AI transparency
Our agents always say they’re AI.
Two promises. You and your customers will always know when you’re dealing with one of our agents. And our agents watch for the AI-made fakes aimed at your business.
Promise 1
How agents identify themselves
- On the phone: every call opens by saying it’s an AI assistant, and offers a person at any time.
- In email and chat: agent-written messages are labelled, with the approving person named where approval was needed.
- In documents: drafts are marked AI-drafted until a person signs off, with Content Credentials where the format allows.
- Never impersonating: agents never pose as a real employee, and we don’t clone staff voices.
Promise 2
Fakes we watch for
- Cloned-voice calls asking for urgent payments. Agents never act on payment instructions given by phone.
- Forged invoices and bank-change letters, compared against each supplier’s own history.
- Synthetic identities in onboarding documents, sent to manual checks when in doubt.
- Prompt injection: hidden instructions in emails and PDFs are logged as attacks, never followed.
Detection is a signal, not a verdict. High-stakes decisions always go to a person.
Agent cards
What each agent may do, and where a person steps in. A full card with exam results from your own data ships with every install.
| Agent | Can do | Never does | Person steps in |
|---|---|---|---|
| Scout | Monitor sources, score tenders and risks, draft briefs | Submit bids or contact suppliers | Every go/no-go |
| Herald | Answer calls and tickets, book, refund within limits | Give medical, legal or financial advice | Urgent calls, refunds over the limit |
| Ledger | Read, match and code bills, stage payment runs | Release payments or post journals | Every payment run |
| Quill | Answer from documents, mark up contracts, draft policy | Sign, send or publish legal documents | All external documents |
| Atlas | Schedule, forecast, set up accounts by role | Grant admin access or exceed a PO budget | Access above role, POs over budget |
| Patch | Propose fixes, run read-only diagnostics, triage alerts | Change production without approval | Every deploy and rollback |
Mapped to Australia’s Voluntary AI Safety Standard
How each of the standard’s ten guardrails shows up in an install.
- Accountability. A named owner on both sides for every floor.
- Risk management. A risk register from the first survey, reviewed monthly.
- Data and security. Onshore hosting and scoped keys, set out in the Trust Centre.
- Testing. An exam on your real history, then drift monitoring.
- Human control. Approval checkpoints and a pause switch on every floor.
- Informing users. The disclosure rules at the top of this page.
- Contestability. Anyone affected by an agent decision can ask for a human review.
- Supply chain. Models and subprocessors disclosed, with notice of changes.
- Records. Decision logs kept for the period agreed with you.
- Stakeholders. Staff briefings before go-live and feedback built into every floor.
Think an agent got it wrong?
Customers and staff can contest any agent decision. We review within five business days and publish a yearly summary of what we learned.
hello@teammate.autos