Key takeaways
- Never demo in production, and never put real customer data in the demo tenant.
- Keep the demo login in a vault, out of prompts, docs and flow steps.
- Allow only the URLs the demo needs, and block destructive actions by name.
- Review refused actions weekly; each one is a flow or a rule to fix.
The guardrail checklist for demo tenants
GUARDRAIL CHECKLIST FOR A DEMO TENANT The tenant [ ] A separate account or workspace used only for demos, never production [ ] Realistic sample data, no real customer names, emails or phone numbers [ ] A dedicated demo user with the smallest role that can show every flow [ ] A reset plan: how and how often the sample data goes back to a clean state Credentials [ ] Demo login stored in a vault, not in a prompt, a doc or a flow step [ ] Two-factor sign-in off for the demo user, or handled by the browser service [ ] Password rotated after anyone with access leaves the team Where the agent may go [ ] URL allowlist: only the hosts and paths the demo needs [ ] Links to billing, account settings, user invites and data exports blocked What the agent may click [ ] Blocked buttons and actions: delete, remove, cancel subscription, invite, export, send email, change password [ ] Read-only mode for forms if saved changes would confuse the next demo [ ] File uploads off unless a flow needs them Watching it [ ] Every action and refused action logged with the session [ ] Someone reviews refused actions weekly and fixes the flow or the rule [ ] Flows re-checked after each product release
Why an AI agent needs stricter rules than a person
A person giving a demo knows not to click Delete account. An AI agent only knows what it is told, and a prompt is not a control: a buyer can ask it to do something odd, or text on a screen can look like an instruction. Limits that matter should be checked in code before each click, with the agent's instructions as a second layer, not the only one.
Use it in TalkDemo
- Demo credentials are sealed and typed into the login form by the browser service; the AI never sees them.
- The agent's Guardrails tab holds the URL allowlist, blocked selectors and actions, and read-only mode.
- Every browser action passes a guardrail check in code before it runs, and refused actions are logged on the session timeline.
- Flow health marks a flow as broken after repeated failures at the same step, so release changes surface quickly.