Discover repetitive work and teach governed automations through chat and screen sharing
Caddi is a workflow automation platform for repetitive back-office work. It connects to existing business tools, helps identify repeated procedures and lets a user teach an automation through conversation and screen sharing. The resulting automations are called loops. This product is distinct from Caddy, the separate personal assistant that works through messages.
Caddi Discovery reads connected activity to identify recurring tasks and estimate where automation may be useful. The official description emphasizes read-only discovery: identifying a pattern is separate from receiving permission to change records. Examples include filing executed agreements, preparing time entries and triaging an inbox. The estimated opportunity should be checked against the team's actual process rather than treated as a guaranteed saving.
A user can demonstrate the work through a screen recording while explaining rules and exceptions. Caddi asks about edge cases and uses the resulting instructions to build the loop. Its stated approach combines AI for judgment-oriented steps with deterministic code for operations that need to follow a fixed path. That distinction is intended to make the process more inspectable than an agent improvising every action.
The platform records runs and the permissions used, and it describes proposing fixes when a connected workflow changes. Process changes require human approval. Pausing a loop or disconnecting a tool provides a way to stop work. Firms should still test exception handling and limit each integration to the access actually required, particularly when documents, payments or client records are involved.
The free Individual plan includes one user, a credit allowance and a limited number of active loops. Paid plans expand capacity, with additional credits and loops billed separately. Advanced integrations and governance capabilities vary by tier. A sensible first workflow is small, frequent and easy to verify. Review its output and run history over several cycles before moving to more consequential work, and confirm the appropriate data-processing terms for sensitive information.
