Choose what your voice agent can use and where data goes.

Knowledge needs approval before the agent can use it.

Every new source is saved unapproved, whether it’s added in the dashboard, over MCP or through the REST API, and approval is a separate step. Each GitHub re-sync needs approval again. Sales knowledge is kept apart from support, and the sales agent never sees GitHub imports.

Kawl Now tells the agent to treat source text as information and never to follow instructions found in it, and it labels the topic a caller types as untrusted. These are instructions to the model rather than hard controls, so review each source before you approve it and include injection attempts in your tests. Approval controls what the agent can use. It doesn’t guarantee that every answer is correct, so read the transcripts too.

Private GitHub repositories work only after we add them to a server-held allowlist for your workspace, with a read-only token.

Credentials stay on the server.

Provider keys are kept on the server. The embed carries only your public workspace ID, and the widget’s public settings expose display details such as names, greeting, label and colour, never your instructions.

You sign in with a single-use email link that expires after 15 minutes. Dashboard sessions use HttpOnly, Secure cookies that last 7 days and end when you sign out.

Workspace tokens start with hl_sk_ and are shown once. They’re stored as SHA-256 hashes, with only a short prefix kept so you can tell them apart. Each one covers a single workspace and can be revoked at any time, and creating or revoking one emails the owner. Each account has one workspace, with no team seats, roles or single sign-on.

Where can calls start?

Customer calls start switched off. Once they’re on, calls start only from the exact website origins you list, through one-use call tickets that must be claimed within 2 minutes. The check relies on the browser’s Origin header, and there’s no CAPTCHA or bot protection yet.

Each workspace takes up to 3 calls at once and 12 new calls a minute, and every call stops at 10 minutes. The free allowance caps the total at 50 conversations.

Where is conversation data stored and processed?

The application runs on Railway in its EU West region. Data is stored in a Postgres database on Supabase in London, in a private schema with row-level security on every table. The product saves conversation transcripts, feedback, knowledge, usage records, automatic insight reports and any team phone number you verify for handover. It doesn’t record call audio.

During a call, audio passes directly between the caller’s browser and OpenAI, which processes the voice conversation. Supabase stores the data and customer sign-in accounts, Railway hosts the application and Resend sends sign-in and notification emails. GitHub is contacted only when you import documentation from it.

After calls, OpenAI analyses customer speech from recent transcripts to produce automatic insights. If you turn on human handover, an early release, Twilio sends the code that verifies your team’s phone number and carries the handover call between the caller’s browser and that phone, and OpenAI writes and voices the private briefing your team hears.

We don’t claim that all data stays in the UK or EU, because OpenAI’s processing region for the voice conversation isn’t documented.

What is saved, and for how long?

Each conversation keeps a two-sided text transcript, its topic, status, length and any feedback. The summary line in your list is the start of the caller’s own words, not an AI write-up. Very long transcripts keep only their last 100 turns.

There are no retention settings or automatic deletion yet. Transcripts are kept until you ask us to delete them, and account export and deletion are handled on request. The dashboard export downloads your latest 100 conversations as JSON, without transcripts.

Email notices cover the free allowance running low, customer calls or human handover being switched on or off, changes to allowed origins, workspace tokens being created or revoked, and a handover that couldn’t be prepared for a caller. Changes to knowledge, approvals, instructions or voice don’t send an email. Notices never include caller names, transcripts or summaries.

What are callers told?

Before a call starts, the widget shows: “By starting, you agree to microphone use and a saved transcript. You’re speaking with AI. Up to 10 minutes.” The agent also tells callers it’s an AI.

What isn’t in place yet?

Use these gaps to decide what your pilot should cover. Start with general product questions and a small group of callers.

• Security certifications such as SOC 2, ISO 27001, HIPAA or PCI, compliance attestations and an uptime SLA.
• High availability. The app runs as a single instance, with call state held in memory.
• Retention settings, team permissions and identity checks for account-specific support.
• Product terms, a privacy notice for the product and a data processing agreement for pilot workspaces.

Who do you contact about security or data?

Contact hello@tenhaw.com with a security or data processing question, or to request deletion. Website enquiry handling is described separately in the privacy notice.

Common questions