An AI assistant for your staff: answers from your own documents

Every company has a person everyone asks. How many days of leave carry over. Which form a supplier needs to fill in. What the warranty covers on the model sold last year. That person answers the same questions all week, and when they are on leave, the answers leave with them. An AI assistant built on the company’s own documents can take most of those questions, if it is set up with care.

The short answer

A custom AI assistant for staff answers questions from your own policies, manuals and price lists, and shows which document each answer came from. It works when the documents are current, when it only sees what each employee is allowed to see, when it says “I don’t know” instead of guessing, and when someone owns keeping it up to date. Without those four, it gives confident wrong answers, which is worse than none.

How it works, briefly

The assistant does not learn your documents the way a person does. When someone asks a question, the system searches your documents for the passages most likely to hold the answer, hands those passages to the AI model, and asks it to answer from them and cite them. This is usually called retrieval. AI providers offer it as a standard feature; OpenAI’s documentation on file search describes one version of it.

Two things follow. The answer can only be as good as the documents it finds. And because the assistant cites its source, a person can check any answer in seconds by opening the document it names.

Where it earns its keep

The best uses are questions that are asked often, answered from documents that already exist, and costly when answered slowly or wrongly. Some typical ones:

Team	Typical questions	Documents it answers from
HR	Leave, allowances, notice periods, procedures	Staff handbook, HR policies
Sales	Specifications, prices, delivery terms, warranty	Price lists, product sheets, terms
Operations	How to do a task, who approves what	Procedures, approval matrix
Customer service	Return rules, exceptions, previous decisions	Service policies, internal notes
New staff	Everything, in their first month	All of the above

Staff-facing assistants are also a sensible first step before a customer-facing one. The mistakes stay inside the company, and staff tell you quickly when an answer is wrong.

The four conditions

Current documents. If the handbook in the system is from 2023 and the leave policy changed in 2025, the assistant will quote 2023, with a citation, and sound certain. Before building, collect the documents, remove old versions and decide which one wins when two disagree. This is usually the largest part of the work, and it improves the company’s paperwork whether or not the assistant succeeds.

Permissions. A salary scale or a disciplinary procedure should not be quoted to someone who is not allowed to read it. The OWASP Top 10 for LLM Applications names sensitive information disclosure and weaknesses in how documents are stored and searched as two of the main risks for exactly this kind of system. The assistant must check who is asking and search only the documents that person may see. Filtering the answer afterwards is not enough.

Admitting ignorance. Instruct and test the assistant to say it does not know, and to suggest who to ask, when the documents do not hold the answer. An assistant that fills gaps with plausible guesses teaches staff to distrust all of its answers. We covered the related problem of assistants that hand over too often in why your Arabic assistant keeps handing over to a human.

An owner. Someone in the company must be responsible for adding new documents and retiring old ones, and for reviewing the questions the assistant could not answer. Those unanswered questions are a free list of gaps in your documentation.

Arabic, English and mixed documents

In Kuwait the handbook may be in Arabic, the supplier contracts in English and the price list in both. Staff ask in whichever language they think in, sometimes mixing the two in one sentence. A well-built assistant searches across languages and answers in the language of the question, while citing the original document. Test this specifically: ask in Arabic about a document that exists only in English, and the other way round.

Scanned documents need attention too. A photographed form or an old PDF without text layer must be converted before it can be searched, and Arabic text recognition needs checking by eye. Our note on automating Kuwaiti paperwork covers the pitfalls.

Testing before staff rely on it

Collect a set of real questions from the people who are asked them today, with the correct answer and the document it comes from. A hundred is a good number. Run them all, and check every answer and every citation by hand. Count wrong answers, missing answers and answers that cite the wrong document. The method is in the 100-example test.

The NIST AI Risk Management Framework makes the wider point: measure an AI system’s behaviour against the use it is actually put to, and keep measuring after launch. Repeat the test whenever major documents change.

Data and privacy

Your documents are processed by the AI provider behind the assistant. Know which provider it is, where the processing happens, whether your data is used to train their models, and how long it is kept. Leave out documents that do not need to be there, such as individual employee files. The questions to ask are set out in where your company data actually goes.

A worked example

Take an illustrative case: a building materials trader in Shuwaikh with forty staff. The sales team phones the office several times a day to check specifications, stock rules and the credit terms for each customer type. The office manager spends much of her morning answering them.

The assistant is built on the product sheets, the price list, the credit policy and the delivery terms. Salespeople ask on their phones, in Arabic or English, and get an answer with the document it came from. Credit limits for named customers are excluded, because that stays with finance. After a month, the office manager reviews the questions it could not answer, finds three procedures that were never written down, and writes them. The assistant now answers those too.

Mistakes to avoid

Loading every file on the shared drive is the most common one. More documents means more outdated and contradictory ones. Start with the ten that answer most questions. The second is launching without citations, which makes every answer impossible to check. The third is forgetting to tell staff what the assistant covers and what it does not, so they either ignore it or trust it with things it was never given.

How Tothiq builds them

Tothiq builds custom AI assistants on your own documents, in Arabic and English, with citations, permissions by role, and a monthly review of unanswered questions. The build is KD 1,250, and the running fee is KD 145 a month on a twelve-month agreement billed quarterly in advance, covering AI usage, hosting and document updates. If what you need is an assistant that answers your customers on WhatsApp, see Tothiq Assistant. Send us your enquiry.

Frequently asked questions

Does the assistant learn from our staff’s questions?

Not by itself. It answers from the documents you give it. The unanswered questions are reviewed so you can add documents, which is how it improves.

Can it answer in Arabic from an English document?

Yes, if it is built for it. Test this before launch, in both directions.

What stops it showing confidential documents?

Permissions checked before searching: it only searches the documents the person asking is allowed to see.

What happens when a policy changes?

The new document replaces the old one in the system, and the old version is removed so it cannot be quoted again.

Can staff use it on their phones?

Yes. It runs in the browser with a company login, so it works on phones without installing an app.