Skip to main content
Published in Engineering Insights

Does Claude Train on Your Company's Data? Plan by Plan

Does Claude Train on Your Company's Data? Plan by Plan
Table of contents
  1. Does Claude train on data from Team and Enterprise plans?
  2. What do Anthropic's Commercial Terms say about training?
  3. Does the Claude API train on what we send?
  4. Do Claude Free, Pro, Max and Ultra plans train on your chats?
  5. Which Claude plans can be used for training, and where does it say so?
  6. Does Claude Code train on our source code?
  7. What about records Claude reads through connectors and MCP servers?
    1. How Gnar connects Claude to company systems
  8. Does Claude on Amazon Bedrock train on our data?
  9. What if staff use personal Claude accounts for company work?
  10. What should a CISO settle before approving Claude company-wide?

Anthropic says it does not train its models on Claude Team, Enterprise or API data by default. Personal Claude Free, Pro, Max and Ultra accounts are different: Anthropic trains on them when the account's training setting is on. Even with it off, Anthropic's Consumer Terms say feedback and material flagged for safety review are used for training.

So the answer to "does Claude train on my data" depends on the plan each person is signed in to. On business plans, it also depends on choices the company makes. Two examples are whether users can send feedback and whether the company opts in to an Anthropic program that shares data.

This guide is written for a CISO who has been asked to approve Claude for the whole company.

The Gnar Company is a Claude Select Services Partner in Anthropic's Claude Partner Network. It is a custom software firm with U.S.-based senior engineers, who use Claude in their own work. The Gnar Company's enterprise AI integration page says: "We connect it to the systems where they already live."

Does Claude train on data from Team and Enterprise plans?#

No, Anthropic says it does not train on Team or Enterprise data by default. Anthropic's Privacy Center article for commercial products says: "By default, we will not use your inputs or outputs from our commercial products" to train its models. The article gives Claude for Work, the Anthropic API and Claude Gov as examples of those products.

Claude for Work covers the Team and Enterprise plans. Anthropic's article on processors and controllers describes "a Claude for Work account (Team or Enterprise plan)."

The default has an exception, and it sits in the next line of the same article. When a user explicitly reports feedback or bugs, or otherwise chooses to allow it, Anthropic says it "may use your chats and coding sessions to train our models." The article's example of feedback is the thumbs up or down button.

Feedback sends Anthropic more than the rating. The article says a thumbs up or down stores "the entire related conversation" in Anthropic's back end for up to 5 years. It also says Anthropic de-links feedback from user and customer IDs before Anthropic uses it.

A company can turn that button off. Anthropic's article says a Primary Owner or Owner of a Team or Enterprise plan can stop members from sending feedback through the thumbs up or down button. The control is the Rate chats setting, found under Organization settings and then Data and Privacy.

What do Anthropic's Commercial Terms say about training?#

The Commercial Terms put it in one sentence: "Anthropic may not train models on Customer Content from Services." The Commercial Terms of Service define Customer Content as Inputs and Outputs together.

Inputs are submissions to the Services by the customer or its users. Outputs are the responses the Services generate to those Inputs.

The training line names no exception. The feedback case is stated in Anthropic's commercial training article, quoted in the section above. The Terms treat feedback in a separate clause, A.3: "If Customer provides (in its sole discretion) Anthropic with feedback regarding the Services, Anthropic may use that feedback at its own risk and without obligation to Customer."

The Terms say they govern a customer's use of Anthropic API keys and other Anthropic offerings that reference them. Anthropic's processor article places Team and Enterprise accounts under the Commercial Terms, where "the customer is the 'Controller' of the data submitted by its Users."

Note the version you read. As of November 2026, the Commercial Terms show an effective date of Jun 17, 2025. The commercial training article shows a date of August 18, 2026. A reviewer who records both dates can tell later whether either page has changed.

Does the Claude API train on what we send?#

No, Anthropic says it does not use API inputs or outputs for training by default. The commercial privacy article covers the Anthropic API along with Claude for Work. So the same default and the same feedback exception apply to it.

Anthropic also documents an opt in route for API customers: the Development Partner Program. Its help article says the program lets an organization "voluntarily choose whether you allow us to use your Claude Code sessions to improve our services, including model training."

In the help article's words:

  • What is shared: "only your Claude Code input and output tokens from the first-party Claude API are provided to us."
  • Who it covers: "This setting applies to all members in your organization."
  • How long it is kept: "Your data will be stored securely for up to two years."
  • What happens on leaving: "You can leave the program at any time, but previously provided data will be stored for up to two years."
  • Who cannot join: "Accounts on our zero data retention agreement are not eligible for the program."

On prepaid billing, an admin who is authorized to enter into agreements joins from the Data retention page in the Claude Console. Organizations on invoice billing are told to contact their Anthropic account executive about eligibility.

The setting covers every member, so treat joining as a company decision. Put it on the same list as the Rate chats setting.

Do Claude Free, Pro, Max and Ultra plans train on your chats?#

Yes, Anthropic trains on those accounts when the person has the training setting on. Anthropic's Claude Code data usage page says: "We will train new models using data from Free, Pro, Max, and Ultra accounts when this setting is on (including when you use Claude Code from these accounts)."

Anthropic's article on changing it says to open Settings, then Privacy, and use the toggle under "Help Improve our AI models."

Anthropic's Consumer Terms say that even if you opt out, "we will use Materials for model training" in two situations. One is when you provide Feedback. The other is when your Materials are flagged for safety review "to improve our ability to detect harmful content, enforce our policies, or advance our safety research."

Turning the setting off works going forward. Anthropic's setting article says it "will not use any new chats and coding sessions you have with Claude for future model training." It also says your data "will still be included in model training that has already started and in models that have already been trained."

Retention changes with the setting too. Anthropic's consumer retention article says that if you allow training, it "may retain your data in a de-identified format for up to 5 years in our model training pipelines." It says that period "only applies to new or resumed chats after the model training setting has been enabled."

The same retention article says: "Your Incognito chats are not used to improve Claude, even if you have enabled Model Improvement in your Privacy Settings."

Which Claude plans can be used for training, and where does it say so?#

Anthropic says it trains on personal Free, Pro, Max and Ultra accounts when the training setting is on, and does not train on Team, Enterprise or API data by default. The table puts each plan next to the page that states its answer.

Plan or route What the page says about training Page
Claude Free, Pro, Max and Ultra Used to train new models when the training setting is on; feedback and flagged material are used even when it is off Claude Code data usage docs and Consumer Terms
Claude Team and Enterprise Not used by default; feedback a user sends can be used Privacy Center, commercial
Anthropic API Not used by default; Claude Code tokens can be shared through the Development Partner Program Privacy Center and help center
Claude on Amazon Bedrock AWS and model providers will not use inputs or outputs to train third party models AWS Bedrock FAQ

The first three rows come from Anthropic. The last row comes from AWS, which runs Bedrock. Read each row against its own page before you quote it in a policy.

Does Claude Code train on our source code?#

Claude Code follows the training rules of the account the developer signs in with. The Claude Code data usage page lists commercial users as "Team and Enterprise plans, API, 3rd-party platforms, and Claude Gov." For them, it says "Anthropic does not train generative models using code or prompts sent to Claude Code under commercial terms."

The same sentence carries its condition: "unless the customer has chosen to provide their data to us for model improvement (for example, the Developer Partner Program)." The page's next section is headed Development Partner Program, the program described above.

A developer on a personal Pro or Max account is on the consumer side of that page. For those accounts, the page says training happens "when this setting is on," including Claude Code use.

The page says that if you send feedback with the /feedback command, Anthropic "may use your feedback to improve our products and services." It says transcripts shared through /feedback, /bug or /share "are retained for 5 years." To opt out, it says to "set the DISABLE_FEEDBACK_COMMAND environment variable to 1."

The page adds one limit on the program. It says the Development Partner Program "is available only for Anthropic first-party API, and not for Amazon Bedrock or Google Cloud's Agent Platform users."

For a company, the practical question is which account each developer's terminal is signed in to.

What about records Claude reads through connectors and MCP servers?#

Anthropic's training articles leave raw connector content out of the data they describe, though records copied into a chat may be included. Its commercial article says feedback data "does not include raw content from connectors (e.g. Google Drive), including remote and local MCP servers."

The MCP documentation describes MCP as "an open-source standard for connecting AI applications to external systems." The Gnar Company's integration page defines an MCP server as "a connector that lets an AI model use one of your systems."

The rule is worded per plan. On consumer plans, Anthropic's consumer training article says the chat data it may use "does not include raw content from connectors (e.g. Google Drive), including remote and local MCP servers." On commercial plans, the commercial article says the same of feedback data.

Both articles then add the same condition. Data "may be included if it's directly copied into your conversation with Claude."

How Gnar connects Claude to company systems#

The Gnar Company's integration page says this about a company's AI: "We connect it to the systems where they already live. You choose what it can see and change, and every lookup and change is recorded."

In an AI integration built by The Gnar Company, you set access one system and one action at a time. The Gnar Company can build it in your own account with a cloud provider, such as Amazon Web Services (AWS). You own what gets built, and Gnar keeps no ownership of it.

Start with Gnar Discovery and Design, from two weeks and $15,000. You leave with a build plan that has a fixed price on it. The Discovery page lists the tiers.

Does Claude on Amazon Bedrock train on our data?#

AWS says it does not, for any third party model on Bedrock. The Amazon Bedrock FAQ answers the question directly: "No, AWS and the third-party model providers will not use any inputs to or outputs from Amazon Bedrock to train Amazon Nova, Amazon Titan, or any third-party models."

The same FAQ says "Users' inputs and model outputs are not shared with any model providers." It also says customer content processed by Bedrock "is encrypted and stored at rest in the AWS Region where you are using Amazon Bedrock."

Anthropic's Claude Code page counts "3rd-party platforms" among its commercial users. It also says the Development Partner Program is not available to Amazon Bedrock users.

For a company already on AWS, this is a separate route with its own terms. Quote AWS's page for Bedrock, and Anthropic's pages for Anthropic's own plans.

What if staff use personal Claude accounts for company work?#

Company material typed into a personal Free, Pro, Max or Ultra account follows the consumer rules. So it can be used for training when that person's training setting is on. The setting belongs to the person who holds the account.

The Gnar Company's Claude partner page says, under "Why self-implementation fails": "Teams picked it up on their own, in isolated pockets, and now there's a pile of experiments that never talk to each other." The same page names the cost: "Nobody can see where data is going or who approved it to go there."

For Enterprise customers, Anthropic's Enterprise plan page lists "network-level access control, which prevents access to personal or non-corporate Claude instances from enterprise networks."

What should a CISO settle before approving Claude company-wide?#

Before approving Claude for the whole company, settle which plan each group uses and who owns each training setting. This list is general advice. Each item points to a page quoted above:

  1. The plan for each group. Team, Enterprise, the API, or a personal Free, Pro, Max or Ultra account. The answer to the training question changes with each one.
  2. The Rate chats setting and the Claude Code feedback command. Record whether Rate chats is on or off, and which Owner decided. Record whether developers set DISABLE_FEEDBACK_COMMAND to 1.
  3. The Development Partner Program. Record whether any API organization has joined, since it applies to all members and keeps data for up to two years.
  4. The route for Claude Code. A business seat, the first party API, a cloud platform such as Amazon Bedrock, or a personal account.
  5. How company records reach Claude. By connector, or by people pasting them into chats.
  6. The page dates. Record the date shown on each page you relied on, so the next review can see what changed.

Put the answers where the board can see them. Our AI governance framework post lists what a board approves and what management runs.

Frequently asked questions

Does Claude Team train on our company's data?

No, not by default. Anthropic's Privacy Center says it will not use inputs or outputs from its commercial products, such as Claude for Work, to train its models by default. Claude for Work means the Team and Enterprise plans. The exception is data a user chooses to share, such as thumbs up or down feedback. A Primary Owner or Owner can turn off the thumbs up or down button for members with the Rate chats setting.

Does Claude Pro train on my chats?

Yes, when the training setting is on. Anthropic's Claude Code docs say it will train new models using data from Free, Pro, Max and Ultra accounts when this setting is on. You change it under Settings, then Privacy, at Help Improve our AI models. The Consumer Terms say Anthropic will still use feedback, and material flagged for safety review, for model training even if you opt out.

Does the Claude API train on the prompts we send?

No, not by default. Anthropic's commercial privacy article covers the Anthropic API and says inputs and outputs are not used for training by default. An organization can opt in to the Development Partner Program, which shares only Claude Code input and output tokens from the first party Claude API. Accounts on a zero data retention agreement are not eligible.

If I turn the training setting off, is my old data still used for training?

Yes, in training that has already started. Anthropic says your data will still be included in training that has already started and in models that have already been trained. It also says that after you turn the setting off, it will stop using your stored chats and coding sessions in future training runs. Its Consumer Terms still allow training on feedback and material flagged for safety review.

Does Claude on Amazon Bedrock train on our data?

No, according to AWS. The Amazon Bedrock FAQ says AWS and the third party model providers will not use any inputs to or outputs from Amazon Bedrock to train Amazon Nova, Amazon Titan or any third party models. It also says inputs and outputs are not shared with any model providers.

Is The Gnar Company an Anthropic partner?

Yes. The Gnar Company is a Claude Select Services Partner in Anthropic's Claude Partner Network. It is a custom software firm with U.S.-based senior engineers, who use Claude in their own work. Start with Gnar Discovery and Design, from two weeks and $15,000. You leave with a build plan that has a fixed price on it.

Stay In The Loop

Never miss an update. Get the latest blogs, webinars, and expert insights straight to your inbox from The Gnar.