ChatGPT with Company Data: How to Connect It, Where Numbers Go Wrong and What GDPR Asks

"We want to use ChatGPT with our company data" can mean uploading a spreadsheet to a chat, or an agent that queries the data warehouse with each person's permissions. What works and the privacy risk both change. There are five real ways to do it:

1. Uploading files to the chat.
2. Connectors built into the assistant, reading Drive, SharePoint, email or the CRM.
3. MCP, an open standard for connecting a model to your systems.
4. RAG: the model searches your documents and answers with citations.
5. Text-to-SQL: the model writes a query against the data warehouse and the number comes from it.

The first three work for documents from day one. The last two are what you need when the answer is a number. Everything below was checked against each vendor's documentation as of October 2026.

For a one-off question it works well. According to OpenAI's help center, ChatGPT writes and runs Python on the file you upload, in an environment that cannot reach the internet or other APIs. Spreadsheets and CSVs are capped at roughly 50 MB (OpenAI), and with scanned tables or tables in images it may not read exact values reliably.

It is good for exploring an export or a quick chart. It is not a system: the file is a snapshot, it ignores permissions and everyone ends up with their own numbers.

The next step is connecting the assistant to where your documents already live. All the major ones do it:

• ChatGPT: apps (formerly connectors; settings now list them under Plugins) connect to Google Drive, SharePoint or Slack, with availability depending on plan and region (OpenAI). Company knowledge, which answers across several sources, is only on Business, Enterprise and Edu, and respects each source's permissions (OpenAI).
• Claude: a connectors directory on every plan; on Team and Enterprise an owner enables them. It inherits each person's permissions, unless the connector uses a shared credential (Anthropic).
• Gemini: in Google Workspace it only retrieves Gmail and Drive content the user can see (Google). For outside sources, Gemini Enterprise has connectors to SharePoint, Jira, ServiceNow, Salesforce or BigQuery (Google Cloud).
• Microsoft Copilot: Copilot connectors (formerly Microsoft Graph connectors) cover over a hundred sources, such as Salesforce, SAP or SQL databases, indexed or queried live through MCP, always with the source's permissions (Microsoft).

They are very good at finding and summarizing documents. For numbers they share the file problem: the model reads text or scattered rows, not a table with agreed definitions.

MCP (Model Context Protocol) is an open protocol that lets a model use external tools and data. Anthropic created it and in December 2025 donated it to the Agentic AI Foundation, under the Linux Foundation (Anthropic). One MCP server can serve several assistants:

• In ChatGPT, custom MCP apps with read and write actions are on Business, Enterprise and Edu, and only admins publish them (OpenAI).
• In Claude, custom MCP connectors are on every plan, with one on the free plan (Anthropic).
• Google offers a BigQuery MCP server that caps each query at three minutes and 3,000 rows and recommends its read-only tool (Google Cloud).

MCP solves how to connect, not what it can see. A server that logs in with an account that sees everything shows everything to whoever asks. Queries should run as the person asking, starting read-only.

RAG (retrieval-augmented generation) sits behind almost every "chat with your documents" tool: documents are indexed and, for each question, the system retrieves the closest chunks and the model answers with citations. For policies, contracts or product sheets it works well. For numbers it fails by design: ask about this quarter's sales and it may retrieve a report from six months ago. It does not add up rows. It reads text.

For questions with a number in the answer, the model has to write a query that runs in the data warehouse with the asking user's permissions. That is conversational analytics.

Google already ships it. In BigQuery, data agents are configured with tables, a glossary and verified queries, and only access what the user can see (Google Cloud). In Looker, the agent answers on top of the LookML model, with the metrics already defined there (Google Cloud). We explain how to set both up in our guide to conversational analytics in BigQuery and Looker. Microsoft does something similar with Copilot in Power BI.

With cryptic table names or undefined metrics, the model will write a query that runs and returns the wrong number.

A model can produce a plausible figure that no table contains. Google itself warns that its agents' output can seem plausible and be wrong (Google Cloud). There are two different errors:

• Making it up: it answers without querying. Fix it by requiring every figure to come from a query that ran, shown to the user.
• Calculating something else: the query runs, but "sales" excludes returns or uses order date instead of invoice date.

The defence against the second is a semantic layer, where you define once what an active customer is or how margin is calculated: LookML in Looker, a Power BI semantic model, dbt, or the glossary and verified queries of a BigQuery agent. Before opening it to the team, prepare 20 or 30 real questions with known answers and hold off until it gets them right.

On business plans, OpenAI, Anthropic, Google and Microsoft say they do not train on your data by default; we summarize it, with sources, in Generative AI for sales. OpenAI also says ChatGPT Enterprise data is encrypted at rest with AES-256 and in transit with TLS 1.2 or higher (OpenAI). None of that is GDPR compliance on its own. Also check:

• A data processing agreement. OpenAI, for example, offers a Data Processing Addendum to business customers.
• Where data is stored and processed. OpenAI offers data residency in Europe (EEA and Switzerland) to new Enterprise and Edu customers and eligible API customers (OpenAI). BigQuery data agents can be created in the EU multi-region.
• What is used to improve the product. In BigQuery conversational analytics, Google analyzes anonymized prompts by default to improve it, without training its models, and you can opt out (Google Cloud).
• Who sees what. A connector with a shared credential bypasses your CRM's permissions.
• Data minimization. For sales by region, the model does not need names or emails.