RPA 2026 Deep Dive - UiPath vs Automation Anywhere vs Blue Prism vs Power Automate Desktop vs WinActor vs Brity Works and the Agentic AI Invasion

Prologue — Spring 2026, the RPA Crossroads

Monday morning in May 2026, Yeouido, Seoul. A digital-innovation team at a major commercial bank stares at two numbers on the whiteboard: "1,847 bots in production" and "12 agentic-AI pilots." The team head points and says: "By year end, we sunset half the bots and put LLM agents in their place."

Same moment, Marunouchi, Tokyo. Mitsubishi UFJ Bank's IT department runs 3,200 WinActor scenarios every night. The department head sighs: "WinActor maintenance keeps going up every year. But we cannot turn them off. The minute we do, humans have to do that work again."

New York, UiPath HQ. CEO Daniel Dines has been back over a year. The stock hovers around a quarter of the IPO price (56 USD), and the board has banned the word "RPA" from public messaging. The official tagline is now "The Agentic Automation Platform."

These three scenes are the 2026 RPA crossroads. Nobody can switch the bots off, and nobody wants to buy new RPA. Meanwhile Microsoft Power Automate Desktop, free on Windows, has eaten half the SMB market, and OpenAI Operator plus Anthropic Computer Use are redefining what "screen automation" even means. This article looks the landscape straight in the eye.

1. The 2026 RPA Market Map — Four Camps

The 2026 RPA market sorts into four camps on one page.

┌──────────────────────────────────────────────────────────────────────┐
│  Tier 1 · Global RPA Big 3 (top in share and mindshare)               │
│    UiPath · Automation Anywhere · Blue Prism (SS&C)                   │
│    Enterprise, attended + unattended, own cloud + on-prem             │
├──────────────────────────────────────────────────────────────────────┤
│  Tier 2 · Hyperscalers + adjacent platforms                           │
│    Microsoft Power Automate Desktop · IBM RPA · Pega · Workato        │
│    Tied to M365/CRM/iPaaS, fastest share gain                         │
├──────────────────────────────────────────────────────────────────────┤
│  Tier 3 · Specialty / regional champions                              │
│    NICE · Kofax · Nintex · Salesforce MuleSoft RPA                    │
│    KR: Brity Works (Samsung SDS) · Aircuvate (LG CNS) · AutoEver      │
│    JP: WinActor (NTT) · BizteX cobit · RoboPat DX · Fujitsu RPA       │
├──────────────────────────────────────────────────────────────────────┤
│  Tier 4 · Agentic-AI new entrants                                     │
│    UiPath Agent Builder · AA Agent Studio · Adept · MultiOn           │
│    Anthropic Computer Use · OpenAI Operator                           │
│    "No selectors anymore — the model looks at the screen"             │
└──────────────────────────────────────────────────────────────────────┘

The pattern: Tier 1 has the biggest revenue but the lowest growth, Tier 4 has small revenue but the highest growth, and Tier 2 is taking share from both. The moment Microsoft bundled Power Automate Desktop into Windows 11 and made "free RPA" a default option, the UiPath and Automation Anywhere SMB channel basically collapsed.

This article dissects each camp. Where each is strong, where weak, and which stack to choose in 2026.

2. UiPath — Repositioning as the "Agentic Automation Platform"

UiPath started in Bucharest, Romania, and listed on NYSE in 2021 as the canonical RPA brand. IPO price 56 USD, peak market cap around 35B USD. In May 2026 the stock sits near 12 USD with a market cap of 6 to 7B USD. Down to a quarter.

Three reasons: (1) RPA market growth stalled, (2) Microsoft entered free, (3) LLM agents emerged as a new threat. Founder Daniel Dines returned as CEO in 2024, and through 2025-2026 the company has been repositioning around "Agentic Automation Platform."

The UiPath stack:

• Studio / Studio X / StudioWeb — Automation designers. Studio targets developers with XAML, StudioX targets non-developers, StudioWeb is the cloud-native browser build.
• Robot (Attended / Unattended) — Bots that actually run.
• Orchestrator — Bot management, scheduling, queues, logs, assets.
• AI Center — ML model training, deployment, and use.
• Document Understanding — Document OCR, LLM extraction, classification, and verification.
• Process Mining / Task Mining — Analytics that find what to automate.
• Action Center — Human-approval steps.
• Agent Builder (2025-) — LLM agent builder that wraps existing RPA workflows as tools.

UiPath automations are stored as XAML. Below is a minimal XAML snippet that reads an Excel file and types values into SAP:

<Activity x:Class="ReadExcelToSAP" xmlns="http://schemas.microsoft.com/netfx/2009/xaml/activities">
  <Sequence DisplayName="ReadExcelToSAP">
    <ui:ExcelApplicationCard FilePath="C:\Input\orders.xlsx" Visible="True">
      <ui:ReadRange SheetName="Sheet1" Range="A2:E100" Result="[dtOrders]" />
    </ui:ExcelApplicationCard>
    <ForEach TypeArgument="DataRow" Values="[dtOrders.AsEnumerable()]">
      <Sequence>
        <ui:SAPLogon Server="prod-sap.example.com" Client="100" User="bot_user" />
        <ui:TypeInto Selector="&lt;wnd app='saplogon.exe' cls='Shell_TabWindowClass'/&gt;" Text="[row(0).ToString()]" />
        <ui:Click Selector="&lt;wnd ctrlname='btnSave'/&gt;" />
      </Sequence>
    </ForEach>
  </Sequence>
</Activity>

This XAML exposes the essence of RPA — point at a UI element with a selector (wnd app='saplogon.exe' cls='Shell_TabWindowClass'), then layer clicks and typing on top. And the moment a selector breaks, the bot stops. AI Computer Vision selectors (since 2024) and the Agent mode (since 2025) are precisely UiPath's attempt to solve this.

Pricing: in 2026 a Studio license runs 1,500 to 2,000 USD per user per year; an attended Robot is 1,200 to 1,500 USD/year, an unattended Robot is 6,000 to 8,000 USD/year; Orchestrator starts around 25k USD/year, varying by cloud vs on-prem. Document Understanding is billed per page. A typical mid-market first-year deal lands at 150k to 500k USD.

3. Automation Anywhere — Cloud-Native + Agent Studio

Automation Anywhere (AA) launched in San Jose in 1996, raised 850M USD in 2018 from Goldman Sachs and SoftBank, and shelved a 2021 IPO when the market turned. Still private, took new funding in 2024 Series G.

AA's differentiators:

1. A2019 / Automation 360 — First in the Big 3 to fully move the IDE to the browser. UiPath Studio remains a .NET desktop app; AA is web-only. Big advantage for governance and deployment.
2. IQ Bot → Document Automation — Native IDP, ahead of UiPath in building IDP into the core platform.
3. Bot Store — The largest prebuilt-bot marketplace, ahead of UiPath Marketplace.
4. Agent Studio (2025-) — AA's agentic-AI builder, exposing bots as tools and letting LLMs call them.

Pricing matches or undercuts UiPath slightly. AA is often called stronger on attended bots — the call-center-agent-PC scenario.

4. Blue Prism (SS&C) — The "Original RPA" Aging

Blue Prism started in 2001 in the UK and coined the term "Robotic Process Automation" in 2008 marketing. Listed on LSE in 2017, acquired by SS&C Technologies in 2022 for 1.6B USD and delisted.

Blue Prism's philosophy from day one was 100% unattended, developer-strong, governance-first. Unlike UiPath, Blue Prism never championed citizen developers — it carried the British enterprise belief that "RPA must be controlled by IT."

The Object Studio model is the core idea. UI elements of an application are defined once in an object, and every process reuses that object. Compare this with UiPath, where selectors live inside each activity:

<object name="SAP_Customer_Master">
  <element name="CustomerIdField">
    <type>Edit</type>
    <attributes>
      <attribute name="WindowText" value="*"/>
      <attribute name="ClassName" value="TextEdit" matchType="Equal"/>
      <attribute name="CtrlId" value="100" matchType="Equal"/>
    </attributes>
  </element>
  <action name="ReadCustomerId">
    <parameter direction="out" name="Value" type="text"/>
    <stage name="ReadFromCustomerIdField" type="Read"/>
  </action>
</object>

Since SS&C acquired Blue Prism, more of its sales come bundled with SS&C's asset-management, fund-administration, and insurance practices. New Blue Prism deals as a standalone RPA choice are rare in 2026, but existing financial-services deployments keep humming.

5. Microsoft Power Automate Desktop — RPA Bundled into Windows

Microsoft acquired Softomotive (WinAutomation, ProcessRobot) in 2020. In 2021 Microsoft announced Power Automate Desktop, and from April 2021 every Windows 10/11 user gets Power Automate Desktop for free. The biggest shock to the RPA market.

What is free and what is paid:

• Free: local Power Automate Desktop (attended, tied to the Windows-logged-in user).
• Paid (Per User Plan): cloud triggers, unattended runs, CoE Toolkit, AI Builder.
• Per User with Attended RPA: about USD 40/user/month.
• Per Flow Plan: about USD 100/flow/month.
• Process Plan (Unattended): about USD 150/bot/month.

UiPath unattended is 6 to 8k USD/year. Microsoft sells unattended at 1,800 USD/year. In SMB the game is over.

Power Automate Desktop serializes actions to JSON. Below is part of a "read PDFs in a folder, append rows to Excel" scenario:

{
  "actions": [
    {
      "name": "GetFilesInFolder",
      "module": "Folder",
      "parameters": {
        "folder": "C:\\Invoices",
        "fileFilter": "*.pdf",
        "includeSubfolders": false
      },
      "output": "Files"
    },
    {
      "name": "ForEach",
      "iterator": "CurrentFile",
      "collection": "%Files%",
      "body": [
        {
          "name": "ExtractTextFromPDF",
          "module": "PDF",
          "parameters": { "pdfFile": "%CurrentFile.FullName%" },
          "output": "ExtractedText"
        },
        {
          "name": "AIBuilderExtractFields",
          "module": "AIBuilder",
          "parameters": {
            "modelId": "invoice-extractor-v3",
            "text": "%ExtractedText%"
          },
          "output": "Fields"
        },
        {
          "name": "ExcelLaunchAndOpen",
          "parameters": { "path": "C:\\Reports\\invoices.xlsx" }
        },
        {
          "name": "ExcelWriteCell",
          "parameters": {
            "row": "%LastRow + 1%",
            "column": 1,
            "value": "%Fields.InvoiceNumber%"
          }
        }
      ]
    }
  ]
}

Power Automate Desktop's weakness is still attended-first. Unattended needs Process Plan (150 USD/bot/month) plus separate machine licensing, after which the price approaches UiPath Cloud or Automation Anywhere. Still, for any SMB inside the M365 estate, this is the first pick.

6. IBM Robotic Process Automation — WDG Acquisition, Quiet Position

IBM entered RPA in 2020 by acquiring Brazil-based WDG Soluções em Sistemas. Before that it was IBM RPA with Automation Anywhere (AA OEM); post-WDG IBM moved to its own product. The 2026 official name is IBM Robotic Process Automation.

IBM RPA's strength: it bundles cleanly into IBM Cloud Pak for Business Automation alongside IBM BPM, IBM Operational Decision Manager, IBM Datacap (document capture), and IBM Process Mining. Since WDG days, process mining has shipped together.

Adoption sits inside IBM's existing customer base — banks, governments, telcos. Choosing IBM RPA as a fresh RPA pick is rare in 2026. But for organizations already on IBM CP4BA, it's a natural fit.

7. Workato · Tray.io — Where iPaaS Meets RPA

Workato launched in 2013 and raised a 2021 Series E at 5.7B USD. Core identity is iPaaS (Integration Platform as a Service), but since 2022 Workato sells "Workato RPA" alongside, with the message: "API where you can, RPA where you cannot."

Why this matters: in 2026 about 80% of automation targets in any enterprise have APIs — Salesforce, NetSuite, Workday, Jira, Slack. RPA is the most inefficient way to automate those. An API call is 100 times faster and 100 times more stable than a selector click.

The Workato/Tray.io/Boomi line is: "Do not deploy bots where APIs exist. Use APIs. Save RPA for true API-less surfaces — SAP GUI, legacy mainframe screens." That message is the most painful one for UiPath and AA.

8. NICE · Pega · Kofax · Nintex — RPA in Adjacent Domains

NICE — Israel. Call center and financial trading is the core business, with RPA (Real-Time, Advanced Process Automation) bundled. Strong on attended bots. Coupled with NICE CXone for call-center agent-PC automation.

Pega Systems — US. BPM is the core, Pega Robotic Process Automation (acquired from OpenSpan) is adjacent. Sold into insurance and bank core systems.

Kofax (Tungsten Automation) — Renamed Tungsten Automation in 2024. Core is IDP (Intelligent Document Processing) alongside ABBYY, but Kofax RPA ships too. Strong in document-heavy industries (insurance, banking, government) with the IDP + RPA combination.

Nintex — Australia. Started in SharePoint/M365 workflow automation, extended to RPA (Nintex RPA, formerly EnableSoft Foxtrot). Mid-market with M365 affinity.

All four are companies where RPA is not the core business. They sell it alongside their main practice (contact center, BPM, IDP, M365 workflow).

9. Korea · Samsung SDS Brity Works · LG CNS Aircuvate · Hyundai AutoEver

The Korean RPA market took a different path. The global Big 3 (UiPath, AA, Blue Prism) entered Korea around 2017-2020, and at the same time the local SI Big 4 (Samsung SDS, LG CNS, SK C&C, Hyundai AutoEver) each built their own RPA stacks.

Samsung SDS Brity Works (formerly Brity RPA) — Samsung group plus external customers. Brity Works bundles RPA, ChatBot, and AI into one workplace brand. The Brity Works RPA engine builds bots from screen recordings. Group standard tool.

LG CNS Aircuvate — LG group plus external. Mix of in-house and UiPath OEM. Aircuvate Studio is the designer.

Hyundai AutoEver (HMG ITS) RPA — Hyundai Motor Group standard. Mix of UiPath and proprietary.

Shinhan, KB, Woori, Hana Bank RPA — All four big banks operate hundreds to thousands of bots. Shinhan on UiPath, KB on UiPath plus proprietary, Woori historically heavy on Blue Prism but migrating to UiPath, Hana on UiPath. Internet-only banks (KakaoBank, Toss Bank, K Bank) are API-first and lean less on RPA.

Typical workloads: loan application data entry, call-center wrap-up, foreign-currency SWIFT input, period close, KYC/AML screen scraping, debt-collection notice generation. Anything a human does in front of a PC clicking screens.

Korean specifics: HR/accounting is split between SAP and Korean ERPs (Douzon Bizon, Younglimwon). The latter has weaker API coverage, so RPA dependency is higher than the global average.

10. Japan · NTT WinActor · Fujitsu RPA Express · BizteX cobit · RoboPat DX

Japan's RPA market goes deeper than Korea's. Japanese enterprises' heavy IT dependency in back office, paper/fax culture, and the EUC tradition combined to drive an RPA explosion.

WinActor (NTT-AT) — Dominant number one in domestic Japanese RPA. Built by NTT Advanced Technology, sold and deployed by NTT Data. Japanese UI, Japan-friendly, deployed via Fujitsu/Hitachi/NEC as system integrators. Close to a large-enterprise standard.

BizteX cobit — First cloud RPA in Japan. Browser-based. Targets SMB and mid-market. SaaS pricing.

RPA RoboPat DX (FCE Process & IT) — Brand promises "RPA built directly by business users." Praised for friendly Japanese UI. Spreads through mid-sized companies and regional Japanese businesses.

Fujitsu RPA Express — Fujitsu's own RPA (Fujitsu also resells WinActor). Focused on existing Fujitsu customers.

NEC Software Robot Solution — NEC's own.

UiPath Japan — Standalone entity in Japan. Toyota, Nippon Life, SoftBank, and other large enterprises use UiPath.

Automation Anywhere Japan / Blue Prism Japan — Low share.

Reference cases:

• Mitsubishi UFJ Bank — Mixed WinActor and UiPath. Back office payment processing, FX screens, loan-underwriting screens. Reported thousands of automation scenarios.
• Sompo Japan — Claims handling screens, policy lookup RPA.
• Panasonic — Finance and HR back office.
• Mitsui Sumitomo Insurance — Multiple internal-award winning RPA cases.
• ANA / JAL — Ticketing post-processing, mileage settlement.

Japan specifics: Japanese enterprise screen designs are more conservative and stable than Korean or US peers (legacy screens go a decade unchanged). That means RPA selectors break less, and RPA has a longer lifespan. It also weakens the motive to migrate to LLM agents.

11. Attended vs Unattended — The Two Bot Modes

The first split in RPA is attended vs unattended.

• Attended bot: runs on a human's PC, triggered by the human. Example: while a call-center agent looks up a customer, the bot fills the ERP screen alongside. It shares the user's Windows session.
• Unattended bot: runs without a human on a dedicated VM/machine on a schedule. Example: at 3am pull yesterday's sales from SAP into the BI system. Uses a dedicated bot Windows login.

Unattended pricing is 3 to 5 times attended (UiPath: attended 1,500 USD/year, unattended 6,000 to 8,000 USD/year). Because (1) unattended bots do more, (2) governance cost is higher when humans are not watching, (3) vendor revenue concentrates here.

Roughly 80% of enterprise RPA is unattended, 20% attended. NICE leads in the attended-heavy contact-center segment.

12. The Selector Tragedy — Why RPA Breaks So Often

The most common industry complaint about RPA is "bots break too often." The cause is selectors.

A selector is a string pointing at a UI element. Example: <wnd app='chrome.exe' cls='Chrome_WidgetWin_1'/><html name='SAP'/><input id='customer_id'/>. When what the selector points at moves or changes attributes, the bot stops.

Common breakage causes:

1. UI updates to SAP/Salesforce/Workday (quarterly).
2. Windows dialog position changes, font changes, DPI changes.
3. Resolution changes or monitor-count changes on the bot host.
4. Exception screens at login — ad banners, security warnings, password-expiry prompts.
5. Screen lock from concurrent logins.

Because of this, RPA ops teams routinely build "bots that reduce headcount" and then hire "people to fix the bots." Internal RPA CoEs of 5 to 50 people are common in large enterprises.

LLM-based computer-vision selectors (UiPath AI Computer Vision, Microsoft Power Automate Vision, Anthropic Computer Use) are the direction of the fix. Instead of selectors, you tell the agent "find the customer ID input" in natural language. Accuracy hit the mid-95% range by mid-2026 and is fast becoming the new default.

13. OCR + Document Understanding (IDP) — Document Understanding · Hyperscience · Rossum · ABBYY · Tegaki

A core RPA weakness is the step where a human has to read a PDF/image before typing. IDP (Intelligent Document Processing) attaches to RPA to solve this.

Main IDP players:

• UiPath Document Understanding — Built into UiPath. Form recognizer + Layout Extractor + Intelligent Form Extractor + LLM extractor.
• Hyperscience — Pure-play IDP. Most common in insurance and banking. Strong human-in-the-loop verification.
• Rossum — Czech origin. Specialized in invoice IDP. Deep SAP integration.
• ABBYY FlexiCapture / Vantage — The oldest OCR company. Its OCR engine is an industry baseline.
• Tegaki (Cogent Labs) — Japanese-origin handwriting OCR. De facto standard for Japanese paper-document processing.
• Kofax (Tungsten) IDP — Co-leader with ABBYY.
• Microsoft AI Builder / Azure Document Intelligence — Microsoft camp.
• Google Document AI — Google Cloud camp.
• AWS Textract — AWS camp.

The biggest change in 2024-2026 is that LLM-based extraction became the default. In the old world you had to define "invoice_number is in the second line top-left of this template" with coordinate-based templates. Today you hand the LLM a PDF and ask "extract the fields as JSON." UiPath is migrating Document Understanding to LLM mode rapidly.

14. Process Mining · Task Mining — Finding What to Automate

Another RPA weakness is that picking the highest-ROI processes to automate is guesswork by humans. Process Mining and Task Mining solve this.

Process Mining — Reconstruct actual process flow from ERP/CRM/MES logs. Visualize the gap between the standard process and the real "variants" to expose inefficiency. Key tools:

• Celonis — Germany. The clear leader. Valued near 13B USD (2024 Series D round). Standard at Siemens, Vodafone, Samsung Electronics, Bosch, and more.
• ABBYY Timeline — Acquired by ABBYY.
• Microsoft Process Mining (formerly Minit acquisition) — Power Automate Process Advisor + Process Mining.
• UiPath Process Mining — Acquired by UiPath (formerly ProcessGold).
• Apromore (open source) — Australia.
• Pafnow / IBM Process Mining — IBM.
• SAP Signavio — SAP acquisition, deep ERP integration.

Task Mining — Record user PC screens and analyze work patterns. Where process mining is system-log-based, task mining is user-behavior-based. UiPath Task Mining, Microsoft Task Mining, Celonis Task Mining all exist.

These two have become the standard pre-RPA-investment toolset. "Which bots to build" is no longer a guess — it's a data answer.

15. Citizen Developers + RPA CoE

RPA governance splits into two models.

Citizen-developer model (UiPath/Microsoft preferred) — Business users build their own bots. Designer must be low-code/no-code. Fast ROI, departmental autonomy. The downside: bots explode and governance breaks.

RPA CoE model (Blue Prism/enterprise preferred) — A dedicated internal RPA team builds and operates every bot. Strong governance/quality/security. The downside: bottleneck and slow ROI.

What actually works in 2026 is hybrid — simple bots from citizen developers, complex or mission-critical bots from the CoE, with the CoE owning governance (security review, secret management, bot ID management, monitoring) for the citizen-built bots too.

16. Bot Identities · Secrets · Governance

A core RPA governance problem is bot identity. A bot needs an ID and password to log into SAP, and who manages those credentials is a security question.

Bad patterns (often seen):

• The bot uses a "shared ID" known by several humans.
• Passwords sit in plaintext inside the bot script.
• The bot ID expires every 90 days, and every rotation breaks every bot.

Good patterns:

• Each bot gets a dedicated ID (for example bot_finance_close_01), separate from humans.
• Passwords live in CyberArk / HashiCorp Vault / UiPath Credentials / Power Automate Connection References.
• Bot ID expiry policy differs from human ID (long-lived with strong rotation discipline).
• Every bot run logs into a SIEM (Splunk / Sentinel) for audit.

Most RPA security incidents in finance trace back to bot-ID mismanagement. A 2023 Korean bank incident saw over a thousand attempted fraudulent transactions via a hijacked bot ID (per local press).

17. Cost Economics — License + Infra + Maintenance

The real cost of RPA is not license. It's maintenance. A typical enterprise RPA cost breakdown:

• License: 30 to 40%
• VM/machine infrastructure: 10 to 15%
• Implementation (SI fees): 20 to 30%
• Maintenance (bot repair, selector updates, new-screen adaptation): 30 to 40%

Maintenance cost rivaling license cost is what eats RPA ROI. You replaced 100 people with 100 bots, then hired 20 RPA developers to maintain the bots. Those 20 cost more per head than the original office workers.

This cost shape is the economic basis for the "RPA is dead" narrative. If 20 people are going on payroll anyway, the math now says hire 20 data engineers / AI engineers to build API integrations and LLM agents instead.

18. The Agentic-AI Invasion — UiPath Agent Builder · AA Agent Studio · Power Automate AI Agents

In 2024-2026 the RPA Big 3 each declared an "agentic AI" pivot.

UiPath Agent Builder (2025-) — Exposes existing UiPath workflows as tools and lets an LLM call them. Effectively LangChain/LangGraph layered over RPA tools. Accepts natural-language tasks like "the customer emailed, reply, log to the system, notify the manager."

Automation Anywhere Agent Studio (2024-) — Similar shape. AA bundles its bots, external APIs, and LLMs into one agent runtime.

Microsoft Power Automate AI Builder Agents + Copilot Studio — Microsoft is merging Power Automate and Copilot Studio. Natural-language agents in M365 Copilot trigger Power Automate flows.

Anthropic Computer Use / OpenAI Operator — Threats from outside the Big 3. The model itself looks at the screen and drives mouse and keyboard. No selectors required. Reliability lags pure RPA in May 2026, but is catching up fast.

Who wins?

• Simple repetitive desktop work → Computer Use / Operator win progressively.
• Mission-critical + governance-heavy → UiPath / AA / Blue Prism hold ground for a while.
• Automation inside M365 → Power Automate AI Agents win.
• Automation of API-rich SaaS → Workato/Zapier win.

19. Where the "RPA Is Dead" Narrative Is Right (and Wrong)

A precise look at the industry slogan "RPA is dead, long live agents."

The right parts:

• New RPA adoption has slowed. UiPath NRR (Net Retention Rate) fell from 121% in 2023 to 105% in 2025.
• LLM agents are more flexible. No selector-breakage problem.
• Telling Claude/ChatGPT what to do beats training citizen developers for many use cases.

The wrong parts:

• Thousands of bots already in production. Nobody is turning them off.
• Mission-critical back office cannot accept LLM non-determinism. "Bookkeeping correct 99% of the time" does not fly.
• Governance/audit/regulator demands are far more mature on the RPA side.

The accurate framing: "RPA growth ended. RPA itself is not going away. New automation goes to LLM agents."

20. Vendor Comparison Matrix 2026

21. What to Choose — A Decision Tree

By organization profile:

• Global enterprise + UiPath already deployed → Keep UiPath + pilot Agent Builder.
• Global enterprise + greenfield → Microsoft Power Automate (M365 bundle) or UiPath.
• Mid-market SMB + M365 → Microsoft Power Automate Desktop.
• Mid-market SMB + Google Workspace → Workato or Zapier (downplay RPA).
• Korean finance + conservative → Keep existing UiPath/Blue Prism.
• Call-center centric → NICE.
• Japanese large enterprise + conservative → Keep WinActor.
• Japanese SMB → BizteX cobit / RoboPat DX.
• Korean conglomerate + group standard → Brity Works (Samsung) / Aircuvate (LG) / Hyundai AutoEver (Hyundai).
• "API everywhere" estate → Workato + minimal RPA.
• "Just go straight to LLM agents" → Anthropic Computer Use / OpenAI Operator + gradual RPA sunset.

22. Antipatterns — 11 Ways to Wreck RPA

1. "Lift and shift Excel + macros into RPA" — moving legacy debt into RPA does not pay it down.
2. Plaintext bot passwords.
3. One bot does too much — 50-step scenarios in one workflow.
4. Citizen developers without a CoE — bots explode and governance collapses.
5. CoE without business-user collaboration — wrong use cases get built.
6. Bot ID = human ID — the human leaves and the bot dies.
7. No bot monitoring — nobody notices when the bot is broken.
8. Coordinate-based selectors (x=120, y=340) — instant death on resolution change.
9. RPA detour where APIs exist — 100x slower, 100x more fragile.
10. ROI math that ignores maintenance — projected savings never materialize.
11. Postponing an LLM-agent pilot — catching up two years late is much harder.

23. Field Checklist — RPA CoE 0/30/90 Day Guide

Days 0 to 30:

• Interview to collect 5 to 10 automation candidate processes.
• Score by (repetition x clear rule base x stable systems x no API) and pick the top 3.
• Map step-by-step with Task Mining (or shadow interviews).

Days 30 to 60:

• Implement 2 to 3 pilot bots. Start attended.
• Set the bot-ID policy (dedicated IDs + Vault + SIEM logging).
• Decide citizen-developer vs CoE model.

Days 60 to 90:

• Ship the first unattended bot.
• Stand up the ops dashboard (Orchestrator / PA Cloud / CoE Console).
• Define the bot change-control flow (selector change → impact analysis → redeploy).
• Plan the next-quarter backlog and seed one LLM-agent pilot.

24. Conclusion — How to Treat RPA in 2026

The straight 2026 spring guidance:

1. Keep what is deployed, but stop adding. Bots in production are validated assets. For every new bot, ask first: "could this be an LLM agent instead?"
2. Citizen-developer model is dangerous. Too many teams watched bot sprawl break governance. Stand up a CoE first; give citizen developers strict guardrails.
3. Bot ID and secret management is RPA's softest security spot. Invest there first.
4. Process Mining should have come before RPA. If you missed it, start now. Building bots without knowing what to automate is self-harm.
5. Run one LLM-agent pilot in 2026. The outcome reshapes 2027 strategy.
6. Minimize vendor lock-in. UiPath/AA/Blue Prism all face weak post-IPO numbers. Do not bet everything on one vendor.
7. Price in maintenance from day one. Not "one bot = one fewer human." Reality is "one bot = 0.5 fewer humans plus 0.1 extra RPA operator."

RPA is not dying. But the definition of automation is changing. RPA is one chapter in that change, not the whole book.

References

• UiPath Platform Documentation — Studio, Robot, Orchestrator, Document Understanding, Agent Builder.
• Automation Anywhere Documentation — Automation 360, Document Automation, Agent Studio.
• Blue Prism Documentation (SS&C) — Object Studio, Process Studio, Hub, Interact.
• Microsoft Power Automate Documentation — Power Automate, Power Automate Desktop, AI Builder.
• IBM Robotic Process Automation — IBM RPA (WDG lineage).
• Workato Documentation — iPaaS + RPA integration.
• Celonis Process Mining — Execution Management System.
• ABBYY Vantage / FlexiCapture — IDP / OCR.
• Hyperscience — Pure-play IDP.
• Rossum — Invoice IDP.
• NICE Robotic Automation — Call-center RPA.
• Pega Robotic Process Automation — BPM + RPA.
• Kofax (Tungsten Automation) — IDP + RPA.
• Nintex RPA — M365 workflow.
• WinActor (NTT-AT) — Japan number one.
• BizteX cobit — Japanese cloud RPA.
• RPA RoboPat DX — Japanese front-line RPA.
• Samsung SDS Brity Works — Samsung group RPA standard.
• LG CNS Aircuvate — LG group RPA.
• Anthropic Computer Use — Computer-vision based automation.
• OpenAI Operator — OpenAI autonomous agent.
• Forrester Wave: Robotic Process Automation Q1 2025 — Market analysis.
• Gartner Magic Quadrant for RPA 2024 — Vendor evaluation.
• Everest Group RPA PEAK Matrix 2024 — RPA market analysis.