Skip to main content
Audact
See the offer

Research programme

Aggregate compliance research: what is in it, and who can get to it.

Planned

A privacy-preserving research dataset built from Audact deployments that opted in. It holds aggregate signal about how disclosure behaves in practice (per country, per language, per failure mode), and it holds no personal data. Access requires an opt-in, k-anonymity of at least 50 and an external counsel review per licensing deal.

Contents

Five classes of signal.

Each row is an aggregate over a record class, not a conversation. The dataset answers questions about patterns; it cannot answer questions about a person, and that is by construction rather than by policy.

  • Signal
    Article 50 disclosure effectiveness per country: drop-off, comprehension and opt-out rates by language and jurisdiction.
    Status
    Planned
  • Signal
    Model failure patterns: hallucination, scope violation and bias-trigger categories across providers.
    Status
    Planned
  • Signal
    Enforcement severity per jurisdiction: fines, supervisory-authority decisions and case-law signal.
    Status
    Planned
  • Signal
    Consent-mechanism durability: opt-in against implied against opt-out, across re-contact scenarios.
    Status
    Planned
  • Signal
    Prompt-incident classification: a taxonomy of the failure modes that triggered a policy intervention.
    Status
    Planned

Design

Privacy-preserving by construction.

  • Control
    Opt-in required. No customer data enters the aggregate without it.
    Status
    Planned
  • Control
    k-anonymity of at least 50: every record class contains at least fifty indistinguishable members.
    Status
    Design rule
  • Control
    No personal data. Names, identifiers, voice samples and transcripts are excluded by design.
    Status
    Design rule
  • Control
    Aggregation runs in the Frankfurt region.
    Status
    Planned

Access

Partner licensing.

Privacy-preserving aggregate insight from governed conversations can support EU-aligned research under a partner licence. Every route in runs through the same three conditions: an opt-in, k-anonymity of at least 50, and an external counsel review per licensing deal.

Pricing is on enquiry. Write to research@audact.ai with the question you are trying to answer and the jurisdictions you need.

Questions

Does the dataset contain conversations?
No. It contains aggregates over record classes. Names, identifiers, voice samples and raw transcripts are excluded by design, and a record class is only licensable once it holds at least fifty indistinguishable members.
How does a customer opt in?
The opt-in is planned; there is no opt-in route yet. Nothing enters the aggregate without it, and an opt-in can be withdrawn.
How is this different from the synthetic dataset?
This one aggregates signal from real, opted-in deployments. The synthetic conversation dataset is generated from templates and policy bundles and contains no customer conversation at all. They are separate programmes with separate access routes.

Synthetic conversation data

The other half of the research programme: compliance-labelled conversations generated from templates and policy bundles rather than aggregated from real deployments.

See the synthetic dataset

Related