Simsim
Book a pilot
Security & data flow

What leaves the room, where it goes, and what is kept.

This page is written for the security review. It describes the v1 architecture as built. Where something is a roadmap item rather than shipped, it says so.

YOUR MEETINGZoom / Teams / Meetthe bot joins as a participantMEETING BOT SERVICECapture + speech-to-textUS, EU or Tokyo · no video capturedaudioYOUR TENANT (cloud or in-country)Core agentwake word · routing · answer guard · audit log · permission modeltranscript, speaker namesvoice + visual cardClaude APIreasoning · no training on dataElevenLabstext → speech onlyAudit databasequestions, results, latencyquestion + tool resultsanswer textYOUR SYSTEMS (read-only service account)SalesforceMCPJiraMCPWarehouseMCPYour agentsMCP

Data handling

Meeting audio and video
Audio is processed in the meeting-bot service (US, EU or Tokyo region, chosen per deployment) for live speech-to-text. Video is never captured. The meeting-bot service keeps the transcript and meeting metadata for a configurable window, 7 days by default or zero retention per client. Self-hosted, in-country capture is available for clients with residency requirements.
Transcript
A rolling window is held in memory during the meeting for context. Utterances are written to your audit database and purged after the retention period (default 30 days). Simsim never transcribes its own speech into the record as a human’s.
Questions and answers
Every interaction is stored: speaker, heard text, cleaned question, resolved query in plain English, raw source result (truncated, purged on schedule), spoken text, chat text, latency per hop, review flags.
Reasoning
Questions, the relevant transcript window and tool results are sent to Anthropic’s Claude API under its commercial terms (no training on your data). A second Claude route (AWS Bedrock or Google Vertex) can be configured for failover and regional preference.
Voice
Only the text to be spoken is sent to ElevenLabs. Audio returns to the bot; nothing is retained by us.
Your systems
Queried through MCP connectors using a read-only service account your IT creates with the agreed scope. Credentials live in the client’s environment file on your infrastructure, never in code, never in our repository.
Write actions
Disabled by default. When enabled per client: only tools explicitly declared as mutating can run, only after a spoken plan is confirmed aloud by the requester, with a full success/failure report and an undo window. Silence cancels.

Permission model

v1 · shipped
Pre-approved dataset

Your team decides, per client, which sources may be spoken aloud. Simsim can answer only from those. Sources can be marked public for rooms with guests.

roadmap · with SSO
Asker’s rights

Answer only from what the person asking is entitled to see, resolved through your identity provider and each source’s own permissions.

roadmap · with SSO
Lowest rights in the room

Answer only from the intersection of everyone’s entitlements, recomputed as people join and leave.

External participants

On join, Simsim compares each participant’s email domain with the client’s internal domains. Anyone not provably internal, including participants whose email the platform does not expose, counts as external. Policy per client: refuse to answer for the meeting, answer only from public sources, or allow. The default is public-only, announced once in the chat, so the room knows.

Accuracy controls

  • A spoken number must be traceable to a query result. This is enforced in code after the answer is composed; an untraceable figure is never spoken and the raw data is posted to the chat instead, with the event logged.
  • Derived figures (growth, differences, ratios) are computed by a deterministic calculator tool, so they are traceable too.
  • Ambiguous business terms trigger exactly one clarifying question. Definitions are confirmed by a named person and are readable in the console.
  • “No data” and “I don’t know” are different answers and are never conflated.
  • Anyone in the room can say “that’s wrong”. The answer is flagged loudly and appears in the review queue with its full chain.

Operational controls

  • Hard caps per client on session length, monthly bot-minutes and monthly spend. The bot leaves empty rooms and times out of waiting rooms.
  • Console access is token-protected and not linked from this site. Enterprise SSO (SAML) is available on request.
  • Meeting-bot callbacks are signed; per-meeting URLs are unguessable and expire with the session.
Need a questionnaire filled in?

Send us your vendor security questionnaire with the pilot request and we will return it before the first call.

Book a pilot