Is RPA dead? Why it spread, and what AI changes

“RPA is dead.” As AI agents draw attention, you hear this more often. From around 2017, many Japanese companies and local governments adopted RPA. Many robots still run every day, but others have stopped and were never fixed, or were built and then left unused.

This article looks back, with survey numbers, at why RPA spread in Japan and why its growth stalled. It then asks whether RPA really is dead, and sorts out what AI changes and what it doesn’t.

The short answer: RPA’s job isn’t over. What changes is what happens after a robot is built

• By the numbers, it is not dead: Japan’s RPA market reached 90.3 billion yen in fiscal 2023, up 13% from the year before. 98% of prefectures and every designated city have adopted RPA
• The reasons it spread still hold: no changes to existing systems, staff can build robots themselves, results show up as hours saved, and companies can’t hire more people
• The reasons it stalled come after adoption: robots stop when screens change, too few people can fix them, built robots fall out of use, and work that needs judgment stays out of reach
• AI changes how robots are built and fixed: AI reads the screen to operate it, and steps can be fixed in words. But handing everything to AI agents costs both money and time

Ten years of RPA in numbers

When	What happened
2017	Japan’s three megabanks announced large workload cuts one after another: MUFG the workload of 9,500 people, SMFG the workload of 4,000, and Mizuho about 19,000 jobs
2018	Worldwide RPA software revenue grew 63.1% to $846 million, the fastest-growing segment of enterprise software (Gartner)
June 2018	RPA adoption among Japanese companies with annual sales of 5 billion yen or more reached 22% (MM Research Institute)
September 2018	The Ministry of Economy, Trade and Industry’s DX Report warned of the “2025 digital cliff”: existing systems grown complex and opaque
April 2019	Legal caps on overtime took effect for large companies (April 2020 for small and medium-sized ones)
November 2019	Adoption reached 38%, and 51% among large companies with annual sales of 100 billion yen or more (MM Research Institute)
Fiscal 2019	Japan’s RPA market grew 56.7% to 52.97 billion yen (Yano Research Institute)
February 2020	Gartner found RPA in Japan past the peak of inflated expectations and heading for the bottom of the trough of disillusionment
March 2021	Power Automate Desktop became available on Windows 10 at no extra cost
Fiscal 2023	Japan’s RPA market reached 90.3 billion yen, up 13% (Deloitte Tohmatsu MIC Research Institute)
March 2024	Adoption among companies with sales of 5 billion yen or more was 44%, down 1 point. Among smaller companies it was 15% (MM Research Institute)
October 2025	RPA was in use at 98% of prefectures, 100% of designated cities, and 44% of other municipalities (Ministry of Internal Affairs and Communications)

The two research firms define the market differently, so Yano’s and MIC’s figures can’t be compared directly. Even so, the growth rate clearly fell, from 56.7% in fiscal 2019 to 13% in fiscal 2023.

Four reasons RPA spread

1. No changes to existing systems

RPA operates screens the way a person does. Even with old systems that have no API or other way to connect, it moves data between systems without waiting for changes to them. In 2019, a Gartner analyst called the ability to integrate legacy systems the key driver for RPA projects.

Rebuilding a system takes years and a large budget. After the Ministry of Economy, Trade and Industry’s 2018 DX Report warned of the “2025 digital cliff,” few companies could rebuild right away. In IPA’s DX Trends 2025, 62.7% of companies still had legacy systems in at least part of their business. Companies chose RPA to bridge the gap until they could rebuild.

The work local governments list for RPA in the Ministry of Internal Affairs and Communications survey shows the same pattern: entering data from a front-desk system into the core system, entering application forms into business systems, downloading data from banks’ systems. Most of it is copying data between systems.

2. Staff were expected to build robots themselves

With RPA, you build a scenario by recording operations or arranging screen parts. The promise that business staff who don’t program could start automating on their own was what won many companies over.

In practice, though, the IT department usually built the robots. In Gartner’s 2019 survey, 57% of companies had their IT department develop the robots. The most common setup, at 67%, was IT coordinating and standardizing the program across the company, and RPA was moving from departmental use to company-wide rollout. The small number of people who could build robots later became a reason growth stalled.

3. Results show up as hours saved

“So many hours saved per year.” RPA’s effect can be shown as manual work that disappeared. Sumitomo Mitsui Financial Group says RPA cut 3.5 million hours over fiscal 2017 to 2019, the workload of 1,750 people. Among local governments, Asahikawa reported to the ministry that RPA saves 6,460 hours a year across 53 tasks.

Start with a small task, show the hours saved, and move on to the next. That approach made budgets easy to approve.

4. Companies couldn’t hire more, and had to cut overtime

In 2017, the three megabanks announced large workload cuts one after another. In April 2019, legal caps on overtime took effect for large companies. Companies couldn’t add people and had to reduce overtime. RPA got budget as a way to cut routine clerical work.

In March 2021, Microsoft made Power Automate Desktop available on Windows 10 at no extra cost, widening the base further. In MM Research Institute’s 2024 survey, Power Automate took first place among mid-sized and large companies with a 24% share.

Four reasons RPA’s growth stalled

In MM Research Institute’s surveys, adoption among companies with sales of 5 billion yen or more rose 16 points in a year and a half, from 22% in June 2018 to 38% in November 2019. In March 2024 it was 44%, down 1 point from the survey before.

The reasons growth stalled come after adoption, not before.

1. Robots stop when screens change

Most RPA tools find what to operate by screen coordinates or element IDs, assuming the screen looks as it did when the steps were recorded. Move a button on a business partner’s web system slightly, and the robot stops.

In a 2022 survey of IT teams by the RPA vendor Robocorp, 69% said their bots break at least once a week, and 41% said a fix takes more than five hours. Keep in mind that a vendor ran the survey. Why RPA bots break when screens change covers the causes in detail.

2. Too few people can fix them

In the ministry’s fiscal 2025 survey, the 842 local governments using or trialing RPA named their challenges, up to three each. 71% chose “no one, or not enough people, to work on it,” 42% chose “RPA technology is hard to understand,” and 30% chose “costs are high and budget is hard to get” (Descarty’s tally of the ministry’s published data). In MM Research Institute’s 2024 survey, too, about 70% of companies said they need outside help to automate across the company.

Few people could build scenarios. When the builder moves to another role, no one is left who understands what’s inside. Taking over an RPA bot whose builder has left explains how to take one over.

3. Built robots fall out of use

In the same survey, 730 local governments gave both the number of scenarios they had built and the number in use. Together, they had built 22,801 scenarios and were using 16,412. The other 28%, about 6,400 scenarios, are not in use.

Some of those cover work that no longer exists, so not all of them are failures. But 66% of the 842 local governments said they “have not set” KPIs to measure RPA’s effect. It’s easy to end up where no one is watching which robots are in use and which have stopped. How to build a robot ledger for RPA explains how to list them.

4. Work that needs judgment stays out of reach

RPA repeats set operations exactly as set. Work like reading invoices in varied formats, or sorting email by what it says, is hard to automate and stayed manual. Only the routine part of each process was automated, and people kept making the judgments before and after it.

Is RPA dead? Where the numbers stand now

By the numbers, RPA is not dead.

• Japan’s RPA market grew from 79.5 billion yen in fiscal 2022 to 90.3 billion yen in fiscal 2023. MIC expects 103.4 billion yen for fiscal 2024
• RPA is in use at 98% of prefectures, 100% of designated cities, and 44% of other municipalities
• Adoption among companies with annual sales under 5 billion yen is 15%, up 3 points from the year before

The period of rapid growth, though, is over. Market growth fell from 56.7% in fiscal 2019 to 13% in fiscal 2023, and adoption among mid-sized and large companies is flat at 44%. UiPath, a major RPA vendor, grew its ARR (annual recurring revenue) only 11% in the fiscal year ended January 2026, and now calls itself a leader in agentic automation.

The number of companies using RPA isn’t falling. What has changed is the main worry: from how to adopt RPA to how to keep the robots running.

What AI changes, and what it doesn’t

Expectations for AI agents are high. In August 2025, Gartner placed AI agents at the peak of inflated expectations in its hype cycle for cloud and AI in Japan. Gartner also says AI agents will not replace RPA for the time being: for processes that run often, AI agents can cost up to 30 times as much as an RPA bot, and their results aren’t always the same.

Set against the four reasons RPA’s growth stalled, here is what AI changes and what it doesn’t:

Why growth stalled	What AI changes	What doesn’t change
Robots stop when screens change	AI reads the screen to find buttons and fields, so small changes are less likely to stop it	If the process itself changes, the steps still need fixing
Too few people can fix them	Steps are written in business language and fixed by conversation, with no special development tool to learn	Someone still owns the process and checks the fix
Built robots fall out of use	AI alone doesn’t change this. Robots that are easier to fix are less often left stopped	People still decide the ledger, the numbers to watch, and when to retire a robot
Work that needs judgment stays out of reach	AI reads values from documents and screens	Actions that are hard to undo still need a person’s approval first

Don’t hand everything to AI agents

An AI agent that reads the screen and thinks on every run copes well with screen changes. But it waits for the AI’s response at every operation, so it is slow, and every run costs a usage fee. A process that handles hundreds of items a day doesn’t need AI to think through each one every time.

The practical shape is for AI to learn an operation once, repeat it the way RPA does, and relearn it only when a changed screen blocks the way. It keeps RPA’s speed and the certainty of the same steps every time, while robots stop less and are easier to fix. AI doesn’t replace RPA. It changes RPA’s weak point, what happens after a robot is built. Descarty calls this shape AI RPA.

RPA problems compares RPA, AI agents, and AI RPA. AI agents vs RPA explains how to choose between them step by step.

What to do with the RPA you have

There’s no need to replace every robot at once. Decide robot by robot:

• Repeats set operations at volume on an internal system whose screens rarely change: keep it on RPA
• The other system offers an API or CSV export: replace it with the official integration
• Operates an external web system whose screens change without notice: switch it to AI RPA
• The builder is gone and no one knows what’s inside: put it in the ledger first, then decide
• No longer used: stop it and remove it from the ledger

RPA migration steps explains how to make the switch.

Frequently asked questions

Will AI agents replace RPA?

Not for the time being. Gartner, too, says AI agents will not replace RPA for now. Work that runs set steps at volume, fast, the same way every time, still suits RPA. What changes is how you automate work whose screens change often or that needs judgment.

Is it too late to adopt RPA now?

No. Adoption among small companies is 15% and still rising. But the walls companies hit during the boom came after adoption. Before choosing a tool, check what happens when a screen changes, whether someone other than the builder can fix it, and whether you can see a list of robots and a record of their runs.

Will RPA staff lose their jobs?

AI will take over more of the work of building scenarios out of screen parts. But writing a process’s steps accurately in words, deciding where people approve, and watching the robot ledger and numbers to decide what to fix or stop all remain. Experience running RPA carries over directly.

Change what happens after a robot is built, with Kitewell

Descarty’s Kitewell is AI RPA that AI builds and AI fixes. Business staff build by conversation, and fix by conversation. It covers the four reasons RPA’s growth stalled this way:

• Less likely to stop when screens change: write steps in words, and AI reads the screen to operate it. Once an operation works, Kitewell learns it and repeats the same steps from then on. AI relearns an operation only when the screen changes, and the run record shows when it did
• Fix by conversation: ask the assistant, and it shows the change and a preview of the flow. Nothing is saved until you choose Apply, and Undo takes it back. Business rules and corrections you state in the chat are kept under What Kitewell knows, so whoever takes over starts from the same place
• It tells you when a run stops: failures and approvals waiting are sent to Slack, Microsoft Teams, email, and more. Runs & logs shows failed runs and the success rate
• Judgment through AI reading and human approval: AI reads values from documents and screens. Any step can require approval, and each approval, send-back, or rejection is recorded
• See everything: export a robot ledger to Excel

In the PoC, Descarty builds the workflows together with your team to prove the effect. In production, your staff build and fix the workflows by talking with Kitewell’s assistant.

You can start by choosing one robot that keeps stopping, or one whose builder has left, and verifying it with us as an early access partner. See Early access partners for details.

Summary

• RPA spread for four reasons: no changes to existing systems, staff could build robots themselves, results showed up as hours saved, and companies couldn’t hire more
• Its growth stalled for reasons that come after adoption: robots stop when screens change, too few people can fix them, built robots fall out of use, and work that needs judgment stays out of reach
• By the numbers, RPA is not dead. The market is growing and reaching local governments. The main worry has moved from adopting robots to keeping them running
• AI changes how robots are built and fixed. Rather than handing everything to AI agents, the practical shape is AI repeating what it learned and relearning only when a screen changes

Sources

• PRESIDENT Online: Why the three megabanks’ expense ratios keep creeping up (September 4, 2018, from the March 19, 2018 issue of PRESIDENT; in Japanese)
• SMBC Group 20-year history: Cost control and productivity through business reform (in Japanese)
• Information Age: RPA software market grew 63.1% in 2018, says Gartner (June 24, 2019). The Gartner analyst’s remark on legacy systems is from the same article
• IT Leaders: RPA adoption among large Japanese companies tops 50% (February 10, 2020, on MM Research Institute’s survey; in Japanese). Adoption in June 2018 and November 2019 is from the same article
• Yano Research Institute: RPA market survey 2020 (December 7, 2020; in Japanese)
• IT Leaders: RPA in Japan expands from departments to the whole company; IT develops scenarios at 57% of companies (February 21, 2020, on Gartner’s survey; in Japanese)
• Microsoft: Download Power Automate Desktop for Windows 10 to automate tasks and processes at no additional cost (March 2, 2021)
• IT Leaders: Japan’s RPA market grows steadily at over 10% a year (November 22, 2024, on Deloitte Tohmatsu MIC Research Institute’s survey; in Japanese)
• IPA: DX Trends 2025 (June 26, 2025; in Japanese). Legacy systems are in section 2.5
• MM Research Institute: RPA use in Japan 2024 (July 24, 2024; in Japanese)
• Ministry of Internal Affairs and Communications: Results of the fiscal 2025 survey on AI and RPA in local governments (April 24, 2026; in Japanese). Challenges, KPIs, and scenario counts are Descarty’s tally of the RPA adoption results spreadsheet on the same page (as of October 31, 2025), which leaves out governments that did not agree to publication
• Information & Data Manager: Broken Bots challenge RPA users (July 8, 2022), on Robocorp’s “The State of RPA” survey (May 2022)
• UiPath: UiPath Reports Fourth Quarter and Full Year Fiscal 2026 Financial Results (March 11, 2026)
• ITmedia @IT: RAG and platform engineering enter the trough: Gartner’s hype cycle for cloud and AI (August 13, 2025; in Japanese)
• ITmedia @IT: AI agents will not replace RPA for the time being (January 16, 2026, contributed by Gartner, in Japanese)