AI Workflow Automation Turns to Governed Agents as RPA Retools

The pitch changed this fall

A year ago, most people bought workflow automation as a set of rules: when a form arrives, copy the fields, route the ticket, send the email. The product announcements from the past several weeks describe something different. Vendors now pitch the agent, a system that takes a goal, picks its own steps, and works across applications, as the default unit of business automation. Selling the idea of autonomy is no longer their main job. Their pitch now is control.

The largest platform companies have shown this most clearly. At its "Gemini at Work" event, Google Cloud described a Gemini agent that receives objectives rather than step-by-step instructions, splits the work among sub-agents, and can pass tasks to non-Google models, beginning with Anthropic's Claude.18 Google Cloud CEO Thomas Kurian said such an agent only works if it is connected to a company's personal workflows, systems of record, and enterprise controls.18 Google also says each agent gets a cryptographic identity and an isolated sandbox and can be governed through an Agent Gateway, though those capabilities were not yet publicly available when announced.11

Microsoft is building along similar lines. In October it began rolling out a preview of Work IQ, which connects Copilot and autonomous agents to live Dynamics 365 and Power Platform data, including CRM records, ERP data, and low-code automation flows. Timing for general availability and pricing has not been announced.11 Copilot Studio added "Hooks," a feature meant to make sure workflows start reliably without depending on an agent's judgment about whether to begin.13 A later preview added agent accounts, shared memory, and triggers.1216

Enterprise suites are becoming agent front doors

The big business-application vendors are moving the same way. Workday made Sana the main agentic interface for HR, finance, and IT, and made its Agent Harness generally available with memory, guardrails, regional data controls, model choice, evaluations, and step-by-step audit traces.11 ServiceNow launched an AI Workflow Factory and an Autonomous Engineer to link the discovery of automation opportunities with unattended execution, explicitly to move agents out of experiments and into governed production.12 It also released a third-generation Employee AI Harness that lets employees build agents for daily tasks without code, plus a developer environment called ToolSmith.16 SAP is turning its Joule agent into a broader "agentic work layer" tied to its Autonomous Enterprise initiative.13 Cisco added an agent framework to Webex so autonomous agents can join meetings and spaces and carry out multi-step work across calendars, documents, and third-party apps, with human oversight and audit trails.11

These vendors are converging on a single structure. Each wants to own the layer where people state a desired outcome and an orchestrator assigns agents, tools, and people to deliver it. Each also describes governance in nearly the same terms: identity, permissions, memory, audit trails, evaluations. That shared language suggests buyers have made governance a condition of purchase, and vendors are competing on whose control plane will run most of an organization's agents.

RPA vendors are rebuilding around orchestration

Robotic process automation is the older approach: scripted bots that click through screens and move data between systems. The leading RPA companies are not defending that approach. They are presenting it as one layer inside a larger orchestration product.

UiPath had the most extensive set of releases. At its FUSION conference in September, it made Cartographer generally available. The tool drafts and maintains a "Map of Work," a store of business rules, policies, and expert judgment inside the customer's own tenant that both people and agents can read.10 UiPath split its Maestro orchestrator into two products: Maestro Orchestrate for long-running processes that can hold their state for weeks, and Maestro Automate for high-volume short tasks.10 It also released Delegate, a governed desktop agent that performs tasks described in plain language or shown once, under the same identity and audit controls as the rest of its platform.10 Another product, UiPath for Coding Agents, lets developers build workflows with whichever coding agent they prefer, with evaluations required to pass before each deployment.10 Financial coverage in May described this as an attempt to sit above competing coding agents, such as Claude Code and OpenAI Codex, as the governance layer. That coverage also noted that the near-term financial effect was not yet clear.5

UiPath is also extending its range. Its self-hosted Automation Suite now offers agentic capabilities with cloud-hosted or fully self-hosted models, aimed at government and regulated industries with data-sovereignty requirements.4 The company's public-sector announcement frames agentic automation as the next stage after RPA, which agencies already use for rules-based work like data entry and reporting.1 UiPath says an October long-term-support release adds Conversational Agents and its IXP document-extraction technology to the on-premises product.7 One independent analysis cautioned that running models and agents on-premises usually means buying inference hardware and managing model updates locally, which affects latency and total cost of ownership.6

Automation Anywhere made two moves in the same period. It announced the acquisition of conversational-agent company Boost.ai and released Automation 360 v.42 with Model Context Protocol (MCP) support, so its bots and agents can work with outside agent runtimes.17 Rocket Software acquired Geniez AI, whose agents run directly on mainframes to monitor systems, diagnose incidents, and run workflows from natural-language instructions.11 Taken together, these deals show RPA's established strength in getting into legacy systems being repackaged as the dependable "hands" for reasoning agents.

No-code tools are merging with their agent products

At the no-code end of the market, the platforms are combining their workflow and agent features. On October 6, Zapier folded its Zapier Agents product into "AI by Zapier," a single AI step that can be dropped into any Zap. It keeps the agents' tool-calling and knowledge sources and adds triggers, filters, paths, and run history.22 Zapier now advertises connections to more than 9,000 apps.2225 This is a practical choice: instead of running agents as a separate product, Zapier is putting reasoning inside the trigger-based workflows its customers already use.

n8n is still the main alternative for teams that want control over their data. Comparisons describe it as self-hostable, with native agent nodes and approval gates before sensitive tools run.2223 One earlier analysis cited a $2.5 billion valuation for n8n from October 2025 and argued that its self-hosted option clears compliance requirements that hosted rivals cannot.26 Pricing models now differ widely. n8n charges per workflow execution. Zapier counts tasks per step. Microsoft bills Copilot Studio in credits.22 One hands-on test found that a ten-node workflow counts as ten operations on Make or Zapier but as a single execution on n8n.21 The no-code field is also losing projects: one review reported that FlowiseAI was archived in August 2026.25

Marketing technology shows how far the agent label has spread. A single week of releases included CreatorIQ's agentic system, which automates workflows within approval guardrails, and an ingestion agent from Guideline that reads unstructured media plans so people don't have to re-enter the data by hand.15 Seismic and Zig.ai connected their tools to external large language models through MCP servers.15 Across these releases, MCP appears to be the common connector between applications and agents, in no-code tools, RPA suites, and SaaS products alike.

How the coverage differs: ambition versus maturity

The vendor announcements and the independent reporting don't fully agree. Vendors describe delegated outcomes and finished work.18 Independent briefings say enterprises are still assessing what agents actually change.12 One weekly summary reported that survey data shows plugging in off-the-shelf agents rarely produces end-to-end automation or clear return on investment. It identified data quality, process mapping, and risk controls as gaps that need to be closed first.11 The small-business numbers point the same way: Upwork's survey found 34 percent of leaders at firms with 10 to 99 employees piloting AI-agent workflow automation.14 That figure counts pilots, not full adoption. The most common automations listed are still basic: assigning leads, routing tickets, updating records, processing invoices.14

Security reporting adds to the caution. Gambit Security described fully automated attack services that use open-source AI tools to hit hundreds of sites at low cost per attack.11 Identity vendors have responded quickly. SailPoint introduced agent governance with just-in-time privilege controls,17 and RSA introduced an Agent ID platform for regulated organizations.11

The two kinds of coverage can be read together. Vendors are not ignoring the maturity gap. Most of what they shipped this cycle is meant to close it: Microsoft's deterministic Hooks, UiPath's evaluation-gated deployments, Workday's audit traces.131011 The industry has moved from showing that agents can act to showing that they can be trusted.

The business-model question

The longer-term change may be in pricing. Workday told investors it wants agents paid for out of customers' labor and outsourcing budgets, not just IT budgets. Its chief AI officer, Joel Hellermark, pointed to customers spending tens of millions on outsourcing, and the company expects agent bills to grow with usage rather than with licensed seats.20 Workday also opened "headless" MCP access, which lets agents work in its HR and finance systems without going through its interface, still subject to permissions and approvals.20

If that model spreads, buyers will judge automation by the work it replaces rather than the seats it needs. Pricing will move toward consumption, as it already has in no-code tools.22 Our reading is that orchestration and governance, not the underlying models, are becoming what vendors compete on. The companies that control the audit trail and the process map are likely to capture the budget. For now, the common deployment remains a pilot, with a human approving the agent's actions.