Control and Proof are the layer; Voice is the application on it.
Feature list
Everything the platform does, with the status of each line.
Voice: In controlled testing
Audact is control and proof infrastructure for consequential AI. Voice is the application on sale: inbound and outbound voice, a managed speech stack and knowledge-grounded answers, with Article 50 disclosure at the start of governed interactions, a cryptographic evidence chain underneath it, and brand governance on every candidate answer. This page lists both, with the status of each line rather than a checkmark.
The conversational platform
What the agent can do.
This is the part every modern voice platform has, and it is the foundation. The layer below it is what makes it governed.
- Capability
- Inbound voice
- What it means
- Answer every call with a real-time voice agent, grounded in your knowledge base rather than reading a static script.
- Status
- In controlled testing
- Capability
- Outbound voice
- What it means
- Run outbound campaigns with eligibility checks before the first ring, so the call happens only when it is allowed to.
- Status
- In controlled testing
- Capability
- Managed speech stack
- What it means
- Speech-to-text, the language model and text-to-speech are run for you. No API keys to juggle and no vendors to stitch together.
- Status
- In controlled testing
- Capability
- Knowledge-grounded answers
- What it means
- Answers are grounded in your own knowledge base rather than read from a static script.
- Status
- In controlled testing
- Capability
- Recording and transcript in the interface
- What it means
- Call recording and transcript downloads in the dashboard. It is not in the product today.
- Status
- Planned
- Capability
- Multilingual within the EU
- What it means
- The agent converses in the caller's European language, with the disclosure in that language.
- Status
- In controlled testing
- Capability
- Custom tool-calling
- What it means
- Your agent calls your own tools and APIs mid-interaction; you add the tools yourself.
- Status
- In controlled testing
- Capability
- Webhooks with Retell lifecycle event names
- What it means
- Interaction-lifecycle and compliance events, with retries and an authentication code, to an endpoint you provide.
- Status
- In development
- Capability
- In-browser test call
- What it means
- Talk to your agent over WebRTC from the dashboard, before anything goes onto the phone network. It runs inside the Governed Voice AI environment; the address is sent with the kick-off e-mail. There is no route to it from this website.
- Status
- In controlled testing
- Capability
- Chat (web, WhatsApp, email)
- What it means
- The same agent handling text channels under one set of governance and disclosure rules.
- Status
- Planned
Article 50 compliance
Audible disclosure as a runtime, not a checkbox.
Article 50 of the EU AI Act has applied since 2 August 2026. Where the disclosure gate is enabled for that jurisdiction, the disclosure is delivered at the boundary before the caller hears or reads a word. It is enabled by default; a tenant can switch it off.
- Audible disclosure at the start of the interaction — never only a notice on a website.
- Caller-language detection, with a phrasing library per language.
- A logged event per interaction, hashed into the evidence chain.
- A policy engine per country, because the national layer on top of Article 50 differs per country.
Coverage
One disclosure runtime, a national layer per country, checked against the national text.
The disclosure runtime is one piece of software and it works everywhere. The national layer on top of it is not: it is checked by a person against the national text per country, and a country is only switched on once that layer has been checked. That is slower than a map with 27 flags on it, and it is the reason the flags are not there.
- Country
- Germany
- National policy layer
- Country
- Netherlands
- National policy layer
- In development
- Country
- Belgium
- National policy layer
- Planned
- Country
- Austria
- National policy layer
- Planned
Three things differ per country, and they are separate questions: commercial intake, national policy readiness, production go-live. Netherlands: intake open · policy layer in progress · go-live once that layer is in place. Flanders (BE): intake open · policy layer not established · go-live not established. Germany: intake is a conversation and not a start date; production onboarding opens later · policy layer active · go-live not open yet. Austria: intake is a conversation · policy layer not established · go-live not open yet. France: research only, no intake · policy layer not established · go-live not open. Other EU: not established, so tell us where and we will tell you what we know.
The evidence chain
A receipt per interaction, Merkle-aggregated, with the production export, measured on browser calls.
Every governed interaction generates a compliance receipt: the prompt and the answer as hashes rather than words, the disclosure event, the recorded policy version. Receipts are hashed and Merkle-aggregated, so a later change or reordering is detectable by recomputation against a retained commitment. The production export of a set of those receipts is a different thing from the receipt itself: Retested on 26 August after the deploy: five calls, thirty validated turns, thirty receipts, across two agents and two configurations, measured on browser calls rather than over a telephone line. A call from 19 August, re-exported the next day, carried its receipts as well.
- A receipt per interaction: prompt and answer as hashes, disclosure event, policy version.
- SHA-256 hash chain plus Merkle aggregation.
- Three fields on the receipt can be recomputed by anyone, without a key and without an account; the fourth, inclusion in the Merkle root, is checked against the published root with the inclusion proof from the call's Article 50 report.
- PDF and JSON are the export formats. The export returns the receipts. That is not the same as an outsider being able to check the export: the export envelope refers to a key of ours, and that key is not published.
- PII crypto-shredding aligned to your GDPR retention policy (coming soon).
- Each sentence segment the model produces is checked before it is sent to speech; a segment already spoken is not withdrawn. In controlled testing
The portfolio
What is filed, and what that does and does not mean.
Audact has filed patent-pending families at the UK Intellectual Property Office covering Article 50 attestation, evidence-chain construction, brand-governance state machines and per-jurisdiction policy enforcement. They are filed applications. None of them is granted, and none has been searched or examined.
- Inline proxy, compliance receipt, multi-candidate validation.
- Brand governance and commercial transaction validation.
- Article 50 attestation and disclosure sequencing.
- A regulatory overlay engine per country.
- Follow-on filings on platform governance and risk.
Onboarding
Our engineering team configures your first agent.
Each agency is onboarded personally. The agent is configured end to end before go-live: disclosure wording, voice, knowledge base, and one outbound webhook to an endpoint you provide (In development; see /integrations). You operate it; we engineer it.
- A twenty-minute scoping call, then a contract.
- First governed test call the same day the contract is signed.
- Four hours a month of engineering time for the first ninety days.
- Primary application hosting: Frankfurt; a backup region in Ireland is being set up.
Questions
- What does 'controlled testing' mean here?
- Voice runs, and it runs with real calls, but not yet as a generally available product with a public sign-up. That distinction matters enough to write down rather than to smooth over with the word 'live'.
- Can I use my own model or my own speech vendor?
- No. The speech stack is managed for you: speech-to-text, the language model and text-to-speech are run by Audact. That is a deliberate product decision: the governance layer can only guarantee what it sits in front of.
- Is there anything on this list I cannot use today?
- Yes, and each of those lines carries a status rather than 'working today'. Chat across web, WhatsApp and email is planned and not part of the current offer; machine-readable marking of synthetic audio under Article 50(2) carries the status Planned: no date has been established and it is not in the product today.
The platform
Control and Proof, the application on top of them, and the governance engine underneath every channel, with the status of each part.
Related
The platform Proof A receipt read field by field, with the chain and the Merkle path.
Brand governance What the agent may and may not say, checked before dispatch.
Article 50 What the law requires, with the article numbers.
Sector templates Pre-built flows with the governance nodes already in them.
Pricing Founding rate €1,497 + €499/month, for the first three external agencies.