> ## Documentation Index > Fetch the complete documentation index at: https://docs.helpalive.com/llms.txt > Use this file to discover all available pages before exploring further. # How HelpAlive's AI uses your data > What HelpAlive's agent reads, what it sends to the AI model provider and why, whether it trains models, and how long the provider keeps it. HelpAlive's agent sends data to an AI model only at the moment a user uses the chat or an assist, so it can answer questions about the page they are on and do tasks there. This page explains what data HelpAlive's agent sends to an AI model, when, and why, and what the model provider does with it. It is for privacy and security reviewers, and for admins writing their own privacy notice. ## AI and your data at a glance * The agent reads the current page only when someone uses the chat or an assist. Apart from password, card, one-time-code, file-upload and hidden fields, that page view is not redacted. * Before activity is stored, our servers scan its text and replace names, email addresses, phone numbers, IP addresses and card numbers. The scan is automatic and can miss some. * The model provider doesn't train on data sent through its API. * The page the agent read is not saved to our database. * Chat messages and assist text are kept for the window you set on **Settings → Privacy Management**, 1 year by default. ## Why the agent needs your data The agent answers questions about the page a user is on, and does tasks on that page. To do that, it has to read the page and know what the user is trying to do. Everything below is sent for that purpose, at the moment the user uses the chat or an assist. ## What HelpAlive reads, and when **Always, while the script runs: activity only.** HelpAlive records which pages are viewed, what is clicked, which fields change and which errors appear. This is not a copy of the page. Before activity is stored, our servers scan its text and replace names, email addresses, phone numbers, IP addresses and card numbers. The scan is automatic and can miss some. Password, card and one-time-code fields are never collected. **Only when someone uses the agent: the page.** When a user opens the chat, asks a question or runs an assist, the agent reads the current page as text, the way a person would see it, so it can help. That page goes to our model provider and is not redacted, apart from the fields listed below, because it is what the agent works from. No copy of the page is taken at any other time. ## What is sent, and when The model receives only what the current moment needs. That falls into five kinds of data: | Kind of data | When it is sent | | - | - | | **The page the user is on**: its address, title and content as structured text, **including field values** | When the chat opens, with each question, and at each step of an assist | | **The user's own words**: their message, the conversation so far, and their answers during an assist | With each question and each assist step | | **Profile fields you send** through `identify()`, such as display name, role and plan | When the chat opens and with each question | | **Recent activity**: what the user did on this visit, and the titles of their own recent chats | When the chat opens and with each question | | **Passages from your published docs** that match the question | With each question | Field values are sent as shown, except in password, card, one-time-code, file-upload and hidden fields, and in any field holding a card number. Apart from those fields, the page is not redacted, because it is what the agent needs to see. The `email` field from `identify()` is never sent to the model. An email the agent can see elsewhere, such as in `displayName`, on the page or in a chat, is sent. Your docs' text and your users' questions are also processed to find the right passages. Our model provider and a search-relevance provider do this work. See [Subprocessors](/trust/subprocessors). ## Where answers come from Answers come from the sources you publish on **Train**. Drafts never reach your users. Where your docs are silent, the agent can use general knowledge of your product, and your docs win when the two disagree. See [How HelpAlive works](/product/how-it-works). ## Model training and retention * **The model provider doesn't train on data sent through its API.** That covers the pages, questions and docs HelpAlive sends it. * **We ask the provider not to store responses.** * **The provider may keep requests and responses**, normally for up to 30 days, to check for abuse. * **We may use the kept words of chats and assists, de-identified, to improve HelpAlive for all customers.** To opt out, write to [support@helpalive.com](mailto:support@helpalive.com). See [Privacy Management](/dashboard/privacy-management). The model provider and the other companies that process data for us are listed on [Subprocessors](/trust/subprocessors). ## What we keep afterwards * **The page the agent read** is not saved to our database. A running assist holds it only while the assist is in progress. * **Chat messages and assist text** are kept in our database for the window you set on **Settings → Privacy Management**, 1 year by default. * **Our server logs** can hold users' chat questions and the page text assists read, never values typed into fields, for 30 days. See [Data we collect](/trust/data) for every kind of data and how long it is kept. ## What you control * **Who gets the agent**, and whether it does the steps or shows how, on **Controls**. * **What it answers from**, on **Train**. Only published sources are used. * **How long chat and assist text is kept**, on **Settings → Privacy Management**. * **Deleting a person's or tenant's data**, on **Settings → Privacy Management**. ## Privacy policy and terms These pages explain how HelpAlive handles data in practice. The legal terms are in our [Privacy policy](https://helpalive.com/legal/privacy) and [Terms of use](https://helpalive.com/legal/terms). ## Next Everything collected, redacted and kept. Who processes data, and for what. Hand-backs and confirmations. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.