A private ChatGPT for your company documents

What teams use a private ChatGPT for

It pays off where there is a lot of documentation and the same questions keep coming back. Some examples:

• Maintenance engineers: ‘What torque setting does this part take on this model?’, with the page from the manufacturer's manual, on a phone next to the machine.
• Quality and procedures: ‘What do we do if a batch arrives out of temperature?’, with the current version of the procedure.
• HR: ‘How much compassionate leave do we give under our policy?’, with the paragraph from the staff handbook.
• Legal and purchasing: ‘What notice period is in our contract with this supplier?’, with the clause quoted.
• Sales: spec sheets, compatibility and product terms without ringing the technical office.
• New starters: they ask the assistant before interrupting colleagues, and land on the reference document.

How RAG works, without the jargon

The technique is called RAG (retrieval-augmented generation), and it is what most RAG development services build. It works in three steps. Index: your documents are split into passages and stored in an index that searches by meaning, not just exact words. Search: when someone asks a question, the system finds the relevant passages among the documents that person is allowed to see. Answer: the AI model writes the answer using only those passages and shows where each part came from.

If the passages found do not contain the answer, the assistant says so instead of improvising, and the question is logged as a gap in your documentation. Before opening it to the team, we test it with a set of real questions and check every answer against the document it cites.

Permissions, versions and document quality

The assistant answers with what it finds, so the state of your documentation matters. This is what gets sorted out during the project:

• Permissions: if documents live in SharePoint, Google Drive or another document system, their permissions are inherited. An engineer does not see contracts or HR files.
• Versions: when an old and a new procedure coexist, someone has to decide which one rules. The assistant gives priority to the current version and flags contradictions when it finds them.
• Scans, tables and drawings: scanned PDFs go through text recognition, and tables and drawings need specific handling so information is not lost.
• Keeping it current: new or changed documents are added to the index automatically, and withdrawn ones stop appearing.
• Ownership: part of the initial work is agreeing with you which documents go in, what is out of date and who is responsible for each source.

Where your data lives and who can see it

The index of your documents is hosted on infrastructure we choose with you, in accounts in your company's name, and cloud regions in the UK or the EU can be used. Only the passages needed for each question are sent to the AI model, never the whole library. We use AI providers on business terms, not their consumer versions; what each provider states about training is in our FAQ answer is the AI trained on my data?.

Where documents contain personal data (contracts, personnel files), UK GDPR applies, or the EU GDPR for EU businesses, and whoever processes the data on your behalf acts as a processor under a written contract; the ICO explains what that contract needs to cover. Contracting entity, invoicing and data-processing terms are set out in writing in the proposal, together with where data is processed, how long it is kept and which providers take part. This is not legal advice: if you have a data protection officer or adviser, we give them the technical documentation they need.

Microsoft 365 Copilot, Gemini or a bespoke assistant

If your documents already sit in an office suite with AI built in, start there. Microsoft states that Copilot only surfaces organisational data the user has at least view permission for, and Google states that Gemini in Workspace only retrieves content the user has access to. According to their documentation, neither uses your prompts or data to train its foundation models without your permission. One detail for EU buyers: Microsoft notes that models provided by Anthropic as a subprocessor are currently excluded from its EU Data Boundary.

• If you also need live data from the ERP or CRM (stock, orders), what fits is an AI agent that queries your systems.
• If the person asking is a customer rather than your team, see our AI chatbot for customer service.

How we would run it and what drives cost

We start with a tightly scoped pilot: one area, one set of documents and a group of users with real questions. You see a working demo every two weeks with your own documentation, you get one fixed price for the project, and the code is yours. The rest is on how we work. What moves the cost most:

• Volume and format of documents: digital PDFs versus scans, tables and drawings.
• Number of sources: one folder versus several document systems.
• Permissions: inherited from one system or defined by your own rules.
• Channel: web, Teams, mobile or inside another application.
• Usage: the monthly cost of AI models depends on how many questions are asked; market references are in our guide to the cost of implementing AI in a business.