What happens to your data when you use an AI agent

The short answer

When you message an AI agent, your text passes through three other parties: the chat app, the agent platform and the AI model provider. Each may keep a copy under its own rules, so what happens to your data depends less on the AI itself and more on who holds each copy and what their terms say.

How much that should worry you comes down to what you send (a shopping list is not a client's contract), who holds it (a hosted service, your own server and a provider's API are different arrangements), and what each provider's policy says today. Those policies change more often than most businesses check them.

Where does your message go when you use an AI agent?

From your chat app to the agent's server, then to a model provider, and the reply comes back the same way. A typical hosted agent involves four parties:

You, sending a message from Telegram, LINE or a web chat.
The chat app, which carries the message. Telegram and LINE are messaging platforms with their own privacy terms, separate from the AI.
The agent platform, which runs the engine, stores your memory and conversation history, and manages the connection to the model.
The model provider, which processes the message to generate a reply.

Each may store something. The chat app keeps the message thread. The platform keeps conversation history and memory. The provider processes the request, and whether it keeps it, and for how long, depends on the product and the terms. A complete answer to "where does my data go?" names all four.

Do AI providers train on your data?

It depends which product you use. On consumer chat plans your conversations can be used for training unless you opt out (OpenAI) or decline (Anthropic). Business plans and the API aren't used for training by default. The details below were checked against each provider's pages on 24 October 2026.

Consumer plans

These are personal plans such as ChatGPT Free, Plus and Pro, or Claude Free, Pro and Max.

OpenAI says it may use your ChatGPT content to train its models. You can opt out by turning off "Improve the model for everyone" under Settings, then Data controls. Temporary Chats aren't used for training. If you give thumbs-up or thumbs-down feedback, the whole conversation may be used even if you've opted out. Deleted chats are scheduled for permanent deletion within 30 days, with exceptions for security and legal reasons.
Anthropic changed its consumer terms in 2025. Under that update, announced on 28 August 2025 with a deadline of 8 October 2025 for existing users, Claude Free, Pro and Max users choose whether their chats and coding sessions can be used to train models. If you allow it, Anthropic may keep that data for up to five years. If you don't, the 30-day retention period still applies. You can change the choice at any time in Privacy Settings.

Business plans and the API

These are team and enterprise plans, and API access, where a business or developer connects to the model directly.

OpenAI says that by default it doesn't train on data from ChatGPT Business, Enterprise or Edu, or from its API. API data hasn't been used for training since 1 March 2023 unless the customer opts in, and abuse-monitoring logs are kept for up to 30 days by default.
Anthropic says that by default it won't train on inputs or outputs from its commercial products, including Claude for Work and the API, unless you send feedback or choose to allow it. The 2025 consumer change doesn't apply to these products.

Both providers keep some data longer in specific cases, such as content flagged for safety review or where the law requires it.

Why this matters for agents

The same provider can treat your data differently depending on how you reach it. A personal subscription and an API account are separate arrangements.

If an agent runs on your own ChatGPT or Claude sign-in, your consumer account and its training setting generally apply. If it runs on the platform's API access, the provider's business terms apply between the provider and the platform, and the platform's own terms apply to you. Ask which one you're on.

These terms change, so check the pages linked above before relying on them. If a provider's page doesn't answer your question clearly, treat the silence as information.

Does a private server keep your data private?

Partly. A private server keeps your history and memory away from other customers, but your messages still go to the model provider and through the chat app. What it gives you is your own machine with its own storage, shared with nobody. The hosted vs self-hosted guide covers the wider trade-offs.

What a private server improves:

Isolation. Your conversation history and memory aren't in a shared pool where a misconfiguration could expose them to another customer's agent.
Control. You decide what's installed, what's backed up and who has access.
Location. You know which country the server is in, which matters for some data arrangements.
Exit. In principle you can take the data and leave, provided you have access to the machine.

What it leaves unchanged:

The model provider still processes your messages. A private server doesn't make the AI local. Your text still leaves the machine to be processed, which is the point most people miss when they hear "private server" and assume "private AI".
The chat app still carries the message. Telegram and LINE store the conversation on their side too, under their own terms.
Someone has to maintain it. Backups, updates and access control still need doing. A private server nobody patches is less private than a well-run shared service.
It doesn't make you compliant by itself. Data-protection laws, such as the UK and EU GDPR or your local data-protection law, attach obligations to how you collect, use and share personal data. Server location matters for some transfer rules, but it's one control among several.

Which information needs care?

Set the architecture aside and look at the specific information you're about to send:

Is it personal data? Names, contact details, ID numbers, employment details, anything that identifies a person. If so, you have obligations whatever the tool.
Does it need to leave your organisation at all? Much information can be summarised or anonymised first. A client's name is rarely needed for a drafting task.
Who is accountable if it leaks? If the answer is "the platform", check that the contract says so.
How long will it stay? Conversation history and memory persist by default, so ask how long each is kept and whether you can delete it.

One rule covers most cases: send the minimum the task needs and keep the rest out. Regulators call this data minimisation, and it's one of the core principles of the GDPR.

If you're using an agent to plan your day, draft internal notes or think through a problem, the risk is low; a to-do list holds little worth protecting. The calculation changes when client information, staff records, contracts or health details are involved, and that's when the questions above deserve ten minutes.

Send less, and make what you do send easy to clean up. None of these habits costs much.

Don't paste what you wouldn't email. If you'd hesitate to send it to a colleague outside the company, don't send it to an agent.

Use placeholders for identifiers. "Client A" and "the supplier" work as well as real names for most drafting. Swap the names back in at the end.

Send the relevant part. If a document is long and mostly irrelevant, paste the paragraph that matters. You share less and get the same result.

Review memory. Persistent memory means things you said months ago are still stored. Look through it now and then and delete what's out of date or too sensitive to keep.

Keep work and personal separate. Using a personal subscription for client or work material may breach your employer's policy, and possibly a client agreement. See Using AI at work: what to check with your employer.

Know your exit. Before committing to a platform, ask how you'd get your history and memory out. If the answer is vague, assume you can't.

What does the chat app keep?

The messaging app is a separate party, and it's easy to forget. A message you send through Telegram or LINE exists in the app's systems as well as the agent's, and the app's privacy terms (not the AI provider's) apply to that copy.

On Telegram, default chats are cloud chats: encrypted between your device and Telegram's servers and stored there so they sync across your devices. That isn't end-to-end encryption. Secret Chats are end-to-end encrypted, but they exist only on the two devices involved. A chat with a bot, which is how an agent lives in Telegram, passes through Telegram's servers to reach the bot's owner, so it isn't end-to-end encrypted either. The new to Telegram guide explains the app itself.

So if a piece of content would be a problem sitting in a messaging platform's storage, it's a problem before the AI ever sees it, and a private server for the agent doesn't change what the chat app holds. Chat apps are fine for an agent; be deliberate about what you send through them.

Questions to ask a provider

Where are conversations and memory stored, and in which country?
Are they encrypted at rest and in transit?
Which model providers do you use, and on which terms: consumer sign-in or business API?
Can I see and delete my agent's memory?
What happens to my data if I cancel?
Who inside your company can access my data?
Have you had a breach, and how would you tell me if you did?

Clear, specific answers tell you something about how a provider operates. So does a link to a generic security page in place of answers. The AI agent safety guide has more questions on limits and approvals.

No setup makes data risk zero. Any system that processes text keeps copies in several places, and policies can change after you read them. A reasonable standard is to send less, know where it goes, keep retention short and recheck the terms now and then. Be wary of anyone promising more certainty than that.

Common questions

Does ChatGPT use my conversations for training?

On personal plans (Free, Plus and Pro), OpenAI may use your conversations to train its models unless you turn off "Improve the model for everyone" in Data controls. Temporary Chats aren't used. On ChatGPT Business, Enterprise and Edu, and through the API, OpenAI says it doesn't train on your data by default. On personal plans, thumbs-up or thumbs-down feedback can still share a conversation for training.

Does Claude use my chats to train its models?

On Claude Free, Pro and Max, it depends on the choice you made under Anthropic's 2025 consumer terms. If you allowed training, your chats can be used and kept for up to five years. If you didn't, they aren't used, apart from feedback you send. On Claude for Work and the API, Anthropic says it doesn't train on your data by default.

It's safer to leave it out unless the task needs it. Personal details can end up stored by the chat app, the platform and the model provider, each under different rules. If you must include personal data, use a business account with agreed terms and check your obligations under data-protection law.

Can I delete my data from an AI service?

Usually, yes. Both OpenAI and Anthropic let you delete chats and say deleted chats are removed from their systems within about 30 days, with exceptions for safety and legal reasons. Changing a training setting stops future use but doesn't undo training that has already happened. An agent platform may hold its own copy of your history and memory, so ask it separately.