Approved documents only
The first version can start with a limited document set such as policies, procedures, manuals, templates or sample reports.
Private document assistant
How a private document assistant answers from your own files
A private document assistant helps staff search policies, procedures, manuals, reports and internal knowledge using a familiar ChatGPT-style interface. The difference is that the assistant is built around approved business documents and can show where an answer came from.
This guide explains what that looks like and what to check before a pilot. Gen AI Solutions builds local, offline and self-hosted options for Irish businesses, with processing and access agreed around the information involved.
What it is
Find the policy and see the source
Instead of staff pasting confidential files into a public AI tool, a private document assistant works from a controlled set of files. It can help answer routine internal questions, point people to the right policy, summarise approved material and reduce reliance on memory or folder hunting.
Source references
Answers can include references back to the files or sections used, so staff have a route back to the original material.
Private deployment options
Depending on the use case, the system can run locally, self-hosted or on controlled infrastructure instead of a public AI platform.
Fictional worked example
A question about booking a meeting room
This made-up office policy shows the intended workflow. It is an illustration, not a client result or a live product demonstration.
The approved source
Fictional Office Handbook, version 1, section 3, Meeting rooms.
Book a room in the shared calendar before inviting attendees. For meetings with visitors, notify reception the working day before.
The staff question
What do I need to do before a meeting with visitors?
A source-based draft answer
Book the room in the shared calendar before inviting attendees, and notify reception the working day before the meeting.
Source reference — Fictional Office Handbook, version 1, section 3.
The human check
Open the cited section, check that it is the current policy and confirm it applies to the meeting. A reference helps someone check the answer. It does not prove that the answer is complete or correct.
If someone asks about visitor parking and the approved files do not cover it, the assistant should say it has no supporting source and direct the person to a colleague. That behaviour needs testing in the pilot.
Who it helps
Useful for teams with repeated document questions
Accountants and bookkeepers
Search internal procedures, client service templates and training notes within an agreed setup with appropriate access controls and processing arrangements.
Solicitors and legal offices
Help staff find approved precedents, office procedures and internal guidance while keeping legal documents under control.
Medical, dental and care practices
Search practice policies, admin procedures and staff training notes without using public AI for sensitive material.
Engineering and construction firms
Ask questions across method statements, manuals, project procedures and technical documentation.
Plain-English technical note
Where RAG fits
RAG means retrieval-augmented generation. In plain English, the assistant first looks for relevant passages from approved files. It then uses those passages to draft an answer. This is why a private document assistant can answer from your material rather than relying only on general AI knowledge.
Retrieval does not itself make a system private. Privacy depends on where the search and AI model run, what information is sent to providers, who has access and how the system is maintained. These choices need to cover the whole workflow.
RAG is useful when your business wants answers grounded in its own policies, procedures, manuals, reports or internal knowledge. It is still important for people to review answers before using them for client work, legal work, medical decisions or other high-risk tasks.
When fine-tuning matters
Fine-tuning is not always the first step
Fine-tuning means adjusting a model using examples so it behaves more like a specialist system. It can be useful for repeated, well-defined tasks or very specific language patterns.
Fine-tuning does not replace a current document collection or provide source references on its own. For a policy-search task, testing retrieval over approved files is usually the more useful starting point. Fine-tuning can be considered later if evaluation identifies a specific behaviour that needs improvement.
What gets delivered
Define a small pilot that staff can check
-
1
Use-case selection
We choose one practical document workflow and agree what should stay out of scope.
-
2
Document review
We identify approved files and the people who may use them, then agree processing, access and retention needs.
-
3
Private assistant setup
The assistant is configured around the approved documents using local, self-hosted or controlled options where appropriate.
-
4
Testing and handover
Staff test agreed questions, missing-information cases and source references, then receive guidance on review and updates.
What to establish before a quote
We need to understand the file formats, approximate volume and condition of the documents. Scanned pages may need text extraction and checking. Tables, diagrams and poor scans need specific tests rather than an assumption that every file will work.
The pilot scope should name the document owner, update process, access rules and example questions. It should also define when the assistant must say that the sources do not support an answer. Accuracy on a small approved set does not establish accuracy on every future document or question.
Not sure where to start?
Start with the AI Privacy and Readiness Audit
If you are unsure which documents, tools or workflows are safe to test, the audit gives you a practical first step before any build work.
FAQ
Private document assistant questions
What is a private document assistant?
It is a ChatGPT-style assistant that answers from approved business documents, policies, procedures or manuals. It can be designed so sensitive files stay on your own hardware or controlled infrastructure.
Is this the same as RAG?
RAG is the technical method often used behind the assistant. It retrieves relevant passages from your approved documents before the AI drafts an answer.
Do we need fine-tuning?
Often no. Many projects start with retrieval over approved files. Fine-tuning is considered only when there is a clear reason, such as a specialised repeated task or domain language that retrieval alone cannot handle.
Can it run offline?
Depending on the workflow, hardware and model choice, a private document assistant can be designed for local, offline or self-hosted use.
Can staff trust every answer?
No AI system should be treated as automatically correct. Staff should review important answers, check sources and use human judgement for client, legal, medical, financial or operational decisions.