You have probably heard of UiPath, an RPA company. Put simply, the RPA business is about teaching computers to mimic human mouse and keyboard actions to run repetitive, cross-system tasks for people—like copying invoices from an inbox into an accounting system. The workflows themselves are rigid and have to be pieced together by humans using specialized development tools. As a result, for over a decade, the company made money from two main things: per-seat development tools, and runtime licenses once workflows were up and running.
On August 19, 2026, UiPath launched Maestro Flow, a new workflow file format and runtime entry point. The most eye-catching rule in the announcement: workflow files generated by AI coding assistants like Claude Code, Cursor, Copilot, and Codex can run directly on its production platform without changing a single word.
To appreciate the weight of this move, you first need to understand one thing: in the RPA world, workflow authoring has always been the primary cash cow. If you wanted a computer to run a workflow, someone first had to build it using UiPath’s development tools. These tools charged an annual fee per seat—one developer in an office meant one license fee. This revenue stream sustained the company for over a decade. Now, it announced: the authoring step is given away for free. No matter how many workflows AI writes, there is no charge; money is collected on a per-run basis only after the workflow is executed.
For a company that makes its living selling development tools, giving away workflow authoring for free to third-party AI looks like dismantling its own primary revenue pillar with its own hands. This article walks through the complete backstory of this move: what it used to charge for, which part AI wiped out, how it got here step by step, and why—when you do the math—the numbers actually make sense.
An accounts payable specialist at a company spends three hours every day entering invoices. In the traditional RPA era, there was a standard playbook for this kind of work. First, install a recording tool on the employee’s machine and let it run silently for two weeks, capturing screenshots, where the mouse clicked, and what the keyboard typed. The backend aggregated these logs into a task path map, concluding that invoice entry had the highest repetition rate and best ROI, so automate it first.
Then the consultants stepped in—usually UiPath-certified developers
from the Big Four or system integrators. They dragged and dropped
flowcharts in Studio, an IDE similar to Visual Studio: watch the inbox,
download the PDF attachment, extract fields, switch to the SAP interface
to enter values field by field, and click save. What came out was a
.xaml file—essentially a UI interaction script: find the
button labeled btnSave and click it. The tags used to
locate UI elements were called selectors. Once built, the workflow was
published to the server-side Orchestrator, where robots charged per
license replayed these clicks in virtual machines in the middle of the
night. Whenever SAP updated, renaming a button or shifting a page
element, the selector broke, and the workflow threw an error and halted
at 2 AM. In the past, such failures had to wait for engineers to fix
them when they came to work the next day; later, the platform added an
AI feature called Healing Agent for automated fixes.
This approach left behind two chronic pain points—keep them in mind, because the new engine is built precisely to solve them. The first problem was waiting: when a workflow reached the finance manager approval step, the robot either hogged a license sitting idle for three days, or developers had to awkwardly split the workflow into two parts—the first half exiting after writing to a queue, and the second half polling periodically for approval results, relying on humans to manually reconcile data in the database. Approval is the most routine step in an enterprise, yet in the old architecture, it was the most failure-prone. The second problem was recovery: if a workflow crashed halfway through the night, intermediate progress was not saved, so it had to start all over again from scratch.
The revenue streams were already hinted at earlier: Studio charged annual per-seat fees for developers, and robots were billed per license. The most lucrative part was actually the third component: consulting work helping enterprises get the system running—namely dragging flowcharts and fixing broken selectors mentioned above. The lion’s share of this money went to consulting firms rather than UiPath’s pocket. In its own financial reports, services revenue accounted for only 3.9%, but it was this ring of consultants that brought the software into major enterprises. The total top line in the latest quarter (as of the end of April this year) was $418.4 million.
We wrote an article in June noting that the valuable part of automation is shifting from UI operation scripts to business rules themselves. At the time, UiPath had the most comprehensive product lineup, but its modules spoke past each other. This release finally aligns the story.
The threat is simple enough. With the arrival of coding agents, dragging flowcharts in Studio is something AI can do directly: give it a requirements description, and it writes out the entire automation script. Authoring workflows costs next to nothing, instantly eroding the foundation of a business built on developer seats and consultant hourly rates. UiPath’s ARR growth rate also slid all the way down to around 12% and stalled there.
Its response came in three steps, each on its own a standard
defensive move for an RPA company. In April 2025, it released the Maestro
orchestration engine, orchestrating not just robots, but also AI
agents and human approval steps—aimed squarely at those two pain points.
In May 2026, it released UiPath
for Coding Agents, allowing Claude Code and Codex to connect to the
platform with enterprise compliance requirements. In late July 2026, RPA
with Coding Agents became generally available, enabling coding agents to
directly generate traditional .xaml projects.
Up to this point, the story remained one of defense: since flowcharts were destined to be written by AI anyway, let AI write them, as long as the output stayed in its proprietary format.
The August 19 release changed the playbook. Maestro Flow defined a
new workflow file format, .flow—a JSON node graph
specifying inputs and outputs. The key design is that three different
paths can modify the exact same file: business users drag and drop on a
web canvas, engineers code in VS Code, and coding agents generate it
directly via the official CLI. There is no intermediate format
conversion; what Claude Code generates is the production-ready file
itself, and humans can open the canvas at any time to inspect and
fine-tune that exact same file.
While opening up generation, the platform put in place most of the guardrails required for production environments. Before publishing a workflow, the platform enforces structural and syntax validation; the CLI can pull down organizational security policies to check agent nodes for compliance; after every workflow modification, you can run regressions against built-in test suites; and approvals are built directly as nodes in the workflow, pushed to Slack or Teams waiting for human action.
The monetization point shifted as well. According to the official pricing model, the tooling side no longer charges per seat, and generating workflow files using your own Claude Code subscription incurs no extra charge; each time a workflow executes in a production environment, it consumes one platform billing unit. The product lead put it bluntly: enterprises do not lack agents today; they lack the orchestration to make agents work.
Intuitively, this looks like giving away one’s own development tools for free. If the seat-based era were still intact, this step would be equivalent to tearing down the ticket booth. It looks like surrender.
Founder Daniel Dines did the math on an earnings call in March, which translates to: the lower the cost of building software, the more valuable a platform becomes that can safely govern, orchestrate, and scale that software; the cheaper things are to build, the more things will be built, which will not decrease enterprise demand for orchestration—it will only increase it.
Translated into business terms: as workflow authoring approaches zero cost, monetization opportunities have not disappeared—they have simply shifted from build to runtime. Giving away generation for free is not an act of charity; it is moving the toll booth onto the highway, betting that traffic on the road will only grow.
Not every company can do this math. Facing AI in the same way, Microsoft could not make this choice because the core assets in their respective vaults are completely different. Microsoft’s legacy revenue serves as the gateway in the agent era: M365 enterprise tenants and search ad inventory. Microsoft sells agents as add-ons to its productivity suite, resulting in a paid penetration rate of only 3.3%; Cowork, which is genuinely usage-based and delivers finished outputs directly, is tucked behind an experimental toggle turned off by default—because honest usage-based pricing amounts to bleeding its per-seat subscriptions. From the same company, GitHub Copilot is sold directly to developers with pricing aligned with value, and 20 million users gladly use it. One product has 20 million users while the other has a 3.3% penetration rate; the difference lies not in technology, but in the ledger behind each.
UiPath’s situation is the exact opposite. It holds neither enterprise tenants nor traffic entry points; its original money came from developer seats and a network of consulting partners. Both of these became commodities once coding agents proliferated. The old toll booth was bound to be torn down anyway and could not be defended; tearing it down proactively and betting all monetization on the runtime layer is actually the logical path forward.
From this, a framework emerges that can be applied to any incumbent software company: examine whether its legacy revenue acts as an entry point or a commodity in the agent era. Companies with entry points will set up toll gates and defend them, packaging AI as an add-on on their old ledgers; only companies whose legacy revenue has become a commodity dare to truly charge based on usage and collect revenue at the runtime layer. Automation Anywhere released a nearly identical orchestration and governance framework in May, but did not take an equivalent step in integrating external coding agents—the difference lies along this exact line.
Making the numbers work does not mean there is no risk, and the conditions for winning and losing must also be laid out. First, the bulk of consulting work has long belonged to consulting firms, and coding agents will first squeeze the workload of these consultants—who historically brought UiPath into large enterprises. If partner channels shrink before new usage-based revenue catches up, the transition will stall. Second, the data currently supporting this narrative comes entirely from the company itself: annualized revenue from AI products is approaching $200 million, 60% of $100k+ customers have adopted AI products, adopters spend about 3x more on average than non-adopters, and expansion deals involving AI average about 6x larger than those without. The numbers look healthy, but they span only four or five quarters and have yet to be independently verified.
For a toll booth to collect fees, the highway itself has to be real. What this engine does directly addresses those two chronic issues mentioned above.
The issue of waiting for approvals is solved by durable execution. With every step the workflow takes, the engine records the result into an append-only journal. When reaching an approval node, the engine saves the entire workflow state and exits the process entirely, consuming zero compute resources. Three days later, when the approval is clicked in Slack, the engine replays the journal, fully restores the previous state, and resumes right from the breakpoint. To the workflow author, the code appears to run continuously from start to finish in one breath, while waiting and recovery are entirely handled by the engine.
The midnight crash problem is addressed along the way by the same mechanism. If a process dies, the engine replays up to the last step before the crash and resumes from there. Each node can be configured with retry policies and compensation actions: in a ten-step workflow, if step ten fails, it can automatically roll back what step seven did. Combined with pre-publish mandatory validation, regression test suites, and built-in approval nodes in the workflow, the platform takes care of all four: waiting, recovery, approvals, and auditing.
There is a neat symmetry here. The environment for coding agents is a
deterministic shell wrapping a non-deterministic model: constrained by
context documents, tests, and linters before action, and backed by logs
and reviews after action, tolerating model instability through these
guardrails. Maestro is the exact inverse: a deterministic skeleton
wrapping non-deterministic nodes. How the flow proceeds, retries, and
compensates is hardcoded in .flow; what sits inside the
nodes is the judgment of agents and humans. This also explains why
orchestrating business workflows with a bare coding agent requires
building state, recovery, approvals, and auditing entirely from scratch;
UiPath’s confidence in charging per instance also stems from here: it
charges precisely for the cost of keeping workflows alive in
production.
Boundaries should also be made clear. Change approvals, multi-environment promotion, and version progression for the workflow files themselves have not yet been detailed in official documentation; governance in this area remains empty for now.
The success or failure of this move does not require faith; there are falsifiable metrics to watch.
The most critical thing to watch is a single indicator: the composition of UiPath’s net new ARR over the next two to three fiscal years. If the share of AI and orchestration products continues to rise, the net retention rate holds steady above 109%, and partner channels don’t collapse first, then this move will have worked. Conversely, if seat and consultant channels shrink faster and the $200 million in annualized AI revenue fails to grow into a new pillar, then “orchestration becomes more valuable” remains nothing more than earnings call rhetoric.
There is also an external variable. If Anthropic or OpenAI one day bundles workflow generation with their own runtime, UiPath’s intermediary business built around an open format could easily be captured from upstream—and Microsoft has already closed the loop on this path within its own ecosystem.
Finally, I leave you with that framework. In your own company, or any company you are evaluating: is legacy revenue an entry point or a commodity in the agent era? If it is an entry point, don’t expect it to disrupt itself; if it is a commodity, watch whether it has the guts to move its toll booth onto the highway. UiPath has already placed its opening move on the table.