Build a Company AI Assistant From Your Own Documents: Projects, GPTs and Custom Builds

Every organisation has the same few questions asked again and again. "What is the leave policy?" "Which form do I use for a refund?" "What is the current price for the three-day package?" The answers exist, somewhere in a shared drive, an old email or one experienced person's head.

A company AI assistant answers those questions from your own documents, with a pointer to where the answer came from. It is one of the most useful first AI projects for an organisation in Nepal, and in many cases it can be built in an afternoon with tools you may already pay for. This post covers the three ways to build one, the steps that decide whether it works, and when you need a custom build.

How it works (and why you do not need to "train" anything)

A common misconception is that the AI has to be trained on your data. It does not. The approach every serious assistant uses is retrieval: when someone asks a question, the system searches your documents for the relevant passages and hands them to the model together with the question. The model writes the answer from those passages.

That has three practical consequences:

• Updating is easy. Replace the old price list with the new one and the next answer uses it.
• Answers can be checked. The assistant can say which document it used.
• Nothing leaves your control permanently. Your documents are read at answer time, not baked into a model.

Three ways to build one

Inside a subscription

Claude Projects is where I start most internal assistants. A Project holds documents and standing instructions, and every chat inside it uses them. On paid plans, when the knowledge grows large, Claude switches automatically to retrieval, which lets a Project hold up to ten times more material. On Team and Enterprise plans, Projects can be shared with colleagues. Claude also connects to Google Drive and, since July 2026, to Microsoft 365 with access to mail, calendar and SharePoint files.

ChatGPT offers Projects for grouping files and instructions, Custom GPTs for a shareable assistant, and, on Business and Enterprise plans since April 2026, Workspace Agents, which OpenAI positions as the successor to Custom GPTs for organisations and which connect to tools like Slack and Salesforce.

Google is replacing Gemini Gems with Skills (announced 30 September 2026), and NotebookLM remains a strong option for a document-grounded research assistant, especially for teams already on Google Workspace.

All three are good enough for an internal assistant. Choose the platform your organisation already pays for and administers, rather than adding a new subscription for this.

Building one in an afternoon: the steps that matter

1. Pick one job, not "everything"

"HR policy questions" or "product and price questions for the sales team" works. "Everything about the company" does not. A narrow assistant gives better answers and is easier to test. You can add a second one later.

2. Gather the source documents, and clean them

This is the step that decides whether the assistant is any good, and the one people skip. Before uploading:

• Remove outdated versions. If "Leave Policy 2023" and "Leave Policy 2025 final v2" are both in the folder, the assistant may quote either. Keep one authoritative version of each document.
• Resolve contradictions. If two documents disagree, the assistant cannot know which is right. Neither can a new employee.
• Convert to text where layout does not matter. Plain text or markdown is read more reliably and costs less than scanned or image-heavy PDFs. I explain why in token optimisation habits.
• Fix legacy Nepali files. Preeti-font documents and scans must be converted to Unicode text first.

Ten to thirty clean documents usually beat two hundred messy ones.

3. Write the instructions

The standing instructions are a prompt like any other, and the same frameworks apply. A starting point I use:

You are the internal assistant for [organisation]'s [team]. Answer staff questions using only the documents in this project.

Always name the document your answer comes from. If the documents do not answer the question, say "The documents I have do not cover this" and suggest asking [name or role]. Never guess policy, prices or dates.

Keep answers short: the direct answer first, then the detail. Answer in the language the question was asked in, English or Nepali.

The three rules in the middle paragraph (only from the documents, cite the source, say when you do not know) prevent most embarrassing answers.

4. Test with twenty real questions

Ask the people who will use it for the twenty questions they actually get asked, including awkward ones. Check every answer against the documents. Where it is wrong, the cause is usually a document, not the AI: fix the document and test again.

5. Give it an owner

Assign one person to keep the documents current and review a sample of questions each month. An assistant with no owner drifts out of date within a quarter, and then staff stop trusting it.

Data rules: what goes in and what stays out

Use a business plan for company documents, not employees' personal accounts. Claude Team and Enterprise and ChatGPT Business and Enterprise do not use your data to train models by default. On consumer plans, that depends on each user's settings, which you cannot control.

Even on a business plan, some material does not belong in a general staff assistant: salaries, individual HR cases, customer personal data, unpublished financials. If people with different access rights need different answers, that is the point at which you need a custom build with role-based access, not a shared Project.

When you need a custom build

An assistant inside a subscription is the right first step for most organisations. A custom build becomes worth it when:

• Customers will use it. On your website, in WhatsApp or Messenger, without an account on any AI service.
• Access must vary by role. Managers see what staff do not.
• It needs live data from your systems. Stock levels, booking availability or order status, not only documents.
• You need records. Logs of every question and answer, for quality and compliance.
• Volume is high. Hundreds of conversations a day are cheaper on the API than on seats.

That is the kind of assistant I build as part of my AI automation and custom software work: grounded in the client's own content, instructed to cite sources and hand off when unsure, and connected to the systems that hold live data. The assistant on this site follows the same rule: it answers only from this site's own pages and says so when a question falls outside them.

From there, the natural next step is an assistant that does not only answer but acts, such as drafting the reply or updating the record. That is where AI agents come in.

How I can help

• Afternoon setup. I build your first internal assistant inside the subscription you already use, including the document clean-up and the instructions, and train the owner to maintain it.
• Custom assistants. Customer-facing or system-connected assistants for websites, WhatsApp and internal tools.
• Team training. An assistant only helps if people use it well. A training session alongside the launch makes the difference between a tool people try once and one they rely on.

Tell me which questions your team answers again and again, and I will suggest which of the three approaches fits.