Safe use policy
2026-09
This policy states the limits on what Ciliclaw Digital Workforce may do on its own: which actions complete automatically, which must be confirmed first, and what configuration responsibility rests with company administrators.
1. Scope
This policy applies to every company that deploys and uses Ciliclaw Digital Workforce, and to anyone using it through DingTalk or the desktop app.
Companies must translate these requirements into their actual configuration. Consequences arising from configuration that departs from this policy rest with the company.
2. Limits on automatic execution
Read-only actions — retrieving material, reading files, browsing, drafting — complete automatically within the scope of the task, with no step-by-step confirmation.
Actions with real effects — writing or deleting files, running commands and scripts, sending messages externally, calling interfaces that change data — must carry a confirmation rule and may only run once an authorised person confirms.
For approval, HR, security, and organisational matters the agent gathers information, states the basis, and points at the relevant human entry point. It does not decide in place of an authorised person.
3. Responsibility for output
Documents, spreadsheets, code, and conclusions the agent produces are **work awaiting review**, not final decisions. They should be reviewed by a person before use.
Output cites the data and versions it relied on, and marks the parts the model itself was unsure of, so that a review has something to check.
Companies should name a reviewer for each class of output, particularly anything sent externally, involving money, or involving people.
4. Local execution
Local capabilities involving file writes, command execution, scripts, or application control must record their purpose, resource scope, risk level, and owner, and must carry a confirmation rule.
Adding a write capability or widening a resource scope requires the company’s security review.
All local execution leaves a record tied to the employee, agent, task, permission basis, and confirmation.
5. Knowledge must be trustworthy
Only knowledge in the “effective” state may be used in a formal answer. Draft, in review, returned, pending effect, superseded, expired, withdrawn, and archived states are all unusable.
Every formal knowledge item must have an owner, a reviewer, and a validity window. When an owner can no longer serve, the item must be reassigned.
Where conflicting knowledge exists on the same topic, it is kept out of formal answers first and assigned an owner to resolve.
6. Changes
This policy is updated as the product evolves. Material changes are announced through company administrators, and the update date is shown on this page.