In 2024, a large bank deployed 200 UiPath RPA bots. When one ERP vendor pushed a routine UI refresh, 180 of those bots broke overnight. Two weeks of engineering time later, the automations were back online — until the next portal change. That story, recounted across multiple RPA post-mortems, is the dirty secret behind a market that still claims 30 percent annual growth: traditional RPA is a brittle automation tax, and enterprises are quietly tired of paying it.
This week, Microsoft made that tax optional. Computer-using agents in Microsoft Copilot Studio reached general availability, rolling out across all commercial Power Platform geographies. The pitch is simple and uncomfortable for the RPA incumbents: instead of recording brittle click paths, you describe the workflow in natural language, and an agent uses vision and reasoning to navigate live UIs — even when fields move, layouts shift, or vendor portals get redesigned. For CIOs running 50 to 500 bots across legacy systems with no APIs, this is the first credible alternative to the UiPath/Automation Anywhere stack since 2018.
What Just Shipped
Computer use is now a standard tool inside Copilot Studio, alongside connectors and knowledge sources. Makers create an agent, navigate to Tools → Add tool → Computer use, and describe the task in plain English. Behind the scenes, the agent takes screenshots of the target application, reasons about what it sees, and chooses the next action — click, type, scroll, or wait — using the same visual and language models that power Copilot.
Key facts from Microsoft's announcement and accompanying Power Platform release notes:
- General availability: All commercial geographies, May 2026
- Pricing: Consumed via Copilot Credits at $0.01 per credit, billed at the "agent action" rate (one credit per significant interaction). Capacity packs run $200 per 25,000 credits per month, with up to 20 percent discount on annual prepurchase plans
- Governance built in: Allow-lists for sites and desktop apps, native Power Platform DLP enforcement, environment isolation, Entra-based identity, role-gated publishing (Environment Admin or System Customizer only)
- Auditability: Every session logged with screenshots, click coordinates, and reasoning trace. Logs propagate to Microsoft Purview, Dataverse, and via connector to Sentinel or any SIEM
- Human-in-the-loop: Configurable checkpoints at low-confidence steps and any action flagged sensitive
- Use cases targeted at launch: Legacy ERPs, mainframes, supplier portals without APIs, claims-handling vendor systems, custom line-of-business apps
This is not a separate product. It is a feature inside the Power Platform stack that 230,000+ organizations already license. That distribution advantage is the strategic punchline — Microsoft does not need to win a procurement war against UiPath. It just needs the feature to be "good enough" inside an existing license.
The technology itself isn't novel. Anthropic shipped computer use in late 2024, and OpenAI launched Operator in early 2025. What is new is enterprise-grade packaging: governance, DLP, audit, role-based access, and a billing meter that maps to existing M365 contracts.
Why This Matters
For CIOs and CTOs: The cost-per-automation curve just bent. Industry research from HfS puts RPA total cost of ownership at 70-75 percent maintenance and support, with licensing only 25-30 percent. Typical 50-bot deployments suffer 8-12 broken automations per major UI release, and 30-50 percent of RPA projects miss their original ROI target because of this brittleness tax. A vision-based agent that adapts when a button moves two pixels is structurally cheaper to operate, even if per-action costs are nominally similar.
For CFOs: UiPath enterprise licensing starts at roughly $420 per user per year for attended automation and climbs into the thousands per unattended robot. A 50-bot deployment routinely exceeds $100,000/year before services. Power Automate plus Copilot Studio costs 60-90 percent less per user, and computer use is billed inside the same Copilot Credits envelope — meaning no new vendor, no new procurement cycle, and no separate audit boundary.
For Risk and Compliance: This is the first major computer-use rollout with a complete enterprise control plane. Anthropic's reference implementation puts the security perimeter on the customer. OpenAI's Operator runs on OpenAI infrastructure outside the enterprise DLP boundary. Microsoft's GA wires computer use into the same DLP, conditional access, audit, and identity systems that already govern Power Platform — eliminating the "shadow agent" problem that has slowed enterprise computer-use adoption for the last twelve months.
For Business Operations Leaders: This is the unlock for the 20 percent of internal workflows that never made the API integration backlog. Vendor portals, B2B supplier sites, expense systems, regulator submission portals, claims handling, and bespoke industry software all become automatable by the same business analysts who build Power Automate flows today — without an integration project.
Market Context: Why Now
Three forces converged to make this launch land harder than it would have eighteen months ago.
Force 1: The RPA market is stuck. Gartner expects over 40 percent of agentic AI projects to be cancelled by 2027, and the most common failure mode named in Gartner's brief is legacy system integration cost. UiPath and Automation Anywhere have been pivoting hard into "agentic" branding, but their core revenue still depends on the brittle UI-recording model. Pure-play RPA vendors without a credible vision-agent story are losing renewals.
Force 2: Computer-use models got accurate enough. OpenAI's CUA scores 38.1 percent on OSWorld and 87 percent on WebVoyager; Anthropic's Computer Use scores 22 percent and 56 percent respectively. Microsoft has not published its OSWorld benchmark for the Copilot Studio implementation, but the GA reference customers — Holland America Line, several insurance carriers, and a number of retail chains — report enterprise-acceptable accuracy in production for the narrowly-scoped tasks they target. The line that mattered was crossed sometime in Q1 2026.
Force 3: Governance frameworks caught up. With the EU AI Act reaching full applicability on August 2, 2026 and Colorado's AI law taking effect, enterprises need to deploy AI agents inside an existing audit and DLP boundary. Microsoft's bundling of computer use into Purview-logged, DLP-policed Power Platform environments is a regulatory tailwind, not a technical one. Forrester's 2026 predictions warn that enterprises will defer 25 percent of planned AI spend into 2027 if they cannot demonstrate governance. Microsoft is removing that excuse.
Analyst perspective from the Forrester Total Economic Impact study of Copilot Studio: projected ROI of 106 percent and $25.7 million NPV for the modeled composite enterprise. That study predates computer-use GA. Customer references from the GA blog suggest the computer-use category will drive a step change in NPV because it unlocks the workflows that pure conversational Copilot Studio agents could not address.
Framework #1: Build vs Buy vs Stay Decision Matrix
The decision facing automation leaders is not "RPA or AI agents" — it is a three-way choice depending on the workflow profile. Use this matrix to triage existing automations and new requests.
| Dimension | Stay on RPA (UiPath/AA) | Switch to Computer-Use Agent | Skip Both — Use API Integration |
|---|---|---|---|
| Target system age | Pre-2010, no roadmap to APIs | Any system with stable UI | Modern SaaS with REST/GraphQL |
| UI change frequency | Stable, vendor-controlled | Moderate — vendor portals, B2B sites | N/A — uses backend contract |
| Volume per day | >10,000 transactions/day | 100-10,000 transactions/day | Any volume |
| Latency requirement | Sub-second | 5-30 seconds acceptable | Sub-second |
| Workflow complexity | Linear, rules-based | Branches, exceptions, judgment calls | Linear, deterministic |
| Maintenance budget | Dedicated RPA CoE in place | Business analyst maintainable | DevOps team available |
| Governance posture | Existing RPA controls | Power Platform / Purview required | API gateway + IAM |
| 5-year TCO (est.) | $250-500K per 50-bot fleet | $80-150K equivalent | $50-100K, but build cost upfront |
When to choose RPA (stay): High-volume, sub-second, deterministic workflows on systems that change less than once a year. Mainframe terminal emulation, Citrix-based enterprise apps, and batch ETL fall here.
When to choose computer-use agents (switch): Anything currently sitting on a "deferred — needs reintegration after vendor portal redesign" list. Insurance claims supplier portals, healthcare prior-authorization sites, government regulator submission portals, manufacturing vendor reorder workflows, retail return processing on third-party platforms. If the workflow includes the word "judgment" or "exception" in its description, computer-use beats RPA.
When to skip both (use APIs): Any workflow where the target system has a usable API and the engineering team has capacity. APIs are still cheaper, faster, and more reliable than either RPA or vision agents. Don't romanticize agents — use them for what RPA could never do.
The trap to avoid: Migrating stable RPA workflows to computer use just to "modernize." The TCO math only favors switching when the underlying UI volatility makes RPA maintenance the bottleneck.
Framework #2: 12-Week Computer-Use Pilot to Production Plan
The Microsoft EY $1B FDE program and ServiceNow-Accenture Forward Deployed Engineering launch both point to the same conclusion: 86-89 percent of agentic AI pilots fail to reach production because governance and integration are deferred. A computer-use rollout should not repeat that mistake.
Weeks 1-2 — Discovery and Triage
- Pull existing RPA inventory; tag each bot with "stable UI" or "frequent breakage"
- Survey backlog of unfunded automations rejected for "no API available"
- Select 3-5 candidate workflows scoring high on Framework #1 (switch column)
- Map data flows: what does the agent see, what does it touch, what does it write?
Weeks 3-4 — Governance Foundation
- Configure Power Platform environment dedicated to computer-use pilots
- Apply DLP policies: block consumer cloud uploads, restrict outbound destinations
- Build allow-list of approved websites and desktop apps for the pilot scope
- Assign Environment Admin / System Customizer roles to a named pilot owner
- Configure Purview audit pipeline; verify logs flow to your SIEM
Weeks 5-7 — Build and Bench-Test
- Author the first agent in natural language with one workflow only
- Run 50-100 supervised sessions; tune prompts and confidence thresholds
- Define human-in-the-loop checkpoints for any monetary transaction
- Test exception paths: what does the agent do when a field is missing?
- Measure: success rate, average actions per task, average credits per task
Weeks 8-10 — Controlled Production
- Move to production environment with full audit and DLP enforcement
- Cap daily volume at 10 percent of target — observe failure modes
- Stand up an on-call rotation for the first month (this is non-negotiable)
- Track three KPIs weekly: completion rate, exception rate, cost per completed task
Weeks 11-12 — Scale or Stop
- Compare actual cost per task vs RPA baseline and the original ROI model
- If completion rate is below 90 percent on the target workflow, do not scale — diagnose root cause first
- If completion rate is above 90 percent, expand to remaining pilot workflows and budget Year 2 capacity
- Document decision criteria; this is the playbook for the next 10 workflows
Common Challenges + Solutions
- Challenge: "We don't have a Power Platform CoE." Solution: Use this pilot to build one. The same governance discipline applies to all future agent work.
- Challenge: "The vendor portal redesigns mid-pilot." Solution: This is the test. Measure how quickly the agent recovers vs how long an RPA bot would have been broken.
- Challenge: "Audit team wants every keystroke logged." Solution: Purview captures screenshots, click coordinates, and reasoning. Demo this in week 4.
- Challenge: "Compliance won't approve outbound automation." Solution: Start with read-only workflows (data extraction, status checks) before writing.
- Challenge: "Cost per action is higher than expected." Solution: Computer use is not for high-volume deterministic flows. Re-check Framework #1 — you may have picked a workflow that belongs on APIs.
Case Study: Holland America Line's $0-to-50,000 Weekly Conversations
Holland America Line built its digital concierge "Anna" on Copilot Studio in a three-month MVP. Anna now handles approximately 50,000 customer conversations per week at 93 percent customer satisfaction, supports new and existing bookings, and operates in the US with expansion in flight to Australia, Canada, and the UK. The original implementation did not require computer use — the underlying systems had APIs. But the pattern is instructive for what GA-class computer-use deployments will look like in the back half of 2026.
What Holland America's success demonstrates and what new computer-use buyers should replicate:
- Narrow scope first: Anna started with three workflows (new bookings, add-ons to existing bookings, general questions). Computer-use pilots should do the same.
- Measurable resolution targets: 93 percent CSAT and a stated weekly conversation volume gave the program defensible ROI from week one.
- Three-month MVP cadence: Not three weeks (too brittle), not nine months (loses executive sponsorship). Three months matches the Microsoft GA reference pattern.
- Multi-region playbook: Once one market is stable, the same agent design ships with localized data sources — a model that maps directly to multi-jurisdiction vendor portal automation.
The next wave of these case studies — already being seeded by Microsoft's GA reference accounts in insurance and retail — will report time-to-replace-broken-RPA-bot rather than time-to-first-Copilot-conversation.
What to Do About It
For CIOs: Pull your RPA inventory by Friday. Tag the top 10 bots by maintenance cost over the last 12 months. Those are your candidate migrations. Do not commit to a UiPath renewal until you have piloted computer use against the most expensive broken bot on that list. The window to renegotiate enterprise RPA contracts on better terms is now — incumbents know what is coming.
For CFOs: Model two scenarios for the FY27 automation budget. Scenario A: status quo RPA renewal with 8 percent uplift. Scenario B: 30 percent RPA spend redirected into Copilot Credits, with the savings funding the AI governance gap work that compliance has been asking for. Scenario B is achievable for most enterprises within 18 months.
For Risk and Compliance Leaders: Computer-use agents are the highest-blast-radius agent class. They can click anything a human can click. Make Power Platform allow-lists, DLP enforcement, and Purview audit non-negotiable preconditions before any computer-use agent reaches production. The AI governance frameworks you have been drafting for 2026 should treat computer use as a distinct category, not a footnote under "AI agents."
For Business Leaders: Reopen the backlog of automations your team killed for "no API available." That list is now the highest-ROI pipeline in the organization. Most of those workflows will pay back the credit budget in single-digit months.
