AI agents that use business software: where to start safely
Recent computer-use advances make new workflows possible. Here is how to design a bounded pilot with approval, audit history and recovery.
What has changed?
Recent model releases have highlighted computer use: the ability to interact with software through a visual interface. In October 2026, OpenAI described GPT-6.1 Terra as supporting computer use, while Anthropic positioned Claude Haiku 5.5 for tasks including browser use. Both are provider claims and should be tested against the actual workflow.
An agent can help gather evidence and prepare a proposed action, but access to a screen does not make it reliable or authorised to make business decisions.
Choose the right integration
When a supported API exists, it can offer clearer permissions, validation and response handling than navigating a changing interface. Computer use can be useful when no suitable integration exists, subject to the application's terms and security requirements.
Keep credentials outside prompts, use least-privilege access and treat external content as untrusted. Validate inputs and outcomes in application code instead of relying on the model alone.
- Scope the agent to one workflow and a small set of permitted actions.
- Show the evidence and the exact proposed change to the reviewer.
- Require approval before consequential external actions.
- Log decisions and support retries without duplicate side effects.
Make failure visible
Expect timeouts, changed screens, missing records and ambiguous instructions. Stop safely when the system cannot verify a precondition, and route the case to a person. Record enough context to investigate while avoiding unnecessary personal data.
BuildrLab's SupportOps is a synthetic demonstration of a bounded workflow: it retrieves policy excerpts, checks account information and sends a helpdesk update only after human approval. It is an example of system design, not evidence of production performance.
Define a useful pilot
Agree the allowed task, data sources, approval rules and success measures. Deliver a working workflow with representative tests, a failure and recovery walkthrough, and a handover that explains what is required before production use.
Have a real workflow in mind?
Explore an AI discovery engagement, a focused RAG pilot or fractional engineering support with BuildrLab.