Your own company GPT: chat with your business data without it leaving the building
Short answer
A company GPT is a chat assistant that answers questions from your own documents, procedures and systems, with sources. It consists of a language model, a search across your sources, roles per team and a log. The question is not whether you need one but where the model runs: with a US provider that sees your data, or on a node that belongs to your organisation and runs open-weight models.
What a company GPT is
The term comes from the “custom GPTs” in ChatGPT: a chat you give your own files and instructions. In organisations it usually means more: an assistant that knows the handbook, the contracts, the tickets and the intranet, answers questions from them and names the place the answer came from. Technically this is retrieval-augmented generation: the question is searched against your documents, the hits are passed to the model, the model writes the answer with a reference. See AI on your own documents for the mechanics.
The model learns nothing in the process: it reads afresh for every question. That is why a company GPT can be built without training, and why access remains the most important question.
Three routes, and what they mean
| Variant | Where model and data run | What you have to arrange |
|---|---|---|
| Custom GPT in ChatGPT | At OpenAI in the US; uploaded files are stored there | A DPA only with Business or Enterprise, transfer outside the EU, no role model beyond your own permissions |
| Copilot or a cloud assistant in the tenant | In the vendor’s cloud, within your licences | Which sources the assistant reaches; old sharing permissions become visible. See Copilot or your own AI |
| Your own node with an open-weight model | In your server room or on a dedicated node in a European data centre | Operations, model choice, connecting sources; no provider in the chain |
The first two variants are quick to set up. But they mean a third party processes your questions and documents. For HR files, contracts, patient data or client matters, that is where most projects stall.
The four building blocks
- Model. An open-weight model such as Llama, Mistral or Qwen, in a size your node can carry. A large model for writing, a small one for classifying.
- Sources. File storage, intranet, ticketing, line-of-business applications, connected with the permissions that apply there. Start with one department and one set of documents.
- Roles. Who may ask what, and from which sources? The assistant must never see more than the person asking. See Access control for AI.
- Log. Every question, every answer, every source. For the data protection officer, for troubleshooting, and from December 2027 for the AI Act’s logging duty in high-risk uses.
How to roll it out
- One process first. The customer service knowledge base, the quality manual, the contract templates. With an owner and a measurable goal.
- Test questions up front. Thirty real questions with known answers. Use them to compare models and later to measure the correction rate.
- Sources instead of memory. Answers without a reference count as guesses. The assistant says so itself.
- Pilot with one team, four to six weeks. Only then the next department. The full sequence is in Setting up your own AI environment.
How werqly fits in
werqly is a company GPT on a node that belongs to your organisation: in your own server room or as a dedicated node in werqly Cloud, with open-weight models, connected sources and werqly AI Control for roles, sources per role and the log. Staff chat as they would with ChatGPT, except that questions and documents never leave the building. Calculate the configuration for your organisation below.
Frequently asked questions
Questions on this topic
Is a custom GPT in ChatGPT a company GPT?
It is the easiest start, but the files are stored at OpenAI in the US, and without a Business or Enterprise contract there is no data processing agreement. Usable for public content, not for confidential data.
Does the model have to be trained on our data?
No. A company GPT searches your documents for every question and passes the hits to the model. The model stays unchanged, answers come with sources, and new documents are usable immediately.
How do we stop the assistant revealing too much?
With roles: the assistant only searches sources the person asking may open themselves, and logs every answer. Start with a well-defined set of documents and expand step by step.
What hardware does a company GPT of our own need?
A server with a GPU, sized by the number of simultaneous users and the model size. For most organisations one node is enough; the calculator on this page shows the matching configuration.
Sources
This article is general information, not legal advice. Rules and guidance change; check the sources listed or ask your lawyer or DPO if in doubt.