How AR Labs demos use data, and their limits

This page states, as fact, what happens to what a visitor says or types in AR Labs' public demonstrations. Today that means one demo, Salesbot, a browser voice agent that runs a short sales discovery conversation. It describes the code of the current release as of 13 November 2026. It is not a privacy policy: it makes no promise about how long anything is kept, how it is secured, or how it can be deleted, and a formal privacy notice is not yet published. The demo page does not currently show a data notice before a call starts.

What the demo collects and stores

Salesbot runs on a single rented server that the lab operates. While a call runs, the server receives the visitor's speech (or typed lines), and it writes the conversation to four places on that server.

Where	What is written	When
A database	The answers the agent recorded for each stage, in its own short wording	As soon as a stage is complete
The same database	The full record: the recorded answers, a transcript of both sides of the conversation, and how the call ended	When the agent closes the call. A call the visitor leaves part-way is stored up to its last completed stage
A conversation log file	Everything in the conversation, including the agent's own instructions	Every five seconds during the call, and once more when the visitor disconnects
A per-turn trace file	For each turn, which stage the call was in and the answers recorded so far	Once per turn

The server's diagnostic log, written by the voice framework the demo is built on, can also contain conversation text: the agent's replies, the conversation as sent to the language model, and the answers the model recorded.

On the closing screen a visitor may choose to leave an email address and tick features they are interested in. If they do, the server stores the address, the ticks and the time against that conversation, and emails the lab one message containing the address (as the reply-to), the ticked features and a summary rebuilt from the answers already stored. That is the only email the demo sends today. The demo does not ask for a name or a company, and the code does not turn on call recording at the call-hosting service.

Outside services that process the conversation

The conversation passes through rented services. Each is named here as the code calls it.

flowchart LR
  V["Visitor's browser"] <-->|"audio"| D["Daily call"]
  D <--> S["Lab's server"]
  S -->|"audio"| DG["Deepgram"]
  S -->|"text"| A["Anthropic"]
  S -->|"reply text"| C["Cartesia"]
  S --> DB["Database and log files"]
  S -.->|"only if an address is left"| R["Resend email to the lab"]

Service	Category	What it receives
Daily	Real-time call hosting (WebRTC)	The audio in both directions, and the messages that update the page during the call
Deepgram	Speech recognition	The visitor's audio
Anthropic (Claude Haiku 4.5)	Language model	The conversation as text, at each turn
Cartesia	Speech synthesis	The text of the agent's replies
Resend	Email delivery	The follow-up email described above, only when a visitor leaves an address
Google Fonts	Web fonts	The page's request for its fonts, as with any site that uses them

The code can be set to use OpenAI's gpt-4o-mini as the language model instead of Claude; Claude is the default.

What the demo is, and is not

Salesbot is a working demonstration of structured voice conversation, open to the public so that people can try it. It is not a production service, and it is not offered to handle anyone's business data. It has no telephone numbers, call transfer, CRM or calendar connection. What has and has not been verified about how it behaves is set out in testing a browser voice agent.

The principles behind it

Two rules from our wider agent work apply here. A person decides at each consequential step: outside the conversation itself, the demo's only action is the one follow-up email, which goes to a person at the lab rather than triggering anything further (see human-in-the-loop gates). And nothing is claimed that the evidence does not show, which is why the closing summary is built by code from the visitor's recorded answers rather than written freely by the model, and why this page says plainly what is stored and where.