AI systems

What each AI feature of the Origin Platform does, which AI providers it uses, how people are told they are dealing with AI, who stays in control and what it cannot do. Written for businesses using the platform (deployers under the EU AI Act and the Colorado AI Act) and for their customers. Customer data is never used to train AI models: the table at the end shows how that is enforced for each provider.

Report a problem with an AI answer from the chat ("Report") or write to the business. Businesses can raise AI incidents from AI safety in their app. Last reviewed November 2026.

System cards

Front Desk website chat and widget

Front Desk · Limited risk (transparency duties)

Purpose: Answers visitor questions from the business's own knowledge base, offers booking times and hands over to staff.
Inputs: Visitor messages, the business's knowledge base, services and free times.
Outputs: Short text answers with sources, booking requests, handovers.
AI providers: Anthropic Claude, Google Gemini (paid tier), OpenAI
How people know it is AI: Locked "AI assistant" badge and greeting suffix in 12 languages, "AI can make mistakes" line, truthful answer when asked if it is a bot (rule outside the model).
Human oversight: Confidence threshold with automatic handover, staff take over any chat, visitors and staff can report an answer.
Limits and known risks: Answers only from the knowledge base and can still misread it. Not medical, legal or financial advice.
Evaluation: Golden-question harness (accuracy, groundedness, citations, handover) plus disclosure, bias and manipulation suites.
Not for: Medical diagnosis, legal advice, emergencies.

Front Desk auto-replies on messaging channels

Front Desk · Limited risk (transparency duties)

Purpose: The same assistant replying automatically on WhatsApp, Instagram, Messenger, Telegram and SMS.
Inputs: Customer messages, knowledge base.
Outputs: Text replies.
AI providers: Anthropic Claude, Google Gemini (paid tier)
How people know it is AI: The first automatic reply of every conversation ends with "Automated reply from an AI assistant" in the customer's language.
Human oversight: Optional review before sending; the team inbox shows every AI reply; text that claims to be a person is held.
Limits and known risks: Same as the website chat; channel rules (24-hour window, templates) apply.
Evaluation: Same harness as the website chat.
Not for: Debt collection or threats; anything a person must decide.

Voice Workforce AI phone agent

Voice Workforce · Limited risk (transparency duties)

Purpose: Answers and makes phone calls, books appointments and transfers to people.
Inputs: Caller audio (speech to text), knowledge base, calendar.
Outputs: Synthesised speech, call summaries, bookings.
AI providers: Deepgram (mip_opt_out), ElevenLabs, Sarvam, OpenAI, Anthropic Claude, Google Gemini (paid tier)
How people know it is AI: Says it is an AI at the start of every call (cannot be switched off) and gives the recording notice before recording.
Human oversight: Warm transfer to staff, call review, QA by people. In EU / UK no automated QA or sentiment on calls staff took over.
Limits and known risks: Recognition errors on noisy lines and code-switched speech. Sentiment is read from the transcript text only, never from the voice (no emotion recognition from biometrics).
Evaluation: Latency and accuracy metrics per call; call QA rules.
Not for: Emergency services, debt collection, political calls.

Revenue Engine follow-up drafts

Revenue Engine · Limited risk (transparency duties)

Purpose: Writes personalised follow-ups for leads.
Inputs: Lead details, conversation summary.
Outputs: Email and WhatsApp drafts.
AI providers: Anthropic Claude, Google Gemini (paid tier)
How people know it is AI: Automatically sent AI-written emails carry an AI footer; the first automatic WhatsApp message carries the AI marker.
Human oversight: Drafts by default; auto-send only when the business turns it on; optional review; manipulative pressure is held.
Limits and known risks: Tone can be off; check offers and prices.
Evaluation: Manipulation guard on every automatic send.
Not for: Hiring, credit, insurance, housing, education admission or any decision with legal or similar effects on a person.

Sales Workforce sequence emails

Sales Workforce · Limited risk (transparency duties)

Purpose: Personalises outbound sales emails within a sequence.
Inputs: Prospect business details.
Outputs: Email drafts.
AI providers: Anthropic Claude, Google Gemini (paid tier)
How people know it is AI: AI footer on automatically sent AI-written emails.
Human oversight: Drafts by default; sending windows and caps; optional review.
Limits and known risks: Research can be out of date.
Evaluation: Manipulation guard on every automatic send.
Not for: Hiring, credit, insurance, housing, education admission or any decision with legal or similar effects on a person.

Lead and prospect scoring

Revenue Engine, Sales Workforce · Limited risk (transparency duties)

Purpose: Ranks leads and prospects so the team can prioritise.
Inputs: Lead and company details against the business's own criteria.
Outputs: A score with the reasons.
AI providers: Rules (no external AI)
How people know it is AI: Scores are shown as advice with their reasons.
Human oversight: Advisory only: nothing is accepted, rejected or charged on a score (fixed in the database).
Limits and known risks: Reflects the business's own criteria, including its biases.
Evaluation: Rule review.
Not for: Hiring, credit, insurance, housing, education admission or any decision with legal or similar effects on a person.

Customer Experience triage and reply suggestions

Customer Experience · Limited risk (transparency duties)

Purpose: Classifies support messages, spots unhappy customers from text and suggests replies.
Inputs: Support messages.
Outputs: Category, priority, text sentiment, suggested replies.
AI providers: Anthropic Claude, Google Gemini (paid tier)
How people know it is AI: Suggestions are shown to staff, not sent.
Human oversight: A person sends every reply unless the business automated it.
Limits and known risks: Text sentiment misreads sarcasm and mixed languages. Never applied to staff.
Evaluation: Rule and sample review.
Not for: Monitoring or scoring employees.

Team inbox suggestions and summaries

Front Desk · Minimal risk

Purpose: Suggests replies and summarises conversations for agents.
Inputs: Conversation text.
Outputs: Suggestions, summaries.
AI providers: Anthropic Claude, Google Gemini (paid tier)
How people know it is AI: Shown only to staff.
Human oversight: Staff choose, edit and send.
Limits and known risks: Summaries can miss details.
Evaluation: Staff feedback.
Not for: Evaluating staff performance.

Business Intelligence "Ask your data"

Business Intelligence · Minimal risk

Purpose: Answers questions about the business's own numbers.
Inputs: Aggregated business data.
Outputs: Charts and short explanations.
AI providers: Anthropic Claude, Google Gemini (paid tier)
How people know it is AI: Shows the query and data used.
Human oversight: No automatic actions.
Limits and known risks: Can pick the wrong measure.
Evaluation: Query review.
Not for: Hiring, credit, insurance, housing, education admission or any decision with legal or similar effects on a person.

Back-Office document reading

Back-Office · Limited risk (transparency duties)

Purpose: Reads invoices, receipts and KYC documents into fields for review.
Inputs: Uploaded documents.
Outputs: Extracted fields.
AI providers: Anthropic Claude, Google Gemini (paid tier), OpenAI
How people know it is AI: Fields are marked AI-read until approved.
Human oversight: Every document is approved by a person.
Limits and known risks: Handwriting and poor scans; Aadhaar numbers are masked before AI reading.
Evaluation: Field accuracy on samples.
Not for: Identity verification decisions without a person.

Content Studio post writing

Content Studio · Limited risk (transparency duties)

Purpose: Writes and rewrites social posts in the brand voice.
Inputs: Brand kit, ideas, earlier posts.
Outputs: Post drafts.
AI providers: Anthropic Claude, Google Gemini (paid tier)
How people know it is AI: Marked "AI-generated" in the app; "Made with AI" added to the published text when the business asks.
Human oversight: A second person approves before publishing.
Limits and known risks: Can state untrue things about the business.
Evaluation: Rule check and manipulation check before approval.
Not for: News or claims about real people.

Content Studio image generation

Content Studio · Limited risk (transparency duties)

Purpose: Generates images for posts.
Inputs: Prompt, brand colours.
Outputs: PNG / JPEG images.
AI providers: OpenAI, Google Gemini (paid tier)
How people know it is AI: IPTC DigitalSourceType trainedAlgorithmicMedia in the file, visible "Made with AI" on published posts, YouTube containsSyntheticMedia.
Human oversight: A second person approves before publishing.
Limits and known risks: Can produce realistic scenes that never happened.
Evaluation: Marker presence test on every generated file.
Not for: Deepfakes of real people, news, political content.

Copilot and automations

Platform · Limited risk (transparency duties)

Purpose: Lets staff ask the platform to do tasks.
Inputs: Requests and records the person can see.
Outputs: Answers, proposed actions.
AI providers: Anthropic Claude, Google Gemini (paid tier)
How people know it is AI: Shown as an AI copilot.
Human oversight: Approval for sending, deleting and paying actions.
Limits and known risks: Can misunderstand requests.
Evaluation: Approval logs.
Not for: Hiring, credit, insurance, housing, education admission or any decision with legal or similar effects on a person.

Translation

Platform · Minimal risk

Purpose: Translates customer messages and knowledge.
Inputs: Messages, knowledge.
Outputs: Translations.
AI providers: Sarvam, Google Cloud, Azure, Self-hosted IndicTrans2, Anthropic Claude
How people know it is AI: Translated messages are marked.
Human oversight: Staff see the original.
Limits and known risks: Idioms and names can be mistranslated.
Evaluation: Spot checks per language.
Not for: Legal documents without a person checking.

Knowledge search

Platform · Minimal risk

Purpose: Finds the right knowledge for answers.
Inputs: Knowledge documents.
Outputs: Ranked passages.
AI providers: Google Gemini (paid tier), OpenAI, Cohere, Voyage, Jina, Self-hosted
How people know it is AI: Answers cite their sources.
Human oversight: Knowledge quality page shows gaps and ratings.
Limits and known risks: Can miss documents with unusual wording.
Evaluation: Citation score in the harness.
Not for: -

Practices the platform does not allow

Manipulative or deceptive techniques that distort behaviour. System prompt rule plus the manipulation guard: automatic AI messages with false urgency, scarcity pressure or threats are held for a person.
Pretending to be a human. Disclosure cannot be switched off; "are you a bot?" is answered truthfully by a rule outside the model; AI text claiming to be a person is held.
Emotion recognition in the workplace. Sentiment is text-only and about customers; in EU / UK orgs automated QA scores and sentiment are blocked on calls staff took over.
Social scoring, biometric categorisation, predictive policing, facial scraping. Not offered. No biometric processing; lead scores are advisory and limited to the business's own sales criteria.
Exploiting vulnerabilities (age, disability, situation). Acceptable Use Policy; manipulation guard; human handover on sensitive topics.

See also the Acceptable Use Policy and the sub-processors.

AI providers and training

Provider	How training is prevented	Detail
Anthropic Claude API	Provider terms: no training on API data	Commercial Terms: API data is never used for training.
OpenAI API	Provider terms: no training on API data	API data is not used for training by default; the platform never opts in.
Google Gemini API (paid tier)	Paid tier only (enforced at start-up)	Paid tier (Cloud Billing on the project): prompts and responses are not used to improve Google products.
Google Gemini API (free tier)	Never used for customer data	Free tier data may be used by Google to improve products: never used for customer data.
Google Vertex AI	Provider terms: no training on API data	Google Cloud terms: no training on customer data.
Deepgram	Opt-out flag sent on every request	mip_opt_out=true on every request (packages/voice-providers).
ElevenLabs	Opt-out flag sent on every request	enable_logging=false on every request (zero retention, ELEVENLABS_ZERO_RETENTION=1) or the account training opt-out (ELEVENLABS_NO_TRAINING_CONFIRMED=1).
Sarvam AI	Account opt-out, confirmed before use	Written no-training confirmation from Sarvam (SARVAM_NO_TRAINING_CONFIRMED=1).
AssemblyAI	Account opt-out, confirmed before use	Account training opt-out (ASSEMBLYAI_NO_TRAINING_CONFIRMED=1).
Microsoft Azure AI	Provider terms: no training on API data	Azure terms: no training of foundation models on customer data.
Google Cloud Speech / TTS / Translate	Provider terms: no training on API data	Google Cloud terms: data logging off, no training.
Amazon Polly	Account opt-out, confirmed before use	AWS Organizations AI services opt-out policy (AWS_AI_OPT_OUT_CONFIRMED=1).
Cartesia	Account opt-out, confirmed before use	Written no-training confirmation (CARTESIA_NO_TRAINING_CONFIRMED=1).
Cohere	Account opt-out, confirmed before use	Dashboard data control: training off (COHERE_NO_TRAINING_CONFIRMED=1).
Voyage AI	Account opt-out, confirmed before use	Dashboard training opt-out (VOYAGE_NO_TRAINING_CONFIRMED=1).
Jina AI	Account opt-out, confirmed before use	Written no-training confirmation (JINA_NO_TRAINING_CONFIRMED=1).
Groq	Provider terms: no training on API data	Groq terms: inputs are not used for training.
Mistral AI	Account opt-out, confirmed before use	Paid plan without training (MISTRAL_NO_TRAINING_CONFIRMED=1).
Self-hosted models	Runs on Origin's own servers	Runs on Origin's own servers.