AI Agent Workflows Lean on RPA as Microsoft Unveils Autopilot

The week agents started doing the work

For most of the past two years, the pitch for AI agents has been that they can reason. The news from late September 2026 was about something else: whether agents can reliably carry out work inside the systems businesses already use. A Microsoft product overhaul, a small but telling acquisition by HCLSoftware, and a large set of platform updates from UiPath all point the same way. The hard part of AI workflow automation is now the execution layer: the connectors, robots, permissions, and audit trails that turn a model's decision into a completed business process.

The theme is clear across the coverage. What the coverage disagrees about is how close any of this is to production, and who will collect the revenue.

Microsoft turns Copilot into a builder and a background worker

The biggest launch came on September 25, when Microsoft rebuilt its Copilot app around three tabs: Home, Code, and Autopilot.4148 Home brings chat and Cowork-style delegation together in one place. Code lets people who don't write software describe an app, tracker, dashboard, or automation in plain language and have Copilot build it. Autopilot is a persistent agent that keeps working after the user logs off.50 Satya Nadella called it the company's biggest Copilot update so far and described the product as a "new OS for work."4647

Autopilot is the piece that matters most for business process automation. Several outlets report that it is the renamed Scout, which Microsoft first showed in June as a desktop app. It has now become a cloud-hosted agent with its own Entra ID identity, memory, and workspace inside a customer's tenant.414245 An owner gives it a name, a role, and a goal. It can then watch channels, follow up on threads, run recurring work, and resume projects days later.45 One analysis put it well: the update shifts the unit of AI work from a single prompt to an ongoing job.43

Code is Microsoft's biggest move yet into no-code AI automation. Commentators note that it runs on the same technology family as GitHub Copilot. Apps built with it can be hosted in a company's own tenant on Copilot Managed Runtime, which is in preview.4345 Two weeks later, Microsoft's Copilot Studio blog made a related case for makers: apps, deterministic workflows, and reasoning agents should be built in one place, each used where it fits best.15

Where reports on Microsoft differ: availability and billing

The coverage does not agree on how much of this is usable today. One explainer says Microsoft "launched" Autopilot on September 25.42 Most others stress that the announcement was a staged rollout, not general availability. Home and Code were going to the Frontier early-access program in the following weeks, Autopilot was entering private preview at the end of September, and broad availability for Autopilot was planned by the end of 2026.474544 One account dated September 30 already describes Home and Code as available through Frontier.43 The honest reading is that this is a preview cycle, and buyers should plan for it as one.

The more important change may be the pricing. Microsoft now splits the bill in two. A fixed per-user license covers everyday use, and agent-heavy work in Cowork, Code, and Autopilot is billed by usage in Copilot Credits.4147 One commentary called the announcement less a technical breakthrough than a change of business model. It noted that Microsoft's admin documentation on launch day still listed only Cowork, Cowork-built apps, and the Work IQ API as credit-billed services.49 Another report adds that admins must set up a billing policy before users can run Code at all.44 For finance teams, that turns AI workflow automation from a seat cost into a variable operating expense, with all the forecasting work that comes with it.

RPA gets a second life as the agent's hands

On September 28, HCLSoftware announced that it intends to buy Robotiq.ai, a robotic process automation vendor based in Zagreb, and fold it into its HCL UnO Agentic platform.14 The company's stated reason is the most candid explanation of the week for why agents still need RPA. Robotiq.ai's bots let AI-driven workflows act in applications where APIs don't exist or aren't good enough. In HCL's words, that extends UnO Agentic "from decision-making to execution."1 Robotiq.ai's platform is used by large banks, insurers, and telecom providers, and the company highlights ISO-certified security and audit logs.5 The deal is expected to close in November 2026.4

The reporting splits on price. The press release and several write-ups say financial terms were not disclosed.57 One analysis says HCLTech's filing with India's stock exchanges does give a number: an all-cash deal at an enterprise value of €9 million, subject to adjustments, with no regulatory approvals required.6 The same analysis notes that Robotiq.ai's revenue was flat at €0.9 million for two years before rising to €1.4 million in 2025. That suggests HCL bought a capability more than a scaled business.6 Its sharpest point is the irony that the newest layer of the enterprise stack, agent orchestration, is being completed by one of the oldest.6

That irony is the key to the week. Legacy systems with no API are still common in regulated industries, and models alone can't bridge them. RPA, often written off as yesterday's technology, has become the adapter agents need. Market research released the same week reflects the shift. It projects the AI-in-RPA segment growing from $4.79 billion in 2025 to $5.6 billion in 2026, and says vendors are building platforms that combine agents, robots, and human workflows.10

UiPath bets on orchestration and governance

UiPath, the best-known RPA vendor, used its FUSION event in Las Vegas to argue that orchestration and control matter more than any single agent. Its September 23 platform update named three reasons automation programs stall: queues for developer time, legacy systems that platforms can't reach, and compliance teams needing proof of what agents actually did.36 Its answer included the general availability of UiPath Coding Agents and a new productivity agent called Delegate. It also added an MCP connector and an on-premises relay, plus governance tools: a Runtime Checker that tests agent behavior against policy while the agent runs, an LLM-as-judge guardrail, and compliance packs.36 Notably, UiPath lets customers bring outside coding agents such as Claude Code, Cursor, and Codex into its environment. That is a quiet admission that the model layer is becoming interchangeable.36

The newest piece is UiPath Cartographer, which builds a "Map of Work" by analyzing documentation and interviewing subject-matter experts about how tasks actually get done.37 CEO Daniel Dines argued that agents lack the informal, unwritten knowledge employees pick up over time, so a company needs an accurate picture of its processes before agents can run safely. He described a trust ladder in which an agent earns more autonomy for a class of decisions, such as low-value invoices, only after proving high accuracy.37 He was blunt about the company's position: "We don't want to play in orchestration alone."37 UiPath also expanded a partnership with BDO USA to build agent-driven accelerators for internal audit and controls testing, with professional review kept in the loop.39

Wall Street is less convinced than the product announcements suggest. DA Davidson downgraded the stock to Underperform on September 24.34 Truist and Canaccord both cut their price targets from $17 to $14, and coverage calls UiPath a "show-me" story despite long-term margin targets above 80% gross and around 30% operating.3832 One equity analysis says plainly that the risk is UiPath losing ground to bundled offerings from larger platforms.40 Microsoft's Copilot Code and Autopilot are exactly that kind of bundle.

Vertical agents and the control layer

Outside the platform fight, agents built for specific industries kept reaching general availability. Cognizant made Workflow Agentic Processing generally available for its TriZetto health-plan claims platforms, with more than 100 MCP tools that let agents handle routine pended claims and send exceptions to people.14 Trintech released three finance agents covering data access, accruals, and exception management in the financial close.14 Synopsys introduced long-horizon engineering agents, and Momentic launched scriptless test automation.18

Governance is also becoming a product category of its own. Microsoft's Copilot Studio added "hooks" that can inspect or block an agent's tool call before it runs.15 Noma launched tools for finding and governing agents on employee devices.18 Earlier in September, JetStream introduced per-action authorization meant to stop risky action sequences before they execute, not just log them afterward.11

The reading

Taken together, the late-September news supports a clear conclusion: the competitive line in AI workflow automation has moved from model quality to execution and control. Microsoft wants agents to live inside the productivity suite and be paid for by usage. HCL is betting that RPA is still the only reliable way into systems without APIs, and UiPath is betting that governed orchestration across bots, agents, and people will be harder to replace than any single agent. All three agree that agents need deterministic support. They disagree about who should own it.

For buyers, the practical takeaways follow from that. Much of what was announced is still in preview. Usage-based billing will make costs harder to predict. And the least glamorous parts of a deployment, including connectors, audit logs, and approval rules, are likely to decide whether any of it reaches production.