AI Knowledge Assistant for Your Company: Staff Answers From Your Own Policies, Manuals and SOPs

Every company has the same quiet cost: people asking each other where things are. How many leave days does a new hire get? What did we tell the last customer with this exact error? The answer exists somewhere, in a PDF policy or an old ticket, but finding it means interrupting the one colleague who remembers.

An internal knowledge assistant answers those questions from your own documents, shows where each answer came from, and keeps confidential material away from people who shouldn't see it. This page explains how it works in plain terms, where it fails, what it costs, and when a ready-made tool is the smarter choice.

What it answers, and from what

The assistant is only as good as the material behind it. Typical sources:

• HR rules and policies: leave, travel, expenses, attendance, the employee handbook
• Operating procedures (SOPs): how to receive stock, close a shift, onboard a client, handle a complaint
• Manuals and product specs: technical sheets, installation guides, pricing rules, warranty terms
• Past support tickets and resolved cases: what the problem was and what fixed it
• Internal FAQs and training material

How retrieval works, in plain words

The method is called retrieval-augmented generation (RAG). Four steps:

1. Split. Each document is cut into short passages, usually a paragraph or a section, each tagged with its title, date and who may see it.
2. Index. Every passage is turned into a numeric "meaning fingerprint", so a question about "annual leave" finds a passage titled "vacation entitlement" even though the words differ.
3. Retrieve. When an employee asks a question, the system finds the handful of passages closest in meaning, filtered to what that employee is allowed to read.
4. Answer with sources. A language model writes a short answer using only those passages, and lists the documents it used so the reader can open and check them.

The assistant on this website works the same way: it answers questions about Q Studio from our own knowledge base, in Arabic and English. When we change a fact in that base and re-index, its answers change. No model training is involved, which is also why your documents don't need to be used to train anyone's model.

Citing sources: the feature that builds trust

An answer without a source is a rumour. Every answer should end with the documents it relied on, ideally with the section, so staff can check before acting. Citations also help you debug: a wrong answer shows whether the model misread a good document or the document itself is outdated.

Permissions: who may see what

This is the part generic chatbots skip. A sales rep should not get an answer built from the salary scale; a branch manager should see their branch's procedures, not the board minutes.

We apply permissions at the retrieval step, not by asking the model to "keep secrets". Each passage carries the access level of its source document, and the search only returns passages the asking user is allowed to read. If nothing permitted matches, the assistant says it can't find an answer for them, without revealing that a restricted document exists.

Our own products follow the same rule. In Dawawin HR, our HR and payroll system, the built-in assistant answers in Arabic or English, shows which records an answer came from, and declines to share company payroll figures with an employee-level account, offering only that person's own payslip. Access is enforced in the database, not only in the screens. The public demo runs on a fictional company, so you can try it with different roles; in the demo the answers come from rule-based logic rather than a live AI model, while a client build connects a real model to your documents.

Keeping content fresh

Most wrong answers come from old documents, not from the AI. So freshness is designed in:

• One home per document. We connect to the folder, drive or system where the official version lives, and stop indexing copies.
• Automatic re-indexing when a document is added, edited or removed, plus a scheduled full refresh.
• An owner per topic. Someone in HR owns the HR documents. A short report of "questions the assistant couldn't answer" tells each owner what's missing.

Arabic search: where generic tools struggle

Arabic adds real difficulty to search, and it is where we spend testing time:

• Spelling variants: أ, إ, آ and ا used interchangeably; ة and ه; ى and ي. "اجازة" and "إجازة" must match.
• Dialect wording: staff ask the way they speak. An Egyptian employee may write "مرتب" where the policy says "راتب", or ask "امتى" instead of "متى".
• Mixed language and Arabizi: questions typed in Latin letters, or half Arabic, half English technical terms.
• Scanned PDFs: older Arabic documents are often images, which need text extraction first, and that step needs checking.

The fix is text normalisation, a search model that handles Arabic well, and testing with the real questions your staff type.

Where it fails, and how we handle it

Hosting and data privacy

• Your accounts. Hosting, the search index and the AI provider account are set up in your company's name. You pay the AI provider directly at its own price and see every charge.
• Paid AI tiers. Major AI providers state that on their paid API tiers your prompts and responses are not used to improve their models (Google's Gemini API terms say this for paid services, for example). We use paid tiers for company data and confirm the terms with you.
• Data stays where you decide. Documents and the index live in your cloud account; for stricter rules we can discuss a private-cloud setup during scoping.

Ready-made suite assistant or custom build?

If your company runs on Microsoft 365 or Google Workspace, their built-in assistants (Microsoft 365 Copilot, Gemini in Google Workspace) can answer from files staff already have access to. That is often the right first step: nothing to build, and permissions follow your sharing settings. The catch: loose sharing settings become loose answers, so tidy folder permissions first.

A custom assistant makes sense when:

• your knowledge sits in several places (a drive, a ticketing tool, an ERP, a website) that one suite can't see;
• you need strict role-based answers beyond file sharing, or answers inside your own app;
• Arabic and dialect quality is central and you want to test and tune it;
• you need it on a channel staff already use, such as a WhatsApp number for field teams (see our WhatsApp AI agent).

How we build it

1. Free call: what staff ask most, where the documents live, who may see what.
2. Proof of concept (about 5 days): 50–200 of your real documents, one group of users, a working assistant with citations and permissions.
3. Test together: your team asks real questions; we review right answers, wrong answers and "I don't know" cases.
4. Production (about 3 weeks): connections to your sources, automatic re-indexing, roles, admin view, usage reports.
5. Handover: code, accounts and running notes in your name.

Price and timeline

AI usage is paid by you directly to the AI provider. You approve a written scope and fixed price first; nothing is paid before that. Projects up to $500 are paid 50% to start and 50% on delivery. Projects over $500 start with a 30% down payment, then 40% and 30% at milestones, each paid after you see it working. A first preview arrives within 72 hours of starting.

When we're not the right fit

• Your documents are few and stable. A well-organised intranet page or FAQ may be enough.
• Your suite's built-in assistant already covers it. Use it, and call us if you hit its limits.
• You need on-site workshops or a locally registered vendor. We work remotely.
• You want the assistant to approve leave or decide disciplinary action on its own. We keep a person in those loops.

If your goal is answering customers rather than staff, start with our AI chatbot development page. To automate the tasks around the answers, see AI workflow automation.

FAQ

How many documents do we need to start?

For a proof of concept, 50–200 real documents covering one area, such as HR policies or support procedures, is enough to judge answer quality. More only helps if they are current and not duplicates.

Can it answer questions about our scanned Arabic PDFs?

Usually, once the text is extracted. Scanned Arabic quality varies, so we test a sample in the proof of concept and tell you which files need a cleaner copy. Our AI document processing service covers extraction in depth.

Will employees see documents they shouldn't?

Not if permissions are mapped correctly. Search is filtered by each user's access before the AI sees anything, and we test it with accounts at each role during handover.

What does it cost to run each month?

Hosting for the app and index, usually modest, plus AI usage paid to the provider per question. We estimate both from your headcount before production.

Can it connect to our ERP or HR system?

Yes, when the system has an API or a readable database copy. That turns "what's the leave policy?" into "how many days do I have left?", and is scoped as an extra step. Our ERP development page covers the systems side.

How do we start?

Tell us the questions your staff ask most and where the documents live, on WhatsApp or through the project planner. Within 24 hours you get a written scope for a proof of concept.