Computer-use agents are changing the legacy automation map

One of the most common reasons a workflow stays manual is not that the work is too valuable to automate.

It is that the systems are ugly.

There is an old portal. A vendor dashboard with no clean API. A carrier site. An internal tool that nobody wants to touch. A browser-based process held together by tabs, downloads, and copy-paste.

For a long time, that was a real constraint.

It is becoming less of one.

The market signal changed

The past 18 months produced a more important AI shift than most buyers realize:

• OpenAI introduced Operator on January 23, 2025, describing an agent that can use a browser by clicking, typing, and scrolling through interfaces the way a human does.
• OpenAI followed on March 11, 2025 by releasing agent-building primitives including web search, file search, and computer use in the Responses API.
• Microsoft said on April 23, 2025 that 46% of organizations globally were already using agents to automate workstreams or business processes in its 2025 Work Trend Index.
• Anthropic said on February 25, 2026 that its Sonnet models improved on the OSWorld computer-use benchmark from under 15% in late 2024 to 72.5%.

Taken together, that is a strong signal that interface-level execution is moving from novelty toward practical operations infrastructure.

Not perfect infrastructure. But practical enough that more workflows deserve a fresh look.

Why this matters commercially

Most businesses do not run on clean modern APIs alone.

They run on:

• portals
• spreadsheets
• email threads
• uploaded forms
• partner systems
• internal tools built years ago

That matters because many operations teams still assume those constraints automatically push them back to manual labor.

That assumption is getting weaker.

If an agent can:

• log into the right system
• navigate the interface
• gather information across tabs
• enter structured values
• trigger the next step
• escalate the exception to a human

then a workflow that used to require a person acting as middleware may now be reasonable to automate.

That changes the ROI conversation.

The question is no longer only:

Do these systems have the right APIs?

It is now:

How much of this workflow can be automated with the system access we already have, and where do we still need a cleaner integration?

What buyers should not get wrong

This is not an argument for replacing good integrations with glorified browser bots.

API-first automation is still the better operating model when stable APIs exist.

Why?

Because APIs are usually better for:

• reliability
• auditability
• structured error handling
• long-term maintenance

Computer use changes the map. It does not erase the hierarchy.

The smart pattern is usually:

1. Use APIs wherever they are available and stable.
2. Use computer use for the stubborn surfaces around the workflow.
3. Keep a human path for sensitive or ambiguous cases.
4. Monitor drift aggressively when interface-level steps matter.

That is much different from the old RPA promise that every click path should become the architecture.

Where the new opportunity is strongest

The best candidates are not random desktop tasks.

They are repetitive workflows where the interface problem used to block automation economics:

• insurance verification across payer portals
• logistics status checks across carrier systems
• onboarding tasks that require portal entry and document checks
• compliance and vendor workflows that span uploads, lookups, and approvals
• finance operations tasks that still bounce between inboxes, PDFs, ERP screens, and external systems

These workflows already have:

• clear pain
• clear volume
• known systems
• visible labor cost

That is what makes the new computer-use trend commercially relevant.

Why this helps buyers now

A lot of companies have been waiting for their stack to become cleaner before they automate.

That wait can last forever.

The better move is often to split the problem:

• stabilize the core workflow
• automate what is already reachable
• contain the fragile last-mile steps
• replace those steps with cleaner integrations over time

That approach gets the business value moving sooner.

It also matches how the market is maturing.

The winning buyers are not waiting for perfect autonomy. They are using better agent capabilities to remove the most expensive manual touches first.

The practical takeaway

If your team still says "that workflow cannot be automated because it lives in a portal," that statement is probably outdated.

The better question is:

• Which parts of the workflow are now reachable with computer use?
• Which parts still need API work?
• Where does human review actually belong?
• What is the cost of leaving the workflow manual for another year?

That is a much more useful buying frame than debating AI in the abstract.

Legacy systems are still a constraint. They are just no longer a universal excuse.

Sources

• OpenAI, "Introducing Operator" (January 23, 2025)
• OpenAI, "New tools for building agents" (March 11, 2025)
• Microsoft, "2025 Work Trend Index" summary (April 23, 2025)
• Anthropic, "Anthropic acquires Vercept to advance Claude's computer use capabilities" (February 25, 2026)

If your most painful workflow is trapped between portals, inboxes, and internal systems, book a workflow audit or run the calculator.