Key takeaways
- Demo tenants only: the agent never touches your production data.
- Credentials are sealed and typed by the browser service; the AI never sees them.
- Every action is checked in code against your allowlist and blocked list before it runs.
- Every action, tool call and model step is logged against the session.
The demo environment guardrails you set
| Guardrail | What it does |
|---|---|
| Demo tenant only | The agent logs into a dedicated demo account with sample data, never production |
| Sealed credentials | Stored encrypted, released once per session to the browser service, never sent to the model |
| URL allowlist | The browser cannot open hosts outside the list you give |
| Blocked actions | Buttons and actions such as delete, billing, settings, invites and exports are refused |
| Read-only mode | Forms can be shown without being submitted |
| Action log | Each click, tool call and model step is saved with the session |
Why the checks live in code, not in the prompt
A prompt that says "never click delete" is a request, and a clever sentence from a prospect or a line hidden in a web page can talk a model out of it. TalkDemo checks every browser action in code before it runs, against the rules you configured. If the action is not allowed, it does not happen, whatever the model asked for.
The same thinking applies to what the model reads. Prospect speech, knowledge passages and screen summaries are passed to it as fenced data, not as instructions.
Setting up a good demo tenant
- Create a separate account with realistic, fictional sample data. No real customer records.
- Give the demo user only the permissions a buyer needs to see.
- List the hosts the demo needs (app, docs, maybe a status page) and nothing else.
- Resetting demo data after each session is coming soon; until then, choose flows that do not leave a mess.