# How to give ChatGPT your company's knowledge

> A step-by-step guide to making ChatGPT answer from your company's own material: projects, custom GPTs, company knowledge, synced apps, custom MCP apps and the API. Which plans get what, whether retrieval respects your permissions, the file limits that stop people, and what OpenAI itself says it is bad at.

Canonical page: https://www.heybeagle.com/guides/how-to-give-chatgpt-your-company-knowledge

Last reviewed: 2026-07-30

Make ChatGPT answer from your material. Five paths, one honest ceiling.

Nobody wants what they ask for here. They ask how to train ChatGPT on their documents; what they want is for it to answer correctly from material it has never seen. Those are different problems with different solutions.

## The short answer

You cannot train ChatGPT on your company's documents. Nothing you upload changes the model, and every path in this guide is retrieval instead: the model is handed relevant extracts at the moment you ask. There are five real paths. A project holds a handful of files plus instructions and works on every plan, capped at 40 files on the business tiers and fewer below. A custom GPT holds up to 20 knowledge files of up to 512 MB each and stays the right shape for static, org-wide reference material. Company knowledge, on ChatGPT Business and Enterprise or Edu, holds no files at all: it searches across the apps you have connected and mirrors each person's existing permissions, so two colleagues can get different answers to the same question. A custom MCP app extends that to systems with no off-the-shelf connector. Building it yourself on the Responses API with file search buys total control over chunking and ranking. The wall most teams hit is the file limit, and it is a signal rather than an obstacle: if the answer needs more than 40 documents, you do not have a file problem, you have a systems problem, and only the last three paths solve it.

## At a glance

- **Can we train it on our documents?**: No, and you do not need to. Nothing you upload changes the model's weights. Every option here retrieves your material at question time and puts the relevant part in front of the model.
- **The three in-product paths**: Projects, on every plan including Free, capped at 40 files on Business, Enterprise, Edu and Pro. Custom GPTs, up to 20 knowledge files at 512 MB each, though creating one needs a paid plan. Company knowledge, on ChatGPT Business and Enterprise or Edu only.
- **Who has to approve it**: An admin, and the default is inverted between tiers: apps are enabled by default on Business workspaces but disabled by default on Enterprise and Edu, where someone has to turn each one on before company knowledge works at all.
- **Does it respect our permissions?**: Company knowledge does. OpenAI states it can only reach what each user is already allowed to view, and says plainly that each employee may therefore get a different answer to the same prompt. Files you upload to a project or a GPT have no permission model beyond who can open them.
- **How fresh is it**: Apps with sync build an index in advance. Initial sync can take days, exposes roughly the past 30 days at a partial stage, then completes and refreshes on an ongoing basis. Apps without sync fetch live at question time instead.
- **Where it does not work**: Company knowledge is web only, not the desktop or mobile apps, and it has to be switched on manually in each new conversation.
- **Biggest limitation**: OpenAI's own: synced retrieval is designed for question-answering and search, and performance is limited in scenarios requiring aggregation from numerous sources or very complex queries. It is not a replacement for a warehouse.

## Side by side

Six paths and the eight questions that separate them. Read row two and row three first: file count and permissions are what actually decide this, and almost nobody checks them before starting.

| Capability | A project | A custom GPT | Company knowledge | A custom MCP app | Built on the API | Beagle |
| --- | --- | --- | --- | --- | --- | --- |
| Works on Free and Plus | Yes (Fewer files) | Partly (Paid plan to build) | No (Business and up) | No (Business and up) | Yes (Pay per token) | Yes (No ChatGPT plan needed) |
| Holds more than a handful of documents | No (40 files) | No (20 files) | Yes (Your connected apps) | Yes | Yes | Yes |
| Retrieval follows each person's permissions | No (Whoever opens it) | No (Whoever uses it) | Yes (Mirrors the source) | Partly (You implement it) | Partly (You implement it) | Yes (Your own OAuth) |
| Cites where the answer came from | Partly (If you ask in the prompt) | Partly (If the instructions say so) | Yes (Links to the source) | Partly (Depends on the server) | Partly (You build it) | Yes (Links the records used) |
| Stays current without a re-upload | No (Files go stale) | No (Files go stale) | Yes (Synced or fetched live) | Yes | Partly (You run the sync) | Yes |
| Works outside the web app | Yes | Yes (Web to build one) | No (Web only) | No (Web only) | Partly (Wherever you ship it) | Yes (Slack and Teams) |
| An admin can restrict what it reaches | No | Partly (Workspace settings) | Yes (Per app, per role) | Yes (Admins publish it) | Yes (It is your code) | Yes (Per-tool toggles) |
| Nothing extra to pay | Yes | Yes | Yes (Business seat needed) | Yes (Business seat needed) | No (Per search and per GB) | No (Free tier, then credits) |

## What each one actually is

These are not variants of one feature. They are three different products and two different builds, and teams routinely try to make the first one do the third one's job.

| Approach | What you get | Where it stops |
| --- | --- | --- |
| A project (Minutes. Any plan) | A folder with files, instructions that override your global custom instructions, and optionally its own memory. Available on every plan including Free. The fastest way to make ChatGPT good at one recurring piece of work, and genuinely underrated for it. | Files are capped at 40 on Business, Enterprise, Edu and Pro, 25 on Go and Plus and 5 on Free, with no permission model beyond who can open the project. Project-only memory can only be chosen at creation, and sharing a project forces it permanently. |
| A custom GPT (An hour. Paid plan to build) | Up to 20 knowledge files at 512 MB each, plus instructions, published to your workspace. OpenAI still positions this as the right shape for static reference material curated by an expert and used across teams - the benefits handbook, the brand guide, the pricing rules. | Deliberately stateless: no saved memory, no custom instructions, each conversation starts fresh. It can use apps or actions but not both. Builders cannot see what people asked it, which is good for privacy and awkward for improving it. |
| Company knowledge (An afternoon, plus a sync that takes days) | The one most people actually want. It holds no files: it searches across the apps you have connected, runs several searches to resolve conflicting details, ranks by recency and cites its sources with links back. Critically, it mirrors the permissions of the source system for each user. | ChatGPT Business and Enterprise or Edu only, web only, read-only by design, and it has to be turned on by hand in every new conversation. OpenAI said in 2025 it intended to fold this into normal chat; as of July 2026 the toggle is still there. |
| A custom MCP app (Days. Admin to publish) | For the system that has no connector: a Model Context Protocol server you publish into the workspace, which company knowledge will then search alongside everything else provided it exposes search and fetch. | Business and Enterprise or Edu only, and only admins and owners can publish. Full MCP support including write actions is explicitly still in beta with the warning that permissions and interface may change. Company knowledge uses only the read half. |
| Built on the API (A fortnight, then forever) | File search over vector stores in the Responses API. You control chunking, hybrid ranking weights, metadata filters and score thresholds, and you can put the result anywhere rather than only in a chat window. | You are joining a moving platform. The Assistants API shuts down on 26 August 2026, and reusable prompts and Agent Builder both have published end dates in November 2026. Vector stores have no default expiry, so an abandoned one bills at $0.10 per GB per day past the free gigabyte, and they are not eligible for zero data retention. |
| An AI teammate that reads the systems directly (Minutes. Any Slack or Teams workspace) | The other shape of answer: rather than assembling a corpus for a chat window, put a coworker in the channel that can look things up across the stack when asked and, more usefully, when not asked. Beagle is one of these, working under each person's own OAuth grants. | It is not a search product and it will not index your entire document estate. If what you need is a permission-aware index over millions of files, buy one of those. This is for the case where the answer lives across four systems and one thread. |

## Step by step

Company knowledge is the path most teams should take, so that is the walkthrough. Step one is the one that saves the fortnight.

1. **Work out whether this is a files problem or a systems problem** - Count the documents that would have to be in scope for the questions you actually want answered. Under 40 and reasonably static, a project or a custom GPT will do and you can stop reading after step two. Over that, or spread across the CRM, the wiki, the tracker and the drive, no amount of uploading will get you there and you want company knowledge. The wall: 40 files per project on the business tiers, 20 per GPT
2. **Check the plan before you promise anything** - Company knowledge is available on ChatGPT Business and on Enterprise or Edu, and is not available on Free, Go, Plus or Pro. Business is published at $20 per user per month billed annually with a two-seat minimum; Enterprise pricing is not published at all, so budget for a sales conversation rather than a card.
3. **Turn the apps on, or check whether they already are** - This default catches people out because it is inverted. Apps are enabled by default on Business workspaces, and disabled by default on Enterprise and Edu, where an admin has to enable each one before anyone can use company knowledge. Company knowledge needs at least one enabled app that supports search and fetch, so with no apps on it does nothing at all. Where: Workspace settings, then Apps
4. **Choose how people connect, deliberately** - There are two models and they have different consequences. Self-service means each person authenticates their own account, which is what makes the permission mirroring meaningful. Deploying to the team means members need no setup at all, which is faster and moves the permission question onto whoever configured it. Pick per app rather than as a blanket policy.
5. **Constrain what it can do before you announce it** - Admins can hold an app to reads only or to a named set of actions, limit which account domains staff may connect, and pre-decide how future actions are treated when the vendor adds them. The safe posture for a knowledge rollout is reads only with new actions disabled by default, and you can loosen it later. Default: Read automatically, ask before important actions
6. **Wait for the sync, and know what the stages mean** - Apps with sync build an index on OpenAI's side in advance. The first sync can take up to a few days, and there is an intermediate stage where roughly the past 30 days is searchable and the rest is not. If you demo during that window you will conclude the product does not work. Wait for a complete sync before anyone judges it.
7. **Tell people the two things that make it look broken** - Company knowledge has to be switched on manually in each new conversation, and it is web only - not the desktop apps, not mobile. Almost every internal complaint that it "does not know our stuff" is one of those two things. Put both in the announcement rather than in a follow-up.
8. **Keep projects and GPTs for what they are actually good at** - This is not a migration. A project is still the best home for one person's recurring piece of work, and a custom GPT is still OpenAI's recommended shape for static reference material an expert curates for everyone. Company knowledge is the layer for evolving material that lives in systems. Most teams end up running all three.

## What trips people up

The specific misunderstandings that waste the most time, starting with the one in the phrase everybody uses.

- **"Train it on our docs" is doing real damage** - The phrase makes people expect a one-time upload that permanently changes what ChatGPT knows, and then judge the result against that expectation. Nothing you upload changes the model. Every option here fetches your material at question time, which means the quality question is not "did it learn it" but "did it retrieve the right part", and those have completely different fixes.
- **The file limits are lower than people assume, twice over** - A project takes 40 files on the business tiers, 25 on Go and Plus, 5 on Free, and only 10 can be uploaded at a time. A custom GPT takes 20. Worse for anything visual: images inside documents are discarded on every plan except Enterprise, and even on Enterprise a PDF stored as project or GPT knowledge is processed text-only. The visual retrieval path applies to files uploaded during a conversation, not to stored knowledge.
- **The nouns changed under everyone's feet in July 2026** - What were connectors became apps, and on 9 July 2026 the app directory became the Plugin directory, with plugins as containers that can hold skills, apps and templates. OpenAI's own help centre still says connector in older articles. If you are following a guide written before that date, the words will not match the interface, and that includes some of OpenAI's.
- **The security default is inverted between the tiers** - Apps are on by default on Business and off by default on Enterprise and Edu. Teams upgrading from Business to Enterprise are the ones this bites: everything quietly stops working until an admin enables each app, and nothing in the interface explains why. Teams starting on Business get the opposite surprise, which is worse.
- **Two people, same question, different answers - on purpose** - Because company knowledge mirrors source permissions, OpenAI states directly that each employee may receive a different response to the same prompt. This is correct behaviour and it will still generate a support ticket the first week. Say it out loud in the rollout note. Also worth knowing: permission changes propagate quickly but not instantly.
- **OpenAI has published the ceiling, so use its words** - When someone asks whether this replaces the data warehouse, the answer is in OpenAI's own documentation: sync is designed to work best for question-answering and search, and the most relevant data is sent to the model based on query intent, which limits performance where an answer requires aggregating from numerous sources or a very complex query. Financial aggregation is the example they give.
- **If you build it yourself, check the end dates first** - The platform under this is moving fast enough to matter. The Assistants API shuts down on 26 August 2026 with no automated migration for threads. Reusable prompt objects and Agent Builder both shut down on 30 November 2026, and so does the Evals platform. Vector stores need an expiry policy you set yourself or they bill indefinitely, they accept 21 file types that do not include spreadsheets or CSVs, and they are not eligible for zero data retention - which surprises people who assumed a ZDR agreement covered their whole build.

Retrieval answers the question somebody thought to ask. Most of what a team actually needs is the answer nobody asked for, on the Tuesday it still mattered.

## Or skip the build

If the goal is not "ChatGPT, but briefed" so much as "the work done, using what we know", that is a different product. Beagle sits in the channel, reaches the systems directly, and does not need a corpus assembled first.

- **No corpus to assemble or keep current** - Beagle queries the CRM, the tracker, the drive and the inbox when the question is asked, so there is nothing to upload, no 40-file ceiling and no index to wait days for.
- **It answers where the question was asked** - The answer arrives in the Slack or Teams thread, with the records it used linked, rather than in a separate browser tab someone has to remember to open and switch a toggle on in.
- **It notices as well as retrieves** - The renewal coming up, the account that has gone quiet, the report due before the standup. Retrieval waits for a prompt. The useful half of institutional knowledge is knowing when to bring something up.
- **Every outbound action waits for a nod** - Beagle drafts the reply, the summary and the record update, and holds all of it for a human's approval. Nothing leaves on the AI's own authority.

## In your stack

Teams here usually connect: Drive, Notion, Slack, HubSpot, Linear, Gmail. Beagle connects to ~3,200 tools in total - see https://www.heybeagle.com/integrations.

## FAQ

**Can I train ChatGPT on my company's data?**

Not in the sense the question usually means. Uploading files to a project, a custom GPT or a connected app does not change the model or teach it anything permanently. What actually happens is retrieval: when you ask a question, the relevant extracts are found and put in front of the model as part of that one request. This is good news, because it means material stays current and can be revoked, and it also means the failure mode is a retrieval failure rather than a learning failure.

**What is the difference between a project, a custom GPT and company knowledge?**

A project is a folder of up to 40 files plus instructions, available on every plan, best for one recurring piece of work. A custom GPT holds up to 20 knowledge files at up to 512 MB each, is stateless by design, and is OpenAI's recommended shape for static reference material curated for a whole organisation. Company knowledge holds no files at all: on ChatGPT Business and Enterprise or Edu it searches across your connected apps, mirrors each user's permissions in the source systems, and cites its sources. Most teams end up using all three for different jobs.

**Which ChatGPT plan do we need for company knowledge?**

ChatGPT Business or ChatGPT Enterprise or Edu. It is not available on Free, Go, Plus or Pro. Business is published at $20 per user per month billed annually, or $25 billed monthly, with a two-seat minimum. OpenAI does not publish Enterprise pricing. Note one more gate underneath: company knowledge needs at least one enabled app that supports search and fetch, and apps are disabled by default on Enterprise and Edu workspaces.

**Does ChatGPT respect our file permissions?**

Company knowledge does. OpenAI states that it can only access what each user is already allowed to view, and that it either records directly whether a user has access to a file or syncs the file's permissions and directory information to resolve group membership. The consequence, which OpenAI also states, is that two employees may get different answers to the same prompt. Approving an app in ChatGPT does not grant anyone anything new in the source system. Files uploaded to a project or a custom GPT are different: they have no permission model beyond who can open the project or use the GPT.

**How many files can I upload to ChatGPT?**

It depends where you are putting them. A project takes 40 files on Business, Enterprise, Edu and Pro, 25 on Go and Plus and 5 on Free, with 10 uploadable at a time and unlimited projects per account. A custom GPT takes 20 knowledge files. Individual files can be up to 512 MB, with a two million token ceiling for text documents and a much lower practical limit for spreadsheets. Storage is pooled at 25 GB per user and 100 GB per organisation across chats, projects and GPT knowledge.

**Should we build this on the API instead?**

Only if you need control that the product does not give you: your own chunking and overlap, hybrid ranking weights you can tune, metadata filtering, a score threshold, or the answer delivered somewhere other than a chat window. File search in the Responses API bills at $2.50 per thousand calls with vector store storage free to one gigabyte and $0.10 per GB per day after that. Three things to check before committing: the Assistants API shuts down on 26 August 2026, vector stores do not expire unless you set a policy, and they are not eligible for zero data retention.

## Read next

- [ChatGPT in Slack](https://www.heybeagle.com/guides/how-to-use-chatgpt-in-slack): Getting the answers to where people ask
- [Claude and MCP](https://www.heybeagle.com/guides/how-to-connect-claude-to-your-tools-with-mcp): The same problem, the other vendor
- [What Beagle connects to](https://www.heybeagle.com/integrations): The systems the knowledge actually lives in

Try Beagle free: https://www.heybeagle.com/signup (1,000 credits, no credit card).