Safe by design, not by prompt.
An agent that talks to your customers needs more than good intentions. Here is how Bizacqui protects your data, your leads and your reputation.
Workspace isolation in the database
Every table that holds your data carries your workspace id and is protected by Postgres row-level security. The application connects with a role that cannot bypass it, and the workspace is set from your signed-in session or the job that is running, never from an AI model. Our automated tests prove one workspace cannot read or write another’s data.
Envelope encryption for credentials
Channel and payment credentials are sealed with AES-256-GCM using a fresh data key per secret, wrapped by a master key held outside the database. Each ciphertext is bound to its workspace, so a copied blob is useless elsewhere. Secrets are never logged and never shown to a model.
One workspace per agent run
Each agent run happens in a fresh temporary folder containing only that workspace’s knowledge and context. Two businesses are never mixed in one prompt, and the folder is deleted when the run ends.
Everything audited
Every action an agent proposes, and whether it was sent, held for approval or blocked by a rule, goes into your audit log. Platform staff access to a workspace is logged on our side and shown in yours.
Kill switches
Pause your workspace in one click; platform operators can pause everything globally. Incoming messages are still recorded while paused. Workspaces with unusual bounce or complaint rates are paused automatically.
Two-factor authentication
Time-based one-time codes with recovery codes. Required for workspace owners and platform administrators.
Your data, portable
Export everything in your workspace as a zip of JSON files whenever you like. Delete your workspace and it is permanently erased after a 14-day grace period.
Operational hygiene
TLS everywhere with automatic certificate renewal, structured logs without message bodies or secrets, nightly encrypted database backups with restore tests, health checks and alerting.
Enforced in code, checked before every send.
These are not instructions we hope the AI follows. They are checks in the platform that run before anything leaves, with unit tests behind each one.
WhatsApp is inbound-only
The agent can only reply to people who messaged your number first. Everyone else gets a wa.me invite link.
No cold email from your main domain
Your verified domain is only used for people who came inbound, replied or opted in. Cold outreach runs on a separate sender, or not at all.
Prices stay inside your bands
Every number in a draft is checked against your offers. Anything outside the band, or any price on a quote-on-request offer, goes to a human.
No unapproved commitments
Discounts, guarantees, payment plans, legal terms and delivery dates beyond your templates are caught in code and handed off.
Compliance by default
Opt-out on every cold message, honoured instantly across every channel. Daily send caps. Business contacts only for cold outreach.
Sector safety
Blocked sectors cannot onboard. Restricted sectors get extra handoff triggers and are held in approval mode for at least 30 days.
The AI decides, the platform acts
Agents return a structured decision. The orchestrator validates it against your rules, limits and locks before anything happens, and logs both.
Kill switches
Pause your workspace in one click. Inbound is still recorded while paused, so nothing is lost.
Human threads are untouchable
Once a conversation is handed to your team, the agent cannot write to it until someone returns it.
Facts only from your knowledge
If the answer is not in your knowledge pack, the agent says so internally and hands off. It never makes things up.
Choose the rules for your market.
Each workspace picks its outreach regime. Opt-outs are honoured instantly across every channel, backed by your suppression list and a platform-wide one.
Honest about what it is, careful about what it says.
- Facts come only from your knowledge pack. Unknown questions go to a person.
- If someone sincerely asks whether they are talking to an AI, the agent does not deny it and hands off.
- No medical, legal, financial or investment advice, and no outcome claims, in any sector.
- Low confidence means a human reply, not a guess. You set the threshold.
- Human edits to drafts become examples, so the agent learns your style instead of drifting.
- Role prompts are versioned and can be rolled back. Changes are tested on a small group first.
Security question or a vulnerability to report? Email security@bizacqui.tutapis.com.
Put your acquisition on autopilot, safely.
Start in shadow mode. Watch what the agent would do on your real leads before a single message goes out.
No card required for the trial. Cancel any time.