Every third party that touches a call, named.
Your callers tell us things they would tell a solicitor. Most vendors answer that with the word “enterprise-grade” and a padlock icon. Below is the actual stack, what each component does, and where it processes data, so your practice manager can assess it rather than take our word for it.
What runs a call.
A voice agent is four things in a chain: hearing, thinking, speaking, and doing something with the result. Each is a separate vendor.
| Component | What it does | Processing location |
|---|---|---|
| Anthropic Claude via AWS Bedrock |
The reasoning layer. Understands the matter, decides what to ask next, writes the summary. | Sydney, Australia |
| Deepgram | Speech to text. Converts what the caller says into words the model can read. | United States |
| ElevenLabs | Text to speech. Produces the voice the caller hears. | United States |
| Vapi | Call orchestration. Connects the telephone line to the components above and holds the call together. | United States |
| Make.com | Workflow automation. Moves the finished summary into your calendar, inbox or practice management system. | United States or EU |
This list is current as at October 2026 and we will update it here when it changes. If your firm requires a signed sub-processor list or a data processing agreement, ask and we will provide one.
Not all of it is onshore, and we will not pretend otherwise.
The reasoning layer, which is the component that actually holds the substance of what a caller tells you, runs on Anthropic Claude hosted in AWS Sydney. That part is in Australia.
The speech, telephony and automation layers are international providers that process offshore. That is the honest position for essentially every voice AI product on the Australian market today, including the ones that do not say so.
Under Australian Privacy Principle 8 a firm disclosing personal information overseas has obligations, and this is information your practice manager needs before signing, not after. If full onshore processing is a hard requirement for your firm, tell us on the first call and we will tell you plainly whether we can meet it yet.
What we commit to
- We name every sub-processor and update this page when the list changes
- Call data is never used to train a model, ours or anyone else’s
- Transcripts and recordings belong to the firm and are exported on request
- Retention periods are set by you, not by us
- One firm’s data never informs another firm’s agent
- You can have everything deleted, and we will confirm it in writing
Disclosure and recording.
Two things every caller is entitled to know, handled the same way on every call we run.
- She says she is AI within the first ten seconds, before any details are taken, and confirms it if asked directly.
- Recording is disclosed at the start of the call where recording is enabled for your firm.
- A caller can decline to be recorded or to speak to an AI. She stops and takes a message or transfers, whichever they prefer.
- A caller can ask for their recording to be deleted and we will action it and confirm in writing.
Disclosure protects your firm, not ours.
An AI that lets a caller believe it is human puts a law firm on the wrong side of its own conduct obligations, and it is the first thing that surfaces if a matter later goes bad.
We have had prospects ask whether she can be made to sound human and deny being AI. The answer is no, and it is not a configuration conversation.
What we can give your practice manager.
Ask for any of these and you will have them before you sign, not after.
- A written sub-processor list
- A data processing agreement
- Our retention and deletion policy
- The exact disclosure wording she uses
- Sample transcripts from your own test calls
- A written statement of what we do not do
Email giulio@frontli.com.au and ask. We would rather answer a hard due diligence question early than have it surface in month three.
Hear her take a call from your own firm.
Send us your firm name and what you want handled. We configure an agent on your matter types and call you back from it, so you are judging the real thing rather than a recording we chose.