OWNER-APPROVED ENGLISH VERSION — EFFECTIVE 26 AUGUST 2026. Version 1.0.0. Independent Czech legal review is recommended but has not been obtained. This English document applies only where English satisfies applicable language requirements. Publication does not enable paid signup, Stripe Live, a blocked territory or an unverified product capability.
1. Configured and intended processing routes
DIRI AI can use different routes for a requested training interaction. The authenticated session must bind one training language and disclose the active route at a useful level without exposing secrets.
| Provider | Model/function in current configuration | Data needed | Status on effective date |
|---|---|---|---|
| OpenAI | gpt-realtime-2.1 for realtime speech interaction; gpt-4o-transcribe for transcription; gpt-4o-mini for text fallback; gpt-5.4-mini for feedback | session instructions, selected scenario, necessary audio/text turns, transcript or scoring inputs and technical identifiers | public production calls fail closed; final enterprise/API settings, DPA, retention and regions not approved |
| ElevenLabs | eleven_flash_v2_5 for low-latency text-to-speech; an approved profile may use eleven_multilingual_v2 | generated reply text, language, voice/profile identifiers and technical metadata; output is synthetic audio | public production calls fail closed; final DPA, retention, region, voice rights and language quality gate incomplete |
| DIRI AI deterministic/mock route | scripted or locally generated test responses without a public model call | test scenario and local session state | used for controlled testing/demo paths; must not be represented as production model quality |
Deprecated, experimental, test-only or quality-assurance models must not silently receive public Customer Content. A model identifier in source configuration is not proof that its production use is contracted, affordable, supported or legally approved.
2. Provider roles and safeguards
DIRI AI intends to use model providers as contracted processors for customer-requested functions. Before activation, the Vendor and Transfer Register must verify:
- contracting entity and signed DPA;
- whether content is used to train provider models and the applicable opt-out/enterprise setting;
- provider abuse-monitoring and retention periods;
- configured processing region and remote-support access;
- transfer mechanism and assessment;
- deletion/API controls and incident notification; and
- voice licence/provenance and prohibited-use terms.
No “zero retention”, “EU only” or “not used for training” provider claim may be published without account-level evidence matching the exact API route.
3. Languages and voice quality
The product currently implements and tests Czech, English, Slovak, Polish and German interface or training paths. Provider models may advertise more languages, but DIRI AI does not promise those languages until the complete product flow, native listening quality, accent, pronunciation, transcription, safety and support have passed for each one.
All synthetic voices can vary. A voice trained primarily in another language may retain its native accent or become less stable. The selected training language is fixed for a session so routing and evaluation do not change silently mid-call.
4. Content use and retention
DIRI AI sends only data needed for the chosen function. Customer Content must not be used to train a general-purpose provider model without a separate, voluntary and specific choice and lawful basis. Optional quality review must be minimised, access-controlled and disclosed.
DIRI AI's own retention is in the Retention Policy. Provider copies may follow different short operational or abuse-monitoring periods and must be stated in the effective version after account verification. Removing content from DIRI AI does not by itself prove deletion at a provider; the deletion workflow must cover both.
5. Model changes
DIRI AI may replace a model for security, availability, quality, cost, deprecation or legal reasons. A replacement must pass privacy, quality, language and cost gates before Customer Content is routed to it. A materially different provider, purpose, retention, transfer or decision capability requires updated notice and, where applicable, subprocessor notice or consent.
6. AI output and oversight
Model output can be wrong, inconsistent or biased. It is training feedback, not professional advice or an employment decision. The AI Transparency Notice describes review, challenge and prohibited uses.
Questions: support@diriai.com. Privacy questions: privacy@diriai.com. Effective version: 1.0.0. Effective date: 26 August 2026.