Power Automate Desktop alternative: limits and AI RPA

Power Automate is spreading fast among companies on Microsoft 365. In MM Research Institute’s 2024 survey, Power Automate took first place for the first time, with a 24% share among mid-sized and large companies. Power Automate for desktop (Power Automate Desktop below), which operates screens on Windows PCs, can be used at no extra cost in some cases, which is one reason it spread.

The more you use it, though, the more often you step outside what is free, hit the time limits known in Japan as the “30-day wall”, or run into questions about where data goes. This guide sets out each of these from Microsoft’s documentation, how to tell which work to keep in Power Automate and which to move, and how to migrate.

In short

• What is free is running flows on your own PC, started by a person and signed in with a Microsoft account. Starting from the cloud or running unattended needs a paid account
• The “30-day wall” refers to cloud flow runs stopping after at most 30 days, and approvals after 28
• AI self-healing is in preview and recovers at run time only. It needs an organizational Premium account and permission to move data across regions

What is free, and when you need to pay

Use	What you need
Running desktop flows on your own Windows PC, started by a person	A Microsoft account (no extra cost)
Starting from a cloud flow, or on a schedule	An organizational account and a paid license
Running with nobody present (unattended)	An organizational account and a license for unattended runs
AI self-healing (preview)	An organizational Premium account

Premium is equivalent to ¥2,248 per user per month (billed annually, excluding tax). When automation that started as a flow someone runs on their own PC becomes department work that must run on a schedule without anyone present, it needs a paid contract and a managed environment to run in.

What the “30-day wall” is

The “30-day wall” is a phrase used in Japan for the time limits on Power Automate cloud flow runs. Microsoft’s documentation gives these limits:

Limit	What it means
Run duration: 30 days	A run lasts at most 30 days; after that, all pending steps time out
Approval wait: 28 days	An approval can wait up to 28 days; beyond that, the flow fails
Run history retention: 30 days	Run history is kept for 30 days
Inactive flows turned off: 90 days	Flows whose triggers have not run for 90 days are turned off (with notice 30 days before; flows owned by holders of Premium and similar licenses are exempt)
Trial: 30 days	Paid features can be tried for 30 days

The limit that most often causes trouble is the approval wait. An approver on leave or simply not responding is enough to make the whole flow fail. Work with approvals that span a month-end, or several people checking in turn, needs reminders or escalation built in separately.

AI self-healing (preview)

Power Automate Desktop has a “self-healing” feature in which AI looks for a screen element again when it cannot be found. As of October 2026 it is in preview, and Microsoft’s documentation lists these conditions and caveats:

• Recovery at run time only: self-healing works only at run time and does not fix the flow itself. A person is expected to fix the flow later
• It may pick the wrong element: Microsoft warns that the AI may identify an unintended UI element
• What it needs: an organizational Premium account and Power Automate for desktop 2.66 or later
• Moving data: outside the US, the setting to move data across regions must be on. It uses GPT-4.1 mini and Claude Sonnet 4.5
• Where it is unavailable: government and sovereign clouds

Whether screen contents may be sent to AI outside your region depends on your IT policy. If they may not, you cannot use this feature.

Work to keep in Power Automate, work to move

The work	Recommendation	Why
Stays inside Microsoft 365 and handles data you may send to cloud AI	Keep it in Power Automate	Its integration with Microsoft 365 pays off
Small routine tasks people start on their own PCs	Keep it in Power Automate	It costs nothing extra
Screen contents may not leave the company or the region	Move it to automation that runs with data kept inside	Self-healing depends on moving data, which you cannot do
Approvals drag on, and you struggle with the 28-day limit and home-made reminders	Move it to automation with approvals and records built in	Human decisions are handled inside the flow of the work
Built on a personal PC and fixable only by its builder	Move it to automation the business owners can fix by talking	It no longer stops when its builder moves on
Operates outside web systems whose screens change without notice	Move it to automation that saves what it relearned	No need to recover from the same change on every run

How to migrate

1. Collect the flows: gather desktop and cloud flows, their owners, and run history from the Power Platform admin center and each environment. Run history is kept for only 30 days, so export it early
2. Set the order by fragility and impact: order the moves by the number of failures, how often approvals timed out, and the impact of a stop on the work
3. Rewrite in the language of the work: do not copy the flow action by action; write the flow of the work in words, spelling out where a person decides and what to check
4. Prove it with a PoC: run it on the real screens and data, and check that it keeps going when a screen changes
5. Run in parallel, then switch: run both through a busy period, compare the results, then switch over

How to switch from RPA to AI RPA explains each step in detail.

How Power Automate Desktop and Kitewell differ

Aspect	Power Automate Desktop	Kitewell
Building	Arrange actions into a flow	Write steps in the language of the work. Build them by talking with the AI assistant, and apply them after reviewing the diff
When a screen changes	Self-healing (preview) recovers at run time. A person fixes the flow	When the remembered steps do not work, AI reads the screen again, relearns the operation, and repeats it from then on. The run record shows what was relearned
Data and AI	Outside the US, self-healing needs a setting that moves data across regions	With an in-house or local AI model, screen contents are not sent outside
Approvals	Approvals can wait up to 28 days	A person decides in an approval step, and what was decided, and when, is recorded
Where it runs	Windows PCs	PCs on macOS and Windows, internal servers, and closed networks
Price	Local runs cost nothing extra. Premium is equivalent to ¥2,248 per user per month (billed annually, excluding tax)	Team and Enterprise for organizations are quoted on the number of computers, never per user

Power Automate information is from Microsoft’s documentation as of October 2026. Features change, so check Microsoft’s official site for current information.

Start with one robot

Descarty offers a PoC that replaces one robot that keeps stopping with Kitewell in two weeks (¥300,000, excluding tax). If the agreed success criteria are not met, you pay nothing.

Sources

• MM Research Institute: RPA use in Japan 2024 (July 24, 2024, in Japanese)
• Microsoft: Power Automate pricing, using desktop flows with a Microsoft account, desktop flow requirements
• Microsoft: limits and configuration, known issues with approvals
• Microsoft: self-healing for desktop flows