Compare responsibilities before labels
Robotic process automation, or RPA, usually describes a broader automation practice and product category. Computer use describes a capability an automation system can contain. Microsoft documents desktop flows for automating repetitive desktop and web work. That scope overlaps with tasks an AI interface agent might attempt, but overlap does not make the systems interchangeable.
For a buyer, the first task is to identify responsibilities. Who schedules work, owns credentials, handles queues, reviews exceptions, deploys changes, and proves a business result? A computer-use model API answers only part of that list. A mature existing automation estate may already supply several of these functions.
Find the variability in the process
Write down which parts are rules and which parts require interpretation. Matching an invoice number is different from deciding whether two differently named attachments describe the same service period. The former may have a stable deterministic rule; the latter needs either a better data contract, a person, or carefully evaluated model assistance.
Our editorial preference is to preserve explicit business rules around balances, approvals, record identity, and accounting periods. Let an agent interpret a bounded interface problem only when that interpretation can be checked. Do not move a rule into a prompt merely because prompts are quicker to write than workflow code.
Add an agent as a controlled exception handler
A useful pilot can put computer use inside an existing queue: a job receives one permitted item, opens the relevant interface, prepares a result, and returns an artifact for validation. The queue still owns retries and completion. The controller still owns which actions are allowed. The agent does not silently gain access to adjacent processes.
Treat uncertainty as a first-class result. An unfamiliar popup, missing record, or conflicting amount should return an exception with context. Automatically retrying a consequential submission can create duplicates. Decide which steps are idempotent and which require a read-back before another attempt.
Budget for operating the process
License or API price alone is a weak comparison. Include environment maintenance, model calls, session time, exception review, and the engineering work needed to keep integrations healthy. Compare cost per accepted output at an agreed quality threshold. Include the cost of rejected and abandoned runs.
A replacement decision should also account for the value of existing monitoring and runbooks. Start with one workflow and a reversible rollout. This guide does not claim a general cost saving for AI agents or imply that RPA cannot incorporate AI. The decision is about a specific process and the controls already available.