One thread, two axes
A conversation is either internal (between colleagues) or external (between organisations). And it is either a direct message, a group, a channel, a thread with a client, or a meeting. That is all. Every real situation composes from those two axes: the employee asking HR something, the company answering a supplier, the CRM contact, and the storefront visitor who has no account at all.
Participants are of different kinds — user, organisation, contact, guest — and each one carries its own organisation. That is what lets a thread cross the boundary between two companies without either side seeing more than it should.
End-to-end encryption where it actually matters
Internal conversations are end-to-end encrypted with MLS (RFC 9420), the IETF standard for encrypted group chat: forward secrecy, post-compromise security, and native groups without the key explosion of improvised schemes. On by default for 1:1 messages.
External and agentic conversations are not end-to-end encrypted, and the screen says so. The reason is physical rather than commercial: if the server cannot read the text, it cannot translate, summarise, search or spam-filter it either. Payroll, legal and colleague talk get real confidentiality; the prospecting thread gets the intelligent features. It is the same line Slack draws on one side and Signal on the other.
The database enforces the rule, not just the interface: in an encrypted conversation there is nowhere for plaintext to be written.
What happens when someone leaves
This is where E2EE products usually die in B2B: an employee leaves and the company permanently loses its business conversations. There is a key backup encrypted with a recovery passphrase, so you can restore your history on another device, and optional organisational escrow for companies that must archive.
When escrow is on, the interface tells every participant that this is no longer strict E2EE — not only the administrator. A product that promises encryption while quietly holding an access key is false advertising, however good the cryptography underneath.
Assisted outreach, not automated spam
The assistant can find suitable companies and draft proposals. What separates that from spam is not intent but constraints, and all of them are enforced:
- Only claimed organisations with explicit opt-in and a completed availability profile appear in discovery — categories, county, minimum budget, languages, declared SLA.
- Matching runs in three tiers, cheapest first: deterministic filters (industry code, county, size), vector ranking, and only then the model, on the top candidates alone. You can see why a company was proposed.
- A human approves the first message. They may edit the text, but they cannot remove the label saying an agent drafted it.