A Hong Kong logistics group approves an AI automation project in January. By June the pilot is stalled. The model works. The problem is that the freight forwarding system the team actually uses was written in 2004, has no API, and the vendor quoted seven months and a six-figure sum to build one. The AI has nothing to connect to.
That gap, between what AI can reason about and what it can actually touch, is the single most common reason enterprise automation projects die in Hong Kong. A category of tooling that reached production maturity in 2026 attacks it directly.
What is a computer-use agent?
A computer-use agent is an AI system that operates software the way a person does: it looks at the screen, decides what to click, types, scrolls and reads the result. It needs no API and no integration project. If a human can run the process through a browser or desktop application, a computer-use agent can attempt it.
The distinction that matters commercially is the interface. Every other form of enterprise AI integration requires the target system to expose a structured way in. A computer-use agent treats the user interface itself as the integration layer.
That single property changes which systems are reachable. Terminal emulators, Citrix-published applications, government e-filing portals, ageing ERP front ends and SaaS products with thin API coverage all become candidates without a vendor engagement.
Why did computer-use agents become an enterprise question in 2026?
Because they stopped being previews. Microsoft made computer-using agents in Copilot Studio generally available on 13 May 2026, rolling out across all commercial Power Platform geographies. General availability moved the conversation from research demo to procurement item.
The GA build ships with the controls an enterprise buyer asks for in the first meeting. Per Microsoft's Copilot Studio announcement, it includes Azure Key Vault credential storage, Microsoft Purview audit logging, and configurable human-in-the-loop review, with OpenAI's computer-use model and Claude Sonnet 4.5 available as the reasoning engines.
Competing offerings followed a similar arc through 2026. The strategic point for a Hong Kong operations leader is not which vendor wins. It is that a capability previously classified as experimental now carries a support contract, an audit trail and a per-step price.
How does a computer-use agent actually work?
It runs a perception loop. The agent captures the current screen, a vision-capable model interprets what is on it, the agent selects one action, executes it, then captures the screen again to verify the outcome. The cycle repeats until the task completes or a guardrail stops it.
This is why the technology is priced by step rather than by seat. Each loop is a billable unit of reasoning, and a task that takes a person forty clicks costs roughly forty times a task that takes one.
It also explains the reliability profile. Because the agent reasons about what it sees rather than following recorded coordinates, a moved button or a redesigned form is usually survivable. A genuinely ambiguous screen, by contrast, produces a wrong decision rather than a clean error.
The practical consequence for a COO: the failure mode is not a crash you get paged about. It is a plausible-looking action taken in the wrong place, which is why verification design matters more than model choice.
How is a computer-use agent different from RPA?
Traditional robotic process automation follows a recorded script bound to specific selectors or screen positions. A computer-use agent interprets the screen at run time and decides the next action. RPA breaks when the interface changes; the agent adapts but may reason incorrectly.
The economics differ accordingly. RPA carries low per-run cost and high maintenance cost. Industry analyses of long-running RPA programmes consistently attribute the majority of ongoing bot cost to interface changes rather than new business logic.
Computer-use agents invert that. Per-run cost is meaningfully higher because every step consumes model inference, while maintenance drops because there is no selector library to repair after each vendor release.
The correct enterprise conclusion is not replacement. RPA remains the cheaper and more predictable choice for high-volume, stable, structured processes. Agents earn their cost on lower-volume, exception-heavy, frequently-changing work, which is exactly the category most Hong Kong back offices have never automated.
Which processes should you give a computer-use agent first?
Apply four tests. The process must be screen-bound with no viable API, verifiable so a machine can confirm the outcome, reversible so a mistake can be undone, and bounded so a single run has a defined end state. A process failing any one of the four is a poor first candidate.
The screen-bound test protects the business case. If a clean API exists, an ordinary integration will be cheaper and more reliable, and choosing an agent instead is paying a premium for nothing.
The verifiable test is the one most organisations skip. A supplier invoice keyed into a legacy accounting front end is verifiable, because the posted record can be read back and compared. A judgement email sent to a client is not.
The reversible and bounded tests define blast radius. Rekeying data into a staging queue is reversible. Submitting a statutory filing is not, and belongs behind a human approval gate rather than inside the agent's autonomy.
Scored this way, the strongest first candidates in Hong Kong tend to be the same four: rekeying between an ageing operational system and a modern finance platform, extracting statements from bank or carrier portals, licence and permit status checks on government sites, and reconciliation across two systems that were never integrated.
How much does a computer-use agent cost to run?
Pricing is consumption-based and measured in steps, not seats. In Copilot Studio, computer use is billed at the agent action rate, and published analyses of the credit model put a single step in the region of US$0.04 on standard prepaid pricing, with prepaid credit commitments cutting the effective rate.
That number is only useful with a step count attached. A forty-step reconciliation run at that rate costs roughly US$1.60 in inference. Run twice daily across a year, it is a low-four-figure annual line item.
The cost that surprises finance teams is not the successful run. It is the retry. An agent that fails at step thirty-five and restarts consumes the first thirty-five steps again, so unreliable processes cost multiples of their nominal price.
Budget accordingly: model the cost per successful completion, not the cost per step, and demand a measured success rate from any pilot before committing to volume.
What goes wrong when enterprises deploy computer-use agents?
Five failure patterns account for most stalled projects. Choosing a process with no machine-checkable outcome. Granting a shared human login. Skipping a shadow-mode period. Measuring activity instead of completion. And treating every agent as though it needs identical governance.
The credential mistake is the most damaging and the most common. Handing an agent a staff member's account destroys attribution, breaks segregation of duties, and makes any subsequent audit unanswerable. Agents need their own identity, scoped to the minimum set of screens the task requires.
The governance mistake is more subtle. Gartner warned in May 2026 that applying uniform governance across all AI agents will itself cause enterprise agent failure, and that failures cluster where organisations do not separate an agent's ability to act from the scope of access it has been granted.
The cost of getting this wrong is now quantified at portfolio level. Gartner's February 2026 analysis projects that more than 40% of agentic AI projects will be cancelled by the end of 2027, attributing the cancellations to operationalisation, governance and unproven business value rather than model capability.
What controls does a computer-use agent need before production?
Six, and they are non-negotiable in a regulated Hong Kong environment. A dedicated non-human identity. Scoped access. Full action-level audit logging. A human approval gate on irreversible steps. A defined stop condition. And a named owner accountable for the agent's output.
--- Dedicated identity. The agent authenticates as itself, never as an employee, so every action is attributable.
--- Scoped access. Permissions cover only the screens the task requires, which caps the damage from a misread screen.
--- Action-level audit log. Every click, keystroke and data field read is recorded, because a screenshot trail is what a PDPO enquiry or an internal audit will ask for.
--- Human approval gate. Payment, submission and deletion steps pause for a person, regardless of how well the agent has performed.
--- Stop condition. The agent halts after a defined number of retries or on any unrecognised screen rather than improvising.
--- Named owner. A person, not a committee, signs off on the agent's scope and reviews its exception log weekly.
Hong Kong adds a specific consideration. Where a screen-bound process touches personal data, the agent is reading and processing that data on your behalf, and the Privacy Commissioner's guidance on AI in the workplace applies to it exactly as it applies to a member of staff. Design the audit trail on the assumption you will one day have to produce it.
How should you start in the next 30 days?
Run a four-week structured evaluation rather than a procurement exercise. Week one, inventory screen-bound processes and score them against the four tests. Week two, pick one and instrument it. Weeks three and four, run the agent in shadow mode and measure.
Shadow mode is the discipline that separates the projects that survive from the ones cancelled at month six. The agent performs the task with no write access while a person does the same work, and the two outputs are compared daily.
The single metric that matters is completion rate without human correction. An agent completing 70% of runs cleanly is a genuine saving. An agent completing 95% of steps but requiring review on every run has moved the work rather than removed it.
Set the go or no-go threshold before the pilot starts, in writing, with the finance sponsor in the room. According to Deloitte's State of Generative AI in the Enterprise research published in early 2026, vendor-supplied agents reach first value in around 38 days against roughly 94 days for in-house builds, so a four-week evaluation is a realistic window rather than an optimistic one.
For related decision frameworks, see UD's guides on how enterprises measure whether AI actually works, non-human identity for AI agents, and build versus buy for enterprise AI.
The strategic takeaway
Computer-use agents do not make your legacy estate modern. They make it reachable, which for most Hong Kong enterprises is the constraint that has actually been blocking automation for a decade.
The organisations that will get value from this in 2026 are not the ones with the largest budget. They are the ones that pick verifiable, reversible processes, give each agent its own identity, and measure completion rather than activity.
That is unglamorous work, and it is where the returns are. We understand AI. We understand you. With UD by your side, AI never feels cold.
Reviewed by the UD enterprise AI team.
Ready to find your first automation candidate?
Knowing the framework is one thing. Finding the process in your own operation that passes all four tests is another. UD has spent 28 years inside Hong Kong enterprise systems, and we'll walk you through every step, from readiness assessment and process scoring to deployment, controls and measurement.
