Talkative uses AI to power features such as voice and chat bots, agent assist, summaries and analytics. We do not build or train our own AI models. We use established third-party large language models (LLMs) for inference only, which means your data is used to generate a response and is not used to train any model. You choose which LLM to use from a self-service configuration in the application.
Every AI feature is optional and configurable. No customer data is shared with third parties outside the sub-processors disclosed in our Data Processing Agreement. Most AI processing runs with zero data retention at the model provider, with the limited exceptions listed under "Data retention" below.
For the complete, always-current version of this information, see our AI data compliance page. Our Data Processing Addendum lists every sub-processor.
AI features and the data they process
Where a row says "See OpenAI rows", the feature runs in the OpenAI region listed for that model in the LLM table below: OpenAI EU region for UK and Germany instances, OpenAI US region for US and Australia instances.
Feature | What it does | Data processed | PII | Model or subprocessor | Region | Retention |
Voice AI | Real-time conversation with an interactive voice bot | Conversation transcript, interaction data | Yes | Your chosen LLM, plus speech-to-text and text-to-speech providers | See tables below | See tables below |
Chat AI | Real-time conversation with a chat bot on web, SMS or social | Conversation transcript, interaction data | Yes | Your chosen LLM | See LLM table | See LLM table |
LLM Chatbot | Real-time conversation with a prompt-driven AI chat bot that can use tools | Conversation transcript, interaction data | Yes | GPT-5 series | See OpenAI rows | 0 days |
Email AI | Suggested responses to emails | Conversation transcript (email contents) | Yes | Your chosen LLM | See LLM table | See LLM table |
AI Agent Assist (KB suggestions, Navi, auto complete) | Real-time agent suggestions based on the ongoing chat | Conversation transcript, interaction data | Yes | Your chosen LLM | See LLM table | See LLM table |
AI Agent Rephrase | Rephrasing options for agent-written responses | Agent input text only | No | GPT-4.1 | See OpenAI rows | 0 days |
AI Agent Training | Fictitious interactions for agent upskilling | Fictitious conversation transcript | No | GPT-4.1 | See OpenAI rows | 0 days |
AI Interaction Summaries | Summaries of customer interactions for internal review | Conversation transcript | Yes | Regionally hosted Llama | See Llama rows | 0 days |
AI Interaction Data Capture | Automatically extracts key fields such as names and issues | Conversation transcript | Yes | GPT-4.1 mini | See OpenAI rows | 0 days |
AI Analytics Assistant | Query customer interaction trends in natural language | Analytics reports | No | GPT-4.1 | See OpenAI rows | 0 days |
AI Insights Report | Periodic reports on trends, issues and agent performance | Conversation transcripts | Yes | Open-weights model hosted on AWS | Mirrors your Talkative region | 0 days |
AI Sentiment Analysis | Evaluates the emotional tone of text interactions | Conversation transcript | Yes | AWS Comprehend | Mirrors your Talkative region | 0 days |
AI Auto Tagging | Reviews chatbot and abandoned interactions to assign appropriate tags | Conversation transcript | Yes | GPT-4.1 mini | See OpenAI rows | 0 days |
AI Auto Suggest Tags | Reviews agent interactions and suggests appropriate tags | Conversation transcript | Yes | GPT-4.1 mini | See OpenAI rows | 0 days |
AI Interaction Alerts | Evaluates interactions against your alert conditions and flags matches | Conversation transcript | Yes | GPT-4.1 mini | See OpenAI rows | 0 days |
AI Voice Call Evaluation | Grades Voice AI calls against evaluation criteria | Voice call audio, conversation transcript | Yes | GPT audio | See OpenAI rows | 0 days |
Call Recording Transcription | Transcribes recorded phone calls | Voice call audio | Yes | Deepgram | US for US and Australia instances; EU for UK and Germany instances | Audio files not stored; logs retained for 90 days |
KB Suggestions | Analyses failed knowledge base lookups to suggest new knowledge base content | Conversation transcript | Yes | GPT-4.1 | See OpenAI rows | 0 days |
KB Retrieval | Finds the relevant sections of your knowledge base to give the LLM context for a customer's message | Conversation transcript | Yes | Cohere Embed and Cohere Rerank via Amazon Bedrock. OpenAI models on localised AI configurations | Embeddings mirror your Talkative region. Reranking runs in Frankfurt for UK and Germany instances, US for US instances and Tokyo for Australia instances. Where best-effort reranker regionalisation has been enabled with your written consent, reranking runs in the US for all instances | 0 days |
Widget Colour Generation | When your account is created with a website URL, a screenshot of the site is used to pick the two best colours for your chat widget | Website screenshot | No | GPT-4.1 | See OpenAI rows | 0 days |
Message Translation | Translates customer messages for agents and detects the customer's language | Conversation transcript | Yes | Google Translate by default; DeepL where enabled with your written authorisation | Google: US. DeepL: global endpoint, no regional data residency | Google: held in memory only, never written to storage. DeepL: deleted after translation |
Voice Translate | Translates a live phone call between customer and agent in real time | Voice call audio, conversation transcript | Yes | DeepL real-time voice translation, plus ElevenLabs text-to-speech | DeepL: global endpoint, no regional data residency. ElevenLabs: US for US and Australia instances; EU for UK and Germany instances | 0 days |
AI Search | Finds the right page in the app from a search term | User's search text | No | Regionally hosted Llama | See Llama rows | 0 days |
LLM Chatbot Evaluations | Grades LLM chatbot conversations against evaluation criteria | Conversation transcript | Yes | GPT-5 series | See OpenAI rows | 0 days |
Chatbot AI Phrase Matching | Matches a customer's reply to the options in a chatbot flow when there is no exact match | Conversation transcript | Yes | GPT-4.1 | See OpenAI rows | 0 days |
KB Test Question Generation | Generates sample customer questions from an AI Insights Report and past interactions for knowledge base test runs | AI Insights Report, conversation transcripts | Yes | GPT-4.1, GPT-4.1 mini | See OpenAI rows | 0 days |
Voice AI Test Runs | Generates test scenarios for simulated calls against a Voice AI bot | Bot prompt, company name | No | GPT-4.1 | See OpenAI rows | 0 days |
LLM Chatbot Test Runs | Generates test scenarios and runs simulated conversations against an LLM chatbot | Bot prompt, company name | No | GPT-4.1 | See OpenAI rows | 0 days |
AI Knowledge Base Prompt Review | Reviews an AI Knowledge Base prompt and gives feedback | Prompt text | No | GPT-5 series | See OpenAI rows | 0 days |
Voice AI Publish Details | Pre-fills a version's label and notes from what changed in the draft | Bot configuration | No | GPT-4.1 | See OpenAI rows | 0 days |
UI Translation | Translates the app's interface text into the user's language | Fixed interface text only | No | GPT-4.1 mini | See OpenAI rows | 0 days |
If your account is configured for localised AI, OpenAI models are used for all of the features above, in the OpenAI region matching your instance. Deepgram, DeepL and ElevenLabs are unchanged.
Choosing your LLM
You select your LLM in a self-service configuration within the application. The models currently available:
Model | Provider | Hosted by | Region | Retention |
GPT-4.1 | OpenAI | OpenAI | OpenAI EU region for UK and Germany instances; OpenAI US region for US and Australia instances | 0 days |
GPT-4.1 mini | OpenAI | OpenAI | OpenAI EU region for UK and Germany instances; OpenAI US region for US and Australia instances | 0 days |
GPT-5 series (some variants available for Voice AI only) | OpenAI | OpenAI | OpenAI EU region for UK and Germany instances; OpenAI US region for US and Australia instances | 0 days |
Gemini Flash | Google | Google (Vertex AI) | Google EU multi-region for UK and Germany instances; Google US multi-region for US and Australia instances | 0 days. Google may log prompts to detect abuse under its Acceptable Use Policy |
Gemini Flash Lite | Google | Google (Vertex AI) | Google EU multi-region for UK and Germany instances; Google US multi-region for US and Australia instances | 0 days. Google may log prompts to detect abuse under its Acceptable Use Policy |
Claude Sonnet | Anthropic | AWS Bedrock | AWS London for UK instances; AWS EU region for Germany instances; AWS Australia region for Australia instances; AWS US region for US instances | 0 days |
Claude Haiku | Anthropic | AWS Bedrock | AWS EU region for UK and Germany instances; AWS Australia region for Australia instances; AWS US region for US instances | 0 days |
Llama 70B | Meta | AWS Bedrock | AWS London for UK, Germany and Australia instances; AWS US for US instances | 0 days |
Llama 8B | Meta | AWS Bedrock | AWS London for UK, Germany and Australia instances; AWS US for US instances | 0 days |
Talkative LLM* | OpenAI open-weights model | Talkative AWS | Any AWS region | 0 days |
*Talkative LLM is an open-weights model, fully hosted and managed by Talkative as a dedicated resource. It is available as a paid add-on; contact us for details.
Voice AI subprocessors
Voice AI combines three components: your chosen LLM (currently OpenAI LLMs are supported for Voice AI, with AWS Bedrock models in progress), a text-to-speech (TTS) provider, and a speech-to-text (STT) provider.
Type | Provider | Region | Retention |
Text-to-speech | ElevenLabs | US for US and Australia instances; EU for UK and Germany instances | 0 days (zero-retention mode) |
Speech-to-text | Deepgram | US for US and Australia instances; EU for UK and Germany instances | Audio files not stored; logs retained for 90 days |
Text-to-speech | Google | US, Europe | 0 days |
Speech-to-text | Google | US, Europe | 0 days |
Speech-to-speech | OpenAI | OpenAI EU region for UK and Germany instances; OpenAI US region for US and Australia instances | 0 days |
We are actively evaluating additional speech-to-text and text-to-speech providers. Contact us if you have specific requirements.
Where AI processing happens
Any data processed by an LLM is also processed and stored within the main Talkative application, which is hosted on AWS. Talkative currently runs instances in the UK (London), Germany, Australia, and the United States. Where possible, the subprocessors above are mapped to the AWS region that matches your Talkative instance. For example, Llama on AWS Bedrock runs in the UK region for customers on the UK instance. Where a provider does not offer a matching region, the tables above state where processing takes place. Other AWS regions are planned; contact us for timelines.
Data retention
The default across our AI features is zero data retention: providers process your data to generate a response and do not store it. Two exceptions apply:
- Deepgram (speech-to-text for Voice AI and Call Recording Transcription) does not store audio files, but retains logs for 90 days.
- Google may log prompts to detect abuse and policy violations as part of providing its generative AI services.
Within the main Talkative application, data is handled according to your instance region and our Data Processing Agreement.