Where Does Your Chatbot's Data Go? OpenAI API Data Policies Explained

Every security review of a website chatbot reaches the same three questions: does OpenAI train on our customers' messages, how long does it keep them, and who else stores them? The answers differ between ChatGPT, the app people chat with, and the OpenAI API, the developer service that chatbot vendors build on. Most of the confusion online comes from mixing the two.

The guide is written for founders, support leads and anyone filling in a vendor security questionnaire. It covers what OpenAI's API terms say about training and retention, what zero data retention and data residency change, which other systems touch a chat transcript, and ten questions to send any chatbot vendor. Every OpenAI rule below was read on OpenAI's own documentation, help center and enterprise privacy page on 19 October 2026.

PepoChat, the AI support agent we build, runs on OpenAI's API, so the last section lists the models it calls and the data it stores, taken from its product code and documentation. Nothing here is legal advice; where a rule matters for a contract or an audit, check the linked source on the day you rely on it.

What happens to a customer's message when a chatbot uses the OpenAI API?

The message makes a round trip. The visitor's browser sends it to the chatbot vendor's servers, the vendor adds context and calls OpenAI's API, and OpenAI returns the generated reply; the vendor and OpenAI can each keep a copy under their own rules.

The OpenAI API is the paid developer service that applications call over the internet to use OpenAI's models, billed by usage and governed by OpenAI's business terms rather than ChatGPT's consumer settings. A support chatbot built on it typically sends four things with each reply: the visitor's latest message, the recent conversation, passages retrieved from your knowledge base, and the bot's instructions. If the bot can look up an order or a subscription, the result of that lookup is sent too.

Your own content takes a second path. An embedding is a list of numbers that represents the meaning of a passage, and chatbots that answer from your documents use an embeddings model to turn every page and file into embeddings when you add them, then embed each visitor question to find the closest passages. That retrieval step, explained in our RAG guide for non-engineers, means your knowledge base passes through the model provider once at indexing time and in pieces at answer time.

So the full answer to where the data goes has three parts: the vendor's own systems, the model provider, and any tools the bot is connected to, such as email, a helpdesk or a CRM. The sections below take them in turn, starting with the rule people ask about most.

Does OpenAI train its models on API data?

No, not by default. OpenAI's data controls documentation says that as of 1 March 2023, data sent to the API is "not used to train or improve OpenAI models" unless the customer explicitly opts in to share it.

Opting in is possible, which is why the question is worth asking a vendor. OpenAI's help center says API organization owners can switch on data sharing, after which members can share feedback and API Playground chats to help improve OpenAI's models. A vendor that has never touched the setting is on the default: no training.

OpenAI's enterprise privacy page, last updated 8 January 2026, gives the same commitment for all business data: inputs and outputs from ChatGPT Business, ChatGPT Enterprise, ChatGPT Edu and the API Platform are not used for training by default, and the customer owns its inputs and outputs where the law allows. The page also says OpenAI encrypts data at rest with AES-256 and in transit with TLS 1.2 or higher, and that the API Platform has been audited for SOC 2 Type 2.

Not used for training does not mean not stored or never read. The same enterprise privacy page says OpenAI may run business data through automated content classifiers and safety tools, and that access to stored API data is limited to authorized employees, for engineering support, abuse investigations and legal compliance, and to contractors bound by confidentiality who review for abuse and misuse. Storage is the subject of the retention section below.

How is ChatGPT different from the OpenAI API?

ChatGPT is OpenAI's own chat app, sold as personal plans and as workspace plans for organizations, and each kind has its own data rules. The API is the separate developer service that chatbot vendors use, and none of ChatGPT's settings or defaults apply to it.

For personal plans, OpenAI's help article on how your data is used to improve model performance says it may use content from Free, Plus and Pro accounts to train its models unless you turn off Improve the model for everyone under Settings, Data controls, or opt out in OpenAI's Privacy Portal. Rating a reply with a thumbs up or down can still bring the whole conversation into training after an opt-out. Temporary Chats are not used for training and may be kept for up to 30 days for safety.

Workspace plans behave more like the API. By default, OpenAI does not use content from ChatGPT Business, Enterprise or Edu to train its models, workspace admins control how long conversations are kept, and response feedback in those workspaces does not permit training on the conversation.

Read on OpenAI's help center, enterprise privacy page and API data controls documentation on 19 October 2026.

The practical difference for a support team is large. A teammate who pastes a customer's email into a personal ChatGPT account is under consumer rules, while a chatbot vendor calling the API is under business terms. Our post on using ChatGPT for customer service covers the workspace options and their limits in more depth.

How long does OpenAI keep API requests?

Up to 30 days by default. OpenAI's data controls documentation says abuse-monitoring logs, which may contain prompts, responses and metadata derived from them, are generated for all API use and kept for up to 30 days, unless the law requires longer retention or it is reasonably necessary to protect OpenAI's services or others from harm.

OpenAI separates two kinds of stored data. Abuse-monitoring logs are records OpenAI keeps to enforce its usage policies, and application state is data that some API features keep so they can work, such as a stored response, an uploaded file or a vector store. The endpoints a chatbot uses decide which of the two apply.

For Chat Completions, which generates replies, and Embeddings, which builds the search index, the documentation's endpoint table lists 30 days of abuse-monitoring retention and no application state. The Responses API keeps response data for at least 30 days by default, and files, vector stores and Assistants objects stay until the developer deletes them.

Endpoint rules from OpenAI's data controls documentation and migration guide, read 19 October 2026.

The store parameter is a per-request switch that tells OpenAI whether to keep the output of a request. OpenAI's migration guide says Responses are stored by default and Chat Completions are stored by default for new accounts, and that setting store to false disables storage in either API, while the data controls table lists no application state for Chat Completions. A vendor that sets store to false on every request does not depend on which reading applies.

Two smaller rules sit outside those tables. Images and files sent as inputs are scanned for child sexual abuse material, and a flagged image is kept for manual review even under zero data retention. Prompt caching can hold encrypted intermediate data on OpenAI's GPU machines, which the documentation says is not kept beyond a 24-hour expiry.

A court order changed these rules for part of 2025. A preservation order in The New York Times' lawsuit required OpenAI to keep consumer ChatGPT and API content it would normally delete, though not ChatGPT Enterprise and Edu data or API traffic under zero data retention. OpenAI's update dated 22 October 2025 says those obligations ended on 26 September 2025 and API data is again deleted after 30 days, while a limited set of April to September 2025 data stays under legal hold.

What is zero data retention, and who can get it?

Zero data retention (ZDR) is an OpenAI control that keeps customer content out of abuse-monitoring logs and makes the API treat the store parameter as false on every request. Only eligible customers get it: OpenAI's documentation says the control requires OpenAI's prior approval and acceptance of additional requirements, arranged through its sales team.

A lighter option sits beside it. Modified Abuse Monitoring also excludes customer content from abuse-monitoring logs, apart from image and file inputs in rare cases, but leaves the full platform available. An approved organization chooses between the two in a Data Retention tab under Settings, Organization, Data controls, for the whole organization or per project.

ZDR has limits worth knowing before you rely on a vendor's claim. Endpoints that are not eligible, such as files, vector stores and Assistants, can still store application state; flagged images are kept for review; and OpenAI reserves the right to make specific models ineligible for a customer, with written notice in advance. The enterprise privacy page describes ZDR as something you request for eligible endpoints when you have a qualifying use case.

For a buyer, the useful question is not whether a vendor says zero retention but which control, on which projects and endpoints, and approved when. Standard API terms are the default for every API customer; a ZDR claim should come with that detail in writing.

Can you choose where chatbot data is stored?

At OpenAI, eligible API customers can pin a project to a region. At the chatbot vendor, the answer depends on where the vendor hosts its own database, which is a separate question with a separate answer.

Data residency in OpenAI's API is a per-project setting that keeps customer content at rest in a chosen region wherever an endpoint needs to store it. OpenAI's documentation lists ten regions: the United States, Europe (EEA and Switzerland), the United Kingdom, Canada, Australia, Japan, India, Singapore, South Korea and the United Arab Emirates. Inference also stays in the region only where OpenAI lists regional processing, such as the United States and Europe.

Residency comes with conditions. Customers must ask OpenAI's sales team about eligibility, any region other than the United States requires approval for abuse-monitoring controls and a Modified Retention amendment, and residency endpoints carry a 10% price uplift for eligible models released on or after 5 March 2026. Residency also excludes system data, such as account details, usage metrics and billing information.

Vendors answer the hosting question very differently. Chatbase, for example, states on its security page that its database and application run on AWS in the us-east-1 region, managed by Supabase and Vercel. PepoChat does not state a hosting region, as the last section explains.

What else touches your chatbot's data besides OpenAI?

Several systems besides the model: the vendor's database, the search index built from your content, email and helpdesk tools, logs, analytics and exports. A 30-day limit at OpenAI means little if another system keeps every transcript forever, so ask about each one.

A sub-processor is a company your vendor uses to process your customers' data on your behalf, such as its cloud host, its model provider or its email service. Under GDPR, a processor may only use sub-processors its customer has authorized, which is why reputable vendors publish the list; our GDPR guide for chatbots explains the controller and processor roles in more detail.

Integrations are the part buyers most often forget. When a bot opens a helpdesk ticket, the ticket can carry a full transcript or only a summary; in PepoChat it carries an AI-written summary and the email the visitor typed, as our guide to creating tickets from chat shows. Once data lands in the helpdesk, the helpdesk's retention rules apply, not the chatbot's.

Ten questions to ask your chatbot vendor about data

Ask about the model, the retention tier, what is sent, where it is stored, for how long, how deletion works and who else processes it. A vendor that answers in writing, with names and dates, is easier to trust than one that answers with a row of badges.

Questions 3, 6 and 7 matter most, because the provider's retention tier and the vendor's own storage location and retention are where long-lived copies sit. A Data Processing Addendum (DPA) is the contract in which a processor commits to handling personal data only on the customer's instructions; OpenAI offers one for API customers, and your vendor should offer one to you.

Some vendors answer several of these on a public page. Chatbase lets you choose models from OpenAI, Anthropic, Google and others, and its security page says it uses the providers' APIs and "doesn't opt-in their data sharing program". So a Chatbase agent running on an OpenAI model sends visitors' messages to OpenAI under API terms, and Chatbase says the providers cannot train on them; the Chatbase alternative page compares the rest of the product.

How does PepoChat handle chatbot data?

PepoChat sends each conversation to OpenAI's API to generate replies and keeps transcripts, contacts and your knowledge base in its own database. The facts below come from PepoChat's product code and documentation as of 19 October 2026.

Models and endpoints. Replies come from OpenAI's gpt-4o-mini through the Chat Completions endpoint, which PepoChat also uses to read retrieved passages, label each conversation's topic from the visitor's first message, power the inbox's Enhance button and transcribe images and scanned PDFs you upload to the knowledge base. Knowledge-base text and search queries are embedded with text-embedding-3-small through the Embeddings endpoint. None of these requests sets a storage or retention option, so they run under OpenAI's standard API terms; PepoChat makes no zero data retention claim.

What is sent and what stays. For each reply PepoChat sends the conversation, the passages retrieved from your knowledge base and your agent instructions, and when an action runs the agent sees only the response fields you allowed for it in custom actions. PepoChat does not use OpenAI's files, vector stores, Assistants or Responses storage: the knowledge base and the conversation history stay in PepoChat's own database. OpenAI's deprecations page lists GPT-4o mini for removal from the API on 14 January 2027, announced on 14 October 2026, so expect the reply model named here to change before that date.

Source: PepoChat's privacy documentation and product code, read 19 October 2026.

Isolation and encryption protect the copies PepoChat keeps. Each workspace's knowledge base, conversations, contacts and credentials sit in separate namespaces, integration keys and OAuth tokens are encrypted with AES-256-GCM, and retrieved content is wrapped as untrusted data in the prompt so an instruction planted on a web page cannot steer the agent. The widget sets no cookies on your site; the security overview lists every control, and our post on AI chatbot security risks puts them in context.

PepoChat's services are Convex for the backend, Vercel for the dashboard and the chat widget, OpenAI for the models, Resend for email, Vapi for voice when you switch it on, and Dodo Payments for billing. PepoChat does not currently claim a hosting region, a signed DPA template or certifications such as SOC 2 or ISO 27001. The privacy documentation suggests listing PepoChat as a processor and the AI model provider as a sub-processor if you operate under GDPR, and asking [email protected] about current status rather than assuming.

What to do next

• Send the ten questions above to every chatbot vendor on your shortlist, and file the written answers with your vendor records.
• Read GDPR and AI chatbots for the controller and processor roles, and build vs buy if you are weighing calling the API yourself, which puts the store and retention choices in your own hands.
• Check PepoChat's privacy and data documentation, then start a free workspace and set Data retention on the Team page before your first real conversation; every feature is on the free plan.