2023.4.0 with claude-sonnet-4-20250514 on three tasks (arXiv, 04-09-2025).
| Task | Execution time RPA / agent | Successful runs RPA / agent | Building time RPA / agent |
|---|---|---|---|
| Data entry | 139.8 s / no valid time | 10/10 / 0/1 | About 40 min / not completed |
| Monitoring | 53.9 s / 109.8 s | 10/10 / 9/10 | About 38 min / 10 min |
| Invoice data | 20 s / 202.8 s | 10/10 / 6/10 | About 240 min / 15 min |
Source: tables 1 and 4 of the full study, pages 9-10.
For invoices, the agent needed 10.14 times the execution time; RPA needed 16 times the building time. Interpret cautiously. Building time was measured once, including developer learning time and an error. The reliability difference was not statistically significant. The invoice task covered four invoices. The model is now over a year old; today’s technology would probably change the agent column. The study does prove you must compare building and execution: measuring only one produces opposite winners.
Are UiPath, Automation Anywhere, and Microsoft themselves moving to agents?
Yes. All three major RPA makers now sell bots, agents, and a coordination layer together. None asks you to discard existing bots. This sequence became visible in eighteen months.
| Date | Maker | What happened | Source |
|---|---|---|---|
| 30-04-2025 | UiPath | “UiPath Maestro™ is now generally available”: coordination of agents, people, and robots | UiPath release notes |
| 03-07-2026 | Microsoft | Copilot Studio agents operate Windows apps and websites themselves, even without APIs | Microsoft Learn |
| 11-07-2026 | Microsoft | Generative AI self-healing for desktop flows, in preview | Microsoft Learn |
| 14-09-2026 | Automation Anywhere | Mozart demonstrated as process design above existing tasks | Automation Anywhere Community |
| 29-09-2026 | UiPath | Maestro patterns library: AI decisions with human review, approval chains, and error handling | UiPath release notes |
Automation Anywhere puts it most clearly: “Seamlessly coordinate agents, bots, APIs, and documents from one unified, governed place to power complex workflows” (Automation Anywhere). In the Mozart demonstration, speakers advise retaining stable work and considering redesign mainly with frequent changes or an already major overhaul (Automation Anywhere, 14-09-2026).
This is a sequence of product releases, rather than measured savings. User reports are in AI orchestration platform reviews. A release does not prove universal customer success. The final step’s direction is striking: UiPath adds approval, rather than more independence. The longest-established automation makers now invest where a person approves.
What does the same administrative task look like with RPA, agents, or both?
For the same task, approval boundary, and desired outcome, the main difference is what happens when information is missing. Consider an installation company receiving a supplier email with an invoice. Its project number is missing; the file identifies the project. After approval, exactly one draft entry must exist in the accounting software. This is a worked example, rather than a customer case.
| Step | RPA with fixed rules | Agents with direct connections | Agents and RPA combined |
|---|---|---|---|
| Read email and attachment | Fixed recognition or connected document recognition | Interpret and structure data | Agent interprets; fixed checks remain |
| Missing project | Predefined search rule, otherwise employee investigation queue | Retrieve candidate projects and support the choice; escalate with insufficient evidence | Same interpretation above existing bots |
| Approval | Build a separate human step | Enforce approval outside model free text | One approval for the specific proposed change |
| Enter invoice | Bot executes checked actions | Authorized action with fixed validation | Existing bot enters where a connection is missing |
| Resume after outage | First check whether entry exists | Same: retry must not create a second entry | Same case number across the chain |
Our design based on UiPath, IBM, Automation Anywhere, and unique references in UiPath Orchestrator.
Notice “Approval”. Whether a bot is attended or unattended says nothing about which entry a person must approve. Every technology needs a separately designed approval step. Most demonstrations omit “Resume after outage”, where duplicate payments originate. See errors and review for responsibility and error handling.
Should you replace working RPA bots with agents?
No. Keep a bot that handles stable, repetitive work correctly. Agents belong before and between that work, where interpretation is needed. Makers advise this themselves (Automation Anywhere, 14-09-2026). System count is no argument for AI. Work across three packages that is identical every time suits RPA, or preferably a direct connection (IBM).
Maintenance shifts rather than disappears with agents. Alongside screens and connections, you maintain decision behavior, permissions, context, and outcome checks. Fewer selector repairs do not prove lower overall management. Accessible building also creates more processes someone must own. More employee-built automations make ownership and shared rules more important.
Compare three routes on the same task: extension only (interpretation above existing bots), full replacement, and a combination. Include ordinary software without AI as a fourth. Platform selection is in choosing an AI orchestration platform. Preproduction comparisons are covered in testing an AI agent before automating.
Which measure decides between RPA and AI orchestration?
Our position: for a small or midsize business, the deciding factor is how much unexpected work returns to your team per hundred correctly completed cases. Compare the facts. RPA reliably succeeds on fixed tasks but takes substantial building time and sends exceptions to an investigation queue (Průcha et al.; UiPath). An agent handles more cases but sometimes chooses incorrectly; self-healing can find the wrong field (Microsoft). Maintenance percentages lack traceable evidence. A small team’s scarce resource is that recurring human work. There is no RPA department to catch exceptions.
Calculate: cost per correct case = (monthly development + licenses + machines + AI usage + management hours + review and exception hours + recovery) divided by correctly completed cases. Count independent completions separately from completions after human handling. Otherwise a system forwarding work scores as well as one handling it.
An assumed example: option A costs €200 more per month than B but saves six human hours at €50. Net benefit: €100. If it saves two hours, the net disadvantage is €100. Break-even is four hours per month. Also measure correctly stopped cases, incorrectly stopped cases, silent errors, and duplicate actions. Correctly stopping without approval is a good outcome, rather than failure.
Where is this heading?
RPA and AI orchestration are converging into one offer. Competition shifts toward who best arranges approval and recovery. Measured pace: over eighteen months, UiPath, Microsoft, and Automation Anywhere moved from standalone bots to coordination of bots, agents, and people. The latest step was Maestro approval chains on September 29, 2026 (UiPath). Meanwhile, the best public performance measurement still uses a May 2025 model (Průcha et al.). Products run far ahead of evidence. See AI agent trends for businesses.
Our expectation: by the end of 2027, at least two of the three major providers will correctly handle an interpretation-heavy administrative workflow more cheaply with agents and RPA combined than existing RPA with human exception handling, without more unnoticed errors.
The reasoning: execution is already reliable; makers now add interpretation and approval. For exception-heavy work, the largest cost is people clearing the investigation queue. This expectation fails if fewer than two providers demonstrate lower cost per correct case with equal or better error outcomes in a fair trial by late 2027. It also fails if necessary controls exist only in enterprise bundles unaffordable for a company with thirty people. We will reassess in September 2027.
By then, Bombos will arrange that measurement on your files: same task, same approval, and a visible count per hundred cases of independent successes and work returned to your team.
What can this not yet do?
No source currently proves agents are cheaper than RPA for Dutch small and midsize businesses. Available evidence covers mechanisms, prices, and product development. The comparison uses three tasks, a model over a year old, and one-time building measurements (Průcha et al.). Independent maintenance research is qualitative and Welsh (DPS Wales). Maintenance percentages come from sellers (EasyData).
Agents sometimes choose incorrectly, including when operating screens. Microsoft’s AI self-healing remains in preview (Microsoft). The faster-built agent executes more slowly in the study: 202.8 versus 20 seconds for invoices. That matters for thousands of identical daily transactions.
A predictable bot can predictably write the wrong value when rules or source data are wrong. “Without errors” (translated), as an RPA comparison claims (VionA, 23-12-2024), does not exist. Neither bots nor agents find knowledge only in someone’s head. It must first be recorded. Bombos has not published its own controlled trial against RPA either. We are setting it up as described above.
How does Bombos approach this?
Bombos keeps working bots and handles the surrounding work: reading incoming material, finding the file, and preparing a proposal you approve in Bombos. The goal is more work, at a higher quality, with the same team.
Bombos uses multiple agents. Chef distributes work: it sees incoming material, recognizes the task, and assigns the right specialist. Wegwijzer guides you and helps set boundaries. Specialists are built around your tasks in your systems. After approval, Bombos reads and writes in packages such as Exact, Dutch accounting software, AFAS, Dutch business software, Twinfield and SnelStart, accounting packages, and Syntess, software for installation companies (your own software).
Each proposal shows its basis. You approve, change, or reject in Bombos; the package then contains only the result. A correction becomes a rule, including for coworkers. Payments, customer messages, and contracts wait for your approval by default. This boundary is technically enforced. Bombos can remove it at your request, at your own risk. If knowledge exists only in someone’s head, Bombos interviews you to record it.
One task is the beginning. Your own team then teaches Bombos the next without needing technical skills. Read what an agent is and how long it works independently in What is an AI agent?.
Sources
Each source was opened on September 30, 2026, and each original quotation appears verbatim in it. Dutch quotations above are labeled as translations.
- UiPath, Agents: About agents, current documentation, no date listed.
- UiPath, Maestro: April 2025 release notes, 30-04-2025.
- UiPath, Maestro: September 2026 release notes, 29-09-2026.
- UiPath, Orchestrator: About Queues and Transactions, current documentation, no date listed.
- UiPath, Plans and Pricing, pricing page, accessed 30-09-2026.
- Automation Anywhere, Agentic Process Automation System, product page, no date listed.
- Automation Anywhere Community, September 2026 Product Club Recap: Adapt Anytime with Mozart Orchestrator, 14-09-2026.
- Microsoft Learn, Self-healing (preview), 11-07-2026.
- Microsoft Learn, Build a custom selector, 12-05-2025.
- Microsoft Learn, Automate web and desktop apps with computer use, 03-07-2026.
- Microsoft, Power Automate pricing, pricing page, accessed 30-09-2026.
- IBM, The Types of RPA, no date listed.
- Průcha, Matoušková, and Strnad, Are LLM Agents the New RPA? A Comparative Study with RPA Across Enterprise Workflows, preprint, 04-09-2025.
- Digital Public Services Wales, Automation and artificial intelligence discovery report and Findings, no date listed, research started in late 2023.
- EasyData, RPA Costs: The Complete Guide for 2026, no date listed; sells automation.
- Blueprint, The Robotic Process Automation ROI in Value Mapping, no date listed; sells automation analysis.
- Codelines, AI agent vs RPA: what fits your small or midsize business?, 29-06-2026.
- SiRo, RPA vs AI: Which Automation Fits Your Business?, 25-05-2026.
- VionA, Difference Between AI Agents and RPA, 23-12-2024.