AI terms and AI use policy

1. What the AI does

The platform uses large language models, speech-to-text, text-to-speech, embeddings and image models from the providers listed on the Subprocessors page to answer questions from your knowledge base, hold voice calls, summarise, score leads, draft messages and content, read documents and analyse data. Model choice can change to keep quality, cost and availability; material changes of provider follow the 30-day sub-processor notice.

2. Your data and model training

2.1 We do not use Client Data, prompts or outputs to train or improve AI models, and we use provider settings that stop providers doing so: business API terms that exclude training (Anthropic, OpenAI, paid Google Gemini / Vertex AI, Azure), Deepgram requests with model-improvement opt-out, ElevenLabs zero-retention mode where the plan allows. The unpaid Gemini API tier is never used in production. Each provider's commitment is listed on the Subprocessors page.

2.2 Providers may keep inputs for a short period for abuse monitoring under their terms [typically up to 30 days]. Where a provider offers zero data retention to us, we use it for sensitive products [Hari: confirm eligibility].

2.3 You can choose India-only processing, which limits AI and speech to providers that process in India where available.

3. Output

• AI output can be inaccurate, incomplete, outdated or biased. You must review it before relying on it where it matters, especially for money, legal, medical, employment and safety matters.
• Between you and us, you own the output generated for you to the extent the law allows. Output may not be unique; others may receive similar output.
• You are responsible for output you publish or send, and for checking that it does not infringe others' rights.

4. Transparency to the people you serve

• Chat assistants show an "AI assistant" label and voice agents say they are AI at the start of each call; these cannot be switched off (EU AI Act Art. 50, California SB 1001 and AB 2905, Utah AI Policy Act, FCC TCPA rules, India MeitY advisories).
• Agents answer truthfully when asked whether they are a bot, and hand over to a human on request where you have staff available.
• Automated replies on messaging channels carry a short "automated reply" marker on the first message.
• AI-generated images, audio and video are labelled and carry content credentials or IPTC metadata where supported; you must not remove them. Synthetic media published to YouTube, Meta, LinkedIn or X is flagged as such where the platform supports it.
• Where you use automated scoring or decisions, tell people as your law requires (GDPR Art. 13-15 and 22, DIFC Regulation 10, Australia Privacy Act ADM transparency from 10 December 2026, Quebec Law 25).

5. Human oversight and prohibited uses

AI scores and recommendations in the platform are advisory. The Acceptable use policy lists prohibited and high-risk uses (employment, credit, insurance, housing, education, essential services) that need human review of each decision or are not allowed. Do not use the platform for emotion recognition of employees or students or any practice banned by EU AI Act Article 5.

6. Reporting problems

If an AI answer is harmful, unlawful or wrong in a way that matters, use the report option on the answer or write to hariprathish@originpixel.in. We review reports, fix knowledge or prompts, and record AI incidents in our incident process. People affected by AI output from your business can also use our grievance officer.

7. Our AI governance

We keep an inventory of AI features with their models, data, risks and evaluation results, run quality evaluations before changes and periodically, monitor answers with feedback, and fall back or hand over when a model is unsure. [Model / system cards per feature will be published at /developers/ai-systems (WP-AI).] We align with NIST AI RMF and ISO/IEC 42001 [no certification held].