BankGPT — computer-use automation

Give an AI agent hands inside apps with no API.

Discover a back-office UI flow once with an LLM, distill it into a typed capability artifact, then replay it deterministically — with human approval for risky actions and live-session handoff when automation gets stuck.

01 — Discovery

The model drives once

An LLM drives a real browser through a back-office flow — observing the accessibility tree, one structured model call per step — until it reaches the goal.

02 — Artifact

The run becomes a contract

The successful run is distilled into a typed, versioned capability artifact: inputs, outputs, locator fallbacks, checkpoints, expected business outcomes. A human reviews it before risky use.

03 — Replay

Zero model calls

A calling agent invokes the capability with typed inputs and gets back typed outputs, a known business outcome, or a debuggable failure — deterministically.