Can ChatGPT answer from my company's technical documents?
Summary
Only if the documents are attached to the chat, kept in a project, or connected through company knowledge on ChatGPT Business, Enterprise or Edu. Otherwise it answers from general knowledge, which holds neither your datasheet revision nor your customer history. Even then, trust only answers that cite the exact document and revision.
OpenAI introduced company knowledge on 2025-10-23 for ChatGPT Business, Enterprise and Edu, with citations in every response (OpenAI, 2025).
Company knowledge respects existing permissions, so ChatGPT can access only what each user is already allowed to view (OpenAI company knowledge help, 2026).
DOCX, PDF and TXT uploads are available on Free and paid ChatGPT plans, and a project holds up to 40 files on Business and Enterprise (OpenAI uploads help, 2026).
NIST AI 600-1 defines confabulation as "the production of confidently stated but erroneous or false content" (NIST AI 600-1, 2024).
When can ChatGPT answer from company technical documents, and when not?
ChatGPT can answer from company technical documents only when those documents are within reach of the conversation: attached files, project files, or apps connected through company knowledge. Anything else, it does not see. The table below decides, condition by condition, whether an answer about your own products can rest on your own documents:
Condition | Verdict | Why |
|---|---|---|
No files attached, no apps connected | No | The answer comes from general knowledge: it reads well but gives what is typical for the product class, not what your datasheet says |
One datasheet attached to the chat | Yes, for that file | ChatGPT works with DOCX, PDF and TXT files attached to a conversation (OpenAI uploads help, 2026); it cannot know whether a newer revision exists |
Datasheets kept in a project | Yes, for the project files | Up to 40 files per project on Business and Enterprise, 25 on Go and Plus (OpenAI uploads help, 2026); uploading a file with an existing name can add a second copy (OpenAI Projects help, 2026) |
Apps connected through company knowledge | Yes, within each user's access | Business, Enterprise and Edu only; it must be turned on in each new conversation, and answers link to the original sources; web only, not the desktop or mobile apps, and on Enterprise and Edu an admin must enable the apps first (OpenAI company knowledge help, 2026) |
Apps connected, company knowledge not selected | Partly | ChatGPT may still use connected apps, but its responses won't include the same depth or detailed citations (OpenAI, 2025) |
Visit notes, complaint history and other records outside connected apps | No | Records in a notebook, a personal file or an unconnected system never reach the answer, so the gap is filled with a generic explanation |
Any answer without a citation to a document and revision | Treat as unverified | NIST lists confabulation as a generative AI risk (NIST AI 600-1, 2024) |
The row that decides most customer questions is the sixth: if the answer depends on a visit note or a complaint that lives outside the connected apps, ChatGPT cannot use it, whatever the plan.
Sort the documents a product question needs into these rows before you ask. If the datasheet sits in a project, delete the old revision before uploading the new one. If the answer needs a visit note or a complaint record, attach it, or answer that part yourself.
Why does ChatGPT give generic answers about our own products?
ChatGPT gives generic answers about your own products because, without your datasheet or records in the conversation, it produces the answer that is most typical for that kind of product. NIST explains that large language models predict the next word, which can produce factually inaccurate output, and that this "is particularly relevant" in domains that require highly contextual or domain expertise (NIST AI 600-1, 2024). The result is a plausible generic answer. Asked for the pour point of your gear oil by trade name, a chat with no datasheet can return a value that is typical for the grade and still wrong for your formulation (illustrative example). Fix the input before the question. Decide which document the answer must come from, put that document in the conversation, and name the product the way the document names it, with its revision or issue date. Then compare any value in the answer with the table in that document, including the test method and unit. A value you cannot find in the document is not an answer from your documents.
Why should an AI answer from technical documents cite its source?
An AI answer from technical documents should cite its source because the citation is the only quick way to confirm, before the answer reaches a customer, that a value came from your current revision and not from general knowledge. With company knowledge, ChatGPT shows the sources it used and the snippets it drew from, and each citation opens the original (OpenAI, 2025).
A citation is a pointer to check, not proof. NIST notes that generative AI output "may also include confabulated logic or citations that purport to justify or explain the system's answer" (NIST AI 600-1, 2024). So open the citation every time the answer will be repeated outside the company.
Three checks are enough for most technical answers:
The cited file is the document the question is about, for the right product and the right market.
The revision or issue date on the cited file is the current one, not a superseded copy left in a drive or project.
The cited passage contains the value as stated, with the same unit, test method and conditions.
Write the document name and revision into the reply you send, so the next person can check the same source.
Which company records does a general AI chat not see?
A general AI chat does not see any record that is not attached, stored in the project or held in a connected app the user can view. Company knowledge reaches only eligible connected apps that support search and fetch, and only what each user is already authorized to view (OpenAI company knowledge help, 2026).
For technical service, the missing records are usually the ones that change the answer: visit notes and measurements from the customer's site, the complaint history for the account, earlier test results, superseded revisions still in use at the plant, and what was agreed with the customer and by whom. Map the records before relying on the chat. For each recurring question, such as a field complaint or a request for an equivalent product, list the records an engineer would open: datasheet, certificate of analysis, last visit report, complaint history. Then mark which of them sit somewhere the chat can reach. Any record left unmarked is a gap that a person must fill or check. Lubinpla answers a team's questions from the team's own accounts, projects, work logs, documents and attached files, and marks the record behind each sourced sentence so a person can open it before replying.
Frequently asked questions
Does ChatGPT train on company documents you upload?
By default, OpenAI does not use content from ChatGPT Business, Enterprise, Edu or ChatGPT for Healthcare workspaces to train its models (OpenAI Memory help, 2026). For personal accounts, OpenAI may use chats to improve models when the setting "Improve the model for everyone" is on; it can be turned off in Data controls (OpenAI Memory help, 2026). Check the plan and that setting before you upload a customer's documents.
What happens when two revisions of a datasheet are in the connected drive?
Company knowledge can run multiple searches to resolve conflicting details and can rank sources by recency and quality (OpenAI, 2025). Recency is not approval: a newer draft can outrank the released revision. Keep superseded and draft revisions out of the searched folders, or label them in the file name, and check the revision on every citation.
Can ChatGPT read the tables in a scanned datasheet PDF?
ChatGPT Enterprise supports visual retrieval for PDFs; other plans and document types use text-based retrieval, which extracts digital text from the file rather than its embedded images (OpenAI uploads help, 2026). A scanned datasheet with no text layer may therefore give an answer with nothing behind it. Use the supplier's digital PDF where you can.