Software too old for an API,
operated through its own screen instead
Some of the software a business depends on most is also the hardest to automate: an old ERP, an accounting package from a decade ago, a factory system with no documented API and no vendor left to ask. A computer-use agent for legacy desktops operates that application's own screen the way a person would, reading what is displayed and clicking and typing in response, inside a scope tested and limited before it ever runs unattended.
The role today
A business built on software from a decade or more ago usually keeps it because replacing it is expensive and risky, not because anyone loves it. That software rarely has an API, sometimes has no vendor left to ask, and the people who understand its quirks best are the ones who have been clicking through it manually for years, repeating the same screens for the same routine tasks.
The second cost is that this manual work does not scale and does not transfer well. When the person who knows the system’s quirks is out, the task either waits or gets done by someone less familiar, more slowly and with more mistakes.
The third is that attempts to automate this kind of system with generic scripting tools often break on small, unpredictable UI quirks that a legacy application is full of, which is part of why this work usually stays manual long after everything else has been automated.
What the agent takes over
The agent operates the legacy application through its own screen: reading what is displayed, clicking, and typing, the same inputs a person would use, inside a script built and tested against the real application first, not a simulation. It runs on a scoped, isolated machine so it never has broader access to other systems than the specific task requires.
When a screen does not match what the script expects, which happens more often with older software, it stops rather than proceeding on a guess, and the run is flagged for a human to review. Every action is screen-recorded, which matters for both debugging and audit in systems this close to a company’s financial or inventory records.
Typical scope: data entry, report generation, routine transactions in systems with no other way in. We recommend starting with the lowest-risk, most repetitive task first and expanding scope only once it has run reliably for a few weeks.
What stays with humans
Anything the application cannot do reliably through its screen, or any step genuinely ambiguous to interpret visually, routes to a human. Credentials and any login step stay under your team’s control. Expanding the agent’s scope to a new task is always a deliberate, tested decision, never an automatic assumption that success on one task means success on another.
Guards
The agent runs on a scoped, isolated machine or VM with no broader access than the task needs. Every action is screen-recorded for a full audit trail. A screen that does not match the expected state halts the run rather than letting the agent proceed uncertainly. Rate limits match what the legacy system can actually handle without strain.
Price and timeline
| Option | Price | What it covers | Timeline |
|---|---|---|---|
| Agency runs it | from $2,500 + support plan | Agent scripted, tested and supervised by us, monthly run review | 3 to 4 weeks |
| Full control, handover-ready | from $4,200 | Same agent on your own scoped machine, documented script, your team runs it | 4 to 6 weeks |
Running cost is usually $20 to $70 a month in compute and model usage, depending on task frequency.
Related
See the AI agents service page and development for the surrounding build. Within this group: browser automation agent applies the same approach to websites, and legacy migration agent and backup and recovery agent cover what often comes before or after this kind of agent. For a related one-time setup, see automate a computer-use agent for legacy systems. Real governed-agent discipline behind this page: the ProBay AI agent team case study.
An old system that still runs the business but has no API in sight? Get in touch and we will look at the actual screens first.
FAQ
How much does a computer-use agent for legacy desktops cost?
From $2,500 for one application and one defined task, live in 3 to 4 weeks. A more complex multi-screen workflow or a second application usually runs $4,000 to $6,500.
How long before it is running the real task?
3 to 4 weeks: scripting against the actual application and testing every screen it needs to recognize takes most of it, since legacy software often has quirks that only show up with real use.
Which kind of software does this work with?
Desktop applications with no usable API: older ERPs, factory or warehouse systems, accounting packages, anything that still requires a person to click through screens. If the software runs and can be viewed on a screen, it is usually a candidate.
What happens if a screen looks different than expected?
The agent detects the mismatch and stops rather than clicking on a guess, which matters more here than almost anywhere, since a wrong click in old financial or inventory software can be expensive to undo. The run gets flagged for a human.
Is this safe to run against a system this critical?
It runs on a scoped, isolated machine or VM, never with broader access than the specific task needs, and every action is screen-recorded for full audit. We also recommend starting with the lowest-risk task first and expanding only once it is proven.