The Systems That Never Got an API
Almost every company still runs on at least one system with no integration interface. A supplier's ordering portal. A government site that only lets you look things up after logging in. An internal screen built twenty years ago whose source code and original developer are both long gone. Someone ends up sitting in front of those screens every day, opening the same pages and retyping the same values.
By 2026, browser and computer-use agents have settled into their own category of automation tooling. Where traditional automation repeats a fixed path, this class of tool perceives the screen through images and accessibility trees and works out its own path toward a goal. The major RPA vendors repositioning their suites around agentic process automation are part of the same shift.
How It Differs From RPA
The most important difference is how each one breaks.
Sorted by the nature of the work, the boundary is clear. Deterministic repetition, where identical input must yield identical output, favors RPA. Work full of inconsistent formats and exception handling favors agents.
The cost structures differ too. RPA loads cost onto initial build and maintenance effort. For a site that redesigns every quarter, annual maintenance routinely exceeds the original build. Agents load cost onto tokens and runtime per transaction. If one case takes 30–60 seconds and dozens of screen-level judgments, per-case cost can reach a few hundred won. Two hundred cases a month is manageable; two hundred thousand is a completely different calculation.
Where to Start
Good fits
Poor fits
And if the system might eventually get an API, design the agent as a temporary bridge from day one. Bake business logic into the agent's prompt and you rebuild everything when the API arrives. Keep the logic in your own system and give the agent only an adapter role: operating the screen.
Design Permissions and Audit Logs First
The most common shortcut under deadline pressure is handing the agent a staff member's own account. Do that and the audit log can no longer distinguish human actions from automated ones, which makes root-cause analysis impossible when something goes wrong.
Operating on the Assumption of Failure
Agents will stop. Redesigned screens, unexpected pop-ups, CAPTCHAs, and expired sessions are the usual causes.
A Rollout Sequence
Starting with one task beats an ambitious company-wide plan.
Drawing on our experience building AI-integrated systems, POLYGLOTSOFT recommends placing a validation layer between the agent and your internal systems. Values the agent reads off a screen are not written straight through; they pass business-rule checks and a human confirmation step first. If you are weighing automation for legacy screens, we can work through it with you — from choosing the right first task to designing permissions and audit trails. Get in touch anytime.