AI tools and diligence files

What happens when buyers paste diligence files into ChatGPT?

The file leaves the data room’s controls and goes to the AI provider, under whatever terms cover that account. Many consumer plans can use chats to improve models unless a setting is off; business plans generally don’t by default. The NDA’s limits still apply, so use business-covered tools, redacted copies and a written policy.

Last updated: 2026-11-09

Key points

  • Once a file is downloaded, the data room can’t see, limit or recall what happens to it. The AI account it lands in decides what happens next.
  • The plan matters more than the brand. The same AI company can train on a personal account’s chats and not on a business account’s.
  • NDAs usually limit use to evaluating the deal, sharing to named kinds of advisers, and require return or destruction. Check how an AI account fits each of those.
  • The safer pattern: a business plan on work accounts, redacted copies for anything with names or account numbers, and AI output treated as a draft.
  • Sellers and brokers can’t control downloads after the fact, so they ask about AI at the NDA stage and limit what leaves the room.

What happens to a file you paste into an AI tool?

A copy goes to the AI provider’s servers, is kept in the chat history and, depending on the plan, may be kept longer, read by reviewers or used to improve models. None of the data room’s controls follow it.

It usually starts with a good reason. A buyer’s analyst has three years of monthly P&Ls, a customer revenue report and twenty service contracts, and a call with the seller on Thursday. Pasting them into a chatbot and asking for a trend table and a list of renewal dates takes ten minutes. Here is what that does:

  • The data room’s limits stop at the download. An access end date, a revoked login or a stage the buyer hasn’t reached does not reach a copy in another service.
  • The record stops too. The room’s activity log shows the download, by whom and when. It cannot show where the file went next.
  • The account’s terms take over. Whether the text trains a model, how long it is kept and whether a person at the provider may read it depend on the plan and its settings.
  • The copy outlives the deal. If the deal dies, the NDA usually asks for information to be returned or destroyed. A chat history in someone’s personal account is easy to forget.

None of this means AI tools are off limits for diligence. It means the choice of tool and account is part of handling the seller’s confidential information, the same as choosing where to save the files.

Do consumer AI plans train on what you paste?

Often they can, unless a setting is turned off. Business plans from the same providers generally don’t train on your data by default. Check the exact plan, because the brand name tells you nothing.

What the providers we could check say on their own pages, read on 2026-11-09:

Whether AI plans train on your chats by default, checked 2026-11-09
Tool and plan Account type Trains on your chats?
Google Gemini app3 Personal Yes, while Keep Activity is on
Gemini in Google Workspace4 Business (qualifying editions) No, without the customer’s permission
Claude Free, Pro and Max1 Personal Only if the user’s setting allows it
Claude for Work and the Anthropic API2 Business No, by default
Microsoft Copilot Chat with a work account5 Business No (foundation models)
OpenAI API6 Business No, since March 1, 2023

Retention differs as much as training. Anthropic says consumer chats are kept for five years if the user allows training, and 30 days if not. Google says Gemini app activity is deleted after 18 months by default, that chats are kept for 72 hours even with Keep Activity off, that a subset of chats is read by human reviewers, and that reviewed chats are kept for up to three years. Google’s own advice is not to enter confidential information you wouldn’t want a reviewer to see.

What about ChatGPT?

OpenAI publishes separate data terms for ChatGPT’s personal plans and for its business plans, and an enterprise privacy page and a help article on model training describe them. Those pages refused automated reading when we checked on 2026-11-09, so we describe ChatGPT only in general terms here: treat a personal ChatGPT account the way you would any consumer plan, find its training setting before anyone pastes a deal file, and read OpenAI’s pages for the exact plan your team pays for. For developers, OpenAI’s API documentation does load, and it says API data has not been used for training since March 1, 2023.

How is a business AI plan different?

A business plan puts the provider under a contract with your firm, usually with training off by default, retention you can see and an administrator who controls who uses it. It only helps if everyone actually signs in with the work account.

Personal and business AI plans compared, in general practice
Question Personal plan Business plan
Who agreed to the terms? The individual user The firm
Training on chats by default Often possible, through a setting Generally off
Who can see the chat history? The user The user, plus admin tools the firm controls
Can the firm delete it at deal end? Only by asking the user Usually, through admin settings
Data processing agreement None Usually offered

Three gaps remain even on a business plan:

  • Personal logins. An analyst who uses the phone app signed in with a personal email is on the personal terms, whatever the firm pays for.
  • Web search and connectors. Features that reach outside the tool can send text under different terms. Microsoft, for example, says web queries from Copilot go to its Bing search service, which has its own data-handling terms, with user and tenant identifiers removed.
  • Feedback buttons. Anthropic says a thumbs-up or thumbs-down can make a conversation available for training even where training is otherwise off, and Google says the same about feedback on Gemini app chats.

Does using AI on diligence files fit the NDA?

Read three parts of the NDA: what the information may be used for, who may receive it, and what happens to it at the end. A personal AI account fits poorly with all three; a business tool under contract fits more like the buyer’s own software.

Lower-middle-market NDAs vary, but most contain the same few parts. Here is how an AI tool lines up against each, in plain terms:

Common NDA terms and the AI question each raises
What many NDAs say The AI question it raises
Use only to evaluate the deal Does training a provider’s model count as another use?
Share only with “Representatives” (employees, advisers, lenders) who need to know Is an AI provider one of them, or a third party?
The buyer answers for its Representatives Who answers for what an analyst did in a personal account?
Return or destroy on request Can every chat and upload be found and deleted?
Tell the seller about any unauthorized disclosure Does a paste into a personal account need to be reported?

Older NDAs rarely mention AI at all, which leaves both sides guessing. The practical fix is to settle it before diligence starts. A buyer can tell the broker which AI tools the team uses and under what terms; a seller can add a line to the NDA naming what is allowed. Either way, the answer is written down before anyone pastes a file.

What are safer ways to use AI on diligence files?

Match the workflow to the document. Use a business-covered tool for originals, a redacted copy for anything with names or account numbers, and your own extracted numbers for financial analysis.

Safer AI workflows for diligence files
Workflow What goes into the AI Good for Watch for
Business plan, work account Original documents Contract and lease term tables Personal logins on phones
Redacted copy A copy with names and numbers removed Customer and employee files Text left under black boxes
Extracted numbers Your own spreadsheet of figures Trends, margins, ratios Typing errors in the extract
Totals only Percentages and sums, no names Customer concentration Losing detail you need later
The data room’s own search Nothing leaves the room Finding documents It finds; it doesn’t analyze

How to make a redacted copy for analysis

  1. Start from a copy. Never edit the file in the data room.
  2. Replace each customer name with a code (Customer 01, Customer 02) and each employee name with a role (Technician 4).
  3. Remove Social Security numbers, bank and card account numbers, home addresses, phone numbers and dates of birth.
  4. Remove the text itself. A black rectangle drawn over a PDF can leave the text underneath readable to an AI. How proper redaction removes the value.
  5. Keep the key that maps codes to names in a separate file that never goes into any AI tool.
  6. Spot-check the copy by searching it for a few real names before you use it.

How to analyze numbers without uploading the statements

For a quality-of-earnings style review of trends, the AI doesn’t need the bank statements. Pull the monthly revenue, cost of goods sold, payroll and owner compensation into a spreadsheet you build, check it against the trailing-twelve-month P&L, then ask the AI to work on the spreadsheet. The result is easier to check, and nothing identifying leaves your hands.

Whatever the workflow, the AI’s output is a draft. A renewal date pulled from a contract, a margin trend or a list of change-of-control clauses gets checked against the source before anyone relies on it. Where AI analysis tools help and where they fail.

What should a deal team’s AI policy say?

Which tools are approved, which accounts may be used, which documents may go in and in what form, how output is checked, and how every copy is deleted at the end. One page is enough.

A policy a buyer team, a search fund or a small PE firm can copy and adapt:

  1. Approved tools. Only [named business AI plan], signed in with a firm account. No personal AI accounts for deal work, on any device.
  2. Settings. Training on our data is off, and the admin has confirmed it. Web search and outside connectors are off for deal workspaces.
  3. Originals allowed. Contracts, leases, corporate records, permits and financial statements without personal data.
  4. Redacted copies only. Customer lists, payroll and census files, anything with Social Security or account numbers, and anything the NDA marks as highly confidential.
  5. Never. Documents the seller has said may not leave the data room, and anything we can view but not download.
  6. Output. AI output is a draft. The person who uses it checks it against the source document and is responsible for it.
  7. Labeling. Files made with AI help are named as drafts and stored in our deal folder, never uploaded back into the seller’s data room.
  8. Advisers. Our accountants, lawyers and lenders follow the same rules or their own written equivalent.
  9. End of deal. When the deal closes, dies or the NDA asks, we delete deal chats, uploads and projects, and the deal lead confirms it in writing.
  10. Mistakes. Anyone who puts deal files in the wrong place tells the deal lead the same day.

A seller or broker can use the same list in reverse: it is the set of questions to ask a buyer before granting downloads.

What can sellers and brokers do?

Ask about AI before the NDA is signed, and limit what can leave the room in the first place. After a download, the seller has no control.

  • Ask at the NDA stage. Which AI tools does the buyer’s team use, on which plan, with training on or off? Put the answer in the NDA or an email.
  • Release in stages. Teaser, then the CIM after the NDA, then the detailed files after the LOI. A buyer who never receives the payroll file can’t paste it anywhere.
  • Keep downloads off by default. Allow them person by person, for the buyer’s accountant or lender, rather than for the whole buyer team.
  • Remove sensitive numbers first. Social Security numbers and full account numbers have no place in diligence copies, AI or not.
  • Watermark what buyers view. A name and time on every page reminds the reader whose copy it is.
  • End access on a date. When the deal moves on, so does the buyer.

Every other way deal data leaks, and the control for each: how deal data leaks.

Worked example: an analyst and a customer list

The firm, people and figures below are invented to show the steps. They are not a client story.

Tessa Marwood runs a search fund and has signed an LOI to buy Bluegrass Turf & Grounds, a commercial landscaping company with $4.8M of revenue, for $3.9M with an SBA loan. Her analyst, Jonah Pryce, downloads the customer revenue report (top 40 customers by name, three years) and pastes it into the personal AI account on his phone to build a churn table before a call with the seller.

  1. Day 1: noticed. Tessa sees the table in the call notes and asks where it came from. Jonah’s account has the training setting on.
  2. Day 1: contained. Jonah deletes the chat and the upload, turns the training setting off and lists exactly what he pasted. Tessa reads the NDA’s Representatives and notice clauses.
  3. Day 1: told. Tessa emails the seller’s broker: what was pasted, into what kind of account, when it was deleted, and what the team is changing. The broker forwards it to the owner.
  4. Day 2: fixed. The team moves to a business AI plan with training off, signed in with firm accounts on every device, and adopts the ten-point policy above.
  5. Day 3: redone the right way. Jonah builds the churn table again from a redacted copy: Customer 01 to Customer 40, revenue by year, no names. The key stays in a spreadsheet that never goes into the AI.

The deal closes on schedule. The broker adds an AI line to the NDA for the next listing, and asks every buyer at the NDA stage which AI tools they use.

What mistakes do deal teams make with AI tools?

Most are account mistakes, not AI mistakes: the right tool used through the wrong login, or the right setting assumed rather than checked.

  1. Using a personal account for deal work. Fix: firm accounts only, on every device.
  2. Assuming the firm’s plan covers everyone. Fix: the admin confirms who is signed in to what.
  3. Uploading the whole data room export. Fix: send only the documents a question needs.
  4. Trusting an opt-out alone. Retention, reviewers and feedback buttons still apply. Fix: read the plan’s retention terms too.
  5. Forgetting web search and connectors. Fix: switch them off for deal work.
  6. Covering text instead of removing it. Fix: redact the value, then search the copy for it.
  7. Treating AI output as a finding. Fix: check every figure and date against the source.
  8. Leaving copies behind. Fix: a dated deletion step at deal end.
  9. Sellers saying nothing. Fix: ask the AI question before the NDA is signed.

How ClosingVault handles this

ClosingVault keeps the AI inside the room and limits what leaves it. Its AI organizes, never opines: it files and finds documents and never writes text.

  • No chatbot, no free AI. ClosingVault never sends your documents to a free AI chatbot. It uses one contracted AI service, TypeSafe, which doesn’t train on customer data, is hosted in the US and is a SOC 2 Type II provider; one customer’s data cannot reach another customer. TypeSafe receives extracted text with long account-style numbers masked to the last four digits, never the file itself.
  • Finding in the room. Ask the Data Room returns documents, not prose, and only those the person asking is allowed to see.
  • Downloads person by person. Everyone else reads in a viewer watermarked with their name and the time, with printing and copying suppressed.
  • Redaction on by default. Social Security numbers and ITINs are removed; bank, loan and card numbers keep only their last four digits; IBANs and labeled driver’s license, passport and state ID numbers and dates of birth are removed. How redaction works.
  • Access that ends. Buyer access always ends: two weeks by default, two months at most. The activity log records every download, by whom and when.
  • Advisers by name. A buyer asks to add their accountant, lawyer or lender, and the room admin approves each one.

Coming soon: an MCP connection, so an analysis model a buyer’s firm already trusts can read the room’s documents directly, within that person’s permissions. It isn’t available yet. ClosingVault itself is not SOC 2 certified; SOC 2 Type II is on the roadmap.

Questions about AI tools and diligence files

Is it OK to put due diligence documents into ChatGPT?

Not into a personal account. Use a business plan your firm controls, signed in with a work account, and check its training and retention settings first. For customer lists, payroll and anything with names or account numbers, work on a redacted copy or on numbers you extracted yourself.

Does ChatGPT train on what I paste into it?

It depends on the plan and its settings. Many consumer AI plans can use chats to improve models unless a setting is turned off, while business plans generally don’t train on your data by default. Read OpenAI’s data terms for the exact plan your team uses before anyone pastes a deal file.

Is a business AI plan safe for confidential deal documents?

It is much safer than a personal plan, because the provider generally doesn’t train on the data by default and a contract covers it. It is only as safe as its use: everyone must sign in with the work account, and features such as web search or connectors may send text elsewhere under other terms.

Does pasting diligence files into an AI tool break the NDA?

Read what the NDA says about purpose, who may receive the information and returning or destroying it. A personal AI account that may train on chats or keep them for years sits poorly with those terms; a business tool under contract looks more like the buyer’s own software. Many deal teams settle it in writing at the NDA stage.

How should we redact documents before using AI on them?

Replace customer and employee names with codes, remove Social Security numbers, account numbers and home addresses, and keep the key that maps codes to names in a separate file that never goes into the AI. Remove the text itself, not just cover it with a black box.

What should a buyer team’s AI policy say?

Which AI tools are approved, that only work accounts may be used, which documents may go in as originals and which only as redacted copies or extracted numbers, that AI output is a draft someone checks, and that every copy is deleted when the deal ends or the NDA requires it.

Can a seller stop buyers from using AI on its documents?

Not once a file is downloaded, but a seller can limit downloads, release documents in stages, remove sensitive numbers before buyers see them, watermark each page with the viewer’s name and ask buyers at the NDA stage which AI tools they use and under what terms.

Does ClosingVault send documents to ChatGPT?

No. ClosingVault never sends your documents to a free AI chatbot. It uses one contracted AI service, TypeSafe, to file and find documents; TypeSafe doesn’t train on customer data, is hosted in the US and is a SOC 2 Type II provider.

Sources

  1. Anthropic, “Updates to Consumer Terms and Privacy Policy”, August 28, 2025: the training choice for Claude Free, Pro and Max, and five-year or 30-day retention: anthropic.com, checked 2026-11-09.
  2. Anthropic Privacy Center, “Is my data used for model training?”: commercial products not used for training by default, and the feedback exception: privacy.claude.com, checked 2026-11-09.
  3. Google, Gemini Apps Privacy Hub: training while Keep Activity is on, human review, 18-month and 72-hour retention, and the advice on confidential information: support.google.com/gemini/answer/13594961, checked 2026-11-09.
  4. Google, Generative AI in Google Workspace Privacy Hub (last updated October 6, 2026): knowledge.workspace.google.com, checked 2026-11-09.
  5. Microsoft Learn, “Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat”: no training of foundation models, and web queries to Bing under separate terms: learn.microsoft.com, checked 2026-11-09.
  6. OpenAI, “Your data” (API data controls): developers.openai.com/api/docs/guides/your-data, checked 2026-11-09.
  7. OpenAI, enterprise privacy page and ChatGPT help article on model training: openai.com/enterprise-privacy and help.openai.com. Both refused automated reading on 2026-11-09 and could not be confirmed; ChatGPT is described above only in general terms.

Last updated: 2026-11-09

How this guide was written: it reflects general lower-middle-market practice, is not legal or tax advice, and was checked against the sources cited above.