1. It depends on the type of service and its terms
Some AI providers publish terms for business and API use under which inputs are not used for training by default. Anthropic says that for its commercial products, such as Claude for Work and the Anthropic API, it does not use inputs or outputs to train its models by default. It adds that it may use them if a customer sends feedback such as a bug report, or otherwise chooses to allow it. It also states that its consumer products, Free, Pro and Max, are handled separately, so the same AI can have different terms depending on the type of contract (Anthropic Privacy Center, page dated August 18, 2026).
OpenAI also states that, since March 1, 2023, data sent to its API is not used to train or improve its models unless the customer explicitly opts in to share it. Separately, abuse-monitoring logs that can contain prompts and responses are kept by default for up to 30 days. Only customers approved in advance by OpenAI can be excluded from those logs (OpenAI developer documentation). In other words, not being used for training and not being stored anywhere are two different things.
Our view We think there is a real difference in risk between staff pasting documents into free personal accounts and a company using a business contract or API after checking its terms. If a company decides nothing, the cost is that staff paste quotations and customer lists into personal accounts even where this is prohibited, and no one can later trace which document went to a service under which terms. A competitor that sets up a business contract and internal rules can also answer the information-handling questions its customers ask. We recommend first listing the AI services in use, with their contract type and their terms on training use and log retention. It is safer to make that list after confirming whether staff are using personal accounts.
2. If personal data is involved, check the law first
On June 2, 2023, Japan's Personal Information Protection Commission issued a notice on the use of generative AI services. It asks businesses that enter prompts containing personal information to confirm that this stays within what is necessary to achieve the specified purpose of use. It also says that entering personal data in a prompt without the person's consent, where the data is handled for any purpose other than producing the response, may violate the Act on the Protection of Personal Information. For that reason, it asks businesses to confirm thoroughly, among other things, that the service provider does not use the personal data for machine learning (Personal Information Protection Commission, in Japanese).
Documents that contain customer lists, employee records or the names of contacts at business partners can fall under this check. In practice, before handing a document to an AI, check where personal information appears, and either mask what is not needed or do not hand over that document at all.
3. How our AI back-office assistant handles documents
The page for our AI back-office assistant states that originals stay on your own equipment and are not stored on our servers. On the standard plan, only the text needed to generate a draft is sent to an external AI. The fully in-house plan is 29,800 JPY per month, excluding tax, and the AI itself also stays on your premises, so drafting is completed in-house. The standard plan is 4,980 JPY per month, excluding tax, for up to three users, and usage fees for the external AI are paid by you directly to the AI provider. A staff member checks each draft before it is sent, so AI output never leaves the company as it is.
Our view Whether to limit what is sent to the text needed for a draft or to choose the fully in-house plan can be decided by whether the documents contain business partner names and unit prices, or customers' personal information. We think the order is to consider the fully in-house plan first for quotations that contain partner names and prices, and to check whether the standard plan is enough for reply drafts from which partner names can be masked. This assumes that you confirm, at the sample stage before adoption, what the text sent to the external AI contains. Starting to send text out without that check means going back later to see whether a confidentiality promise to a partner was affected.
4. What to confirm before you start
- Whether the contract excludes training use: personal or business plan, and whether sending feedback creates an exception.
- Log retention: OpenAI's API keeps abuse-monitoring logs for up to 30 days by default, and turning this off needs prior approval.
- Whether personal information is involved: if so, check that it stays within the purpose of use and that the provider does not use it for machine learning.
- Where originals are kept: whether only text reaches the AI provider or the files themselves.
- Personal accounts: put in writing which services the company allows and which kinds of documents must never be pasted.
Summary
Whether data leaks depends not on the type of AI but on the contract, the handling of logs and the scope of what you hand over. Some providers publish terms that exclude training use, but logs may still be kept separately, and legal checks are needed when personal information is involved. To discuss our AI back-office assistant before adoption, use the contact page.
Questions this article answers
If AI reads our internal documents, could they be used for training and leak?
Some providers do not use inputs for training by default on business products and APIs, as Anthropic does for its commercial products and OpenAI does for its API. Feedback or the customer's own permission creates an exception, and consumer services have different terms. Check the contract type and the training and log terms before you start.
Is it all right to give AI documents that contain customers' personal information?
The Personal Information Protection Commission says that entering personal data in a prompt without the person's consent, where it is handled for a purpose other than producing the response, may violate the Act on the Protection of Personal Information. Before entering it, confirm it is within the purpose of use and that the provider does not use it for machine learning. Mask what is not needed.
Sources
- Anthropic Privacy Center, Is my data used for model training? (2026-08-18)privacy.anthropic.com
- OpenAI, Data controls in the OpenAI platformplatform.openai.com
- 個人情報保護委員会 生成AIサービスの利用に関する注意喚起等について（2023-06-02）ppc.go.jp
Related services
If you want to try this in your own business
To see how the topics in this article would apply to your own operations, and what it would cost, send us a message through the contact form.