Skip to content

An AI assistant for company documents: when it is worth it and when it is not

Yes, if you have many up-to-date documents and the same questions keep coming back: the assistant answers and points to the document; for a handful of documents or for decisions that need a person, it is not worth it.

Pop Robert · Updated:

Yes. An AI assistant can answer from your company’s contracts and procedures and show the document behind each answer, if you have many up-to-date documents and the same questions come up often. It is not worth it for a few documents, for information that changes daily without a single source, or for decisions a person must make.

What an assistant for company documents is

It is a program where a colleague types a question in everyday words, for example “How long do we have to pay a supplier’s invoice?”, and gets a short answer together with the document and the passage it comes from. It does not answer from what the AI model “knows” in general, but from the documents your company has chosen and approved. Specialists call this way of working RAG (answers from retrieved documents).

Compared with an ordinary search in a folder, the assistant finds things by meaning, not just by exact words: “payment terms” and “invoice due date” lead to the same paragraph. Compared with a chatbot that answers from memory, the difference is the reference to the source: whoever reads the answer can check it in the document.

How it works, in three steps

  1. Preparing the documents. The chosen documents are turned into text, split into short passages and stored in a database that can compare them by meaning. Scanned documents first have to be read as text, and a blurry scan gives the wrong text.
  2. The search. For each question, the system looks for the passages closest in meaning and keeps only those the person asking is allowed to see.
  3. The answer. The AI model receives the question and the passages found and writes the answer from them alone, pointing to the document. If it found nothing suitable, it says it has no answer instead of guessing.

Where we have used this mechanism

In Nexus Movie AI, our own product, the search compares the text a visitor types with the film synopses stored in the database and finds results by meaning rather than exact words. On film, series and actor pages, the assistant’s answers rely strictly on what is on the page and say so, so they can be checked. It is a film catalogue, not a contract archive, but the principle is the same: first find the source, then answer from it.

Which documents are a good fit

  • procedures and work instructions;
  • internal rules and company policies;
  • standard contracts and general terms, for example those with suppliers;
  • product manuals, technical data sheets and installation or service instructions;
  • frequently asked questions, with the approved answers.

Documents work well when they have a clear valid version and a person responsible for them. Order matters more than quantity: a few up-to-date procedures help more than a folder full of versions where nobody knows which is the latest.

What is not a good fit

  • information that changes daily and has no single source, for example stock kept in several spreadsheets;
  • data that already lives in a program, such as your accounts or your customer list: that needs a direct connection to the program, not a document search (see the guide on integrations and automation);
  • documents with several versions that contradict each other, until it is settled which one is valid;
  • confidential documents, if access rights cannot be applied in the assistant as well.

When a document assistant is worth it

  • You have many documents. Too many for anyone on the team to know them by heart or to find the right passage quickly.
  • The same questions keep coming back. New colleagues, the sales team or customers ask again and again about the same clauses, deadlines and steps.
  • The answer has to be cited. “I think that is right” is not enough: whoever asks needs to see the document and the passage, so they can check.

When it is not worth it

  • You have few documents. One well-organised document with a table of contents, or a list of frequently asked questions, solves the same problem more simply and more cheaply.
  • The information changes daily, without a single source. The assistant would answer from what was written yesterday. First you need to settle where the correct information lives.
  • The answer is a decision. The assistant can find the termination clause, but it cannot decide whether you terminate a contract, how you interpret it in a dispute or whether you approve an exception. Those stay with a person.

If you are not sure which group your process falls into, an AI consulting session starts from a single process and ends with written recommendations, including the simpler option if AI is not justified.

What the risks are

A wrong answer, given with confidence

An AI model can give an incomplete or wrong answer, worded as convincingly as a correct one. Pointing to the source does not remove the risk, but it makes it visible: whoever gets the answer sees the passage and can check it. For important questions the rule has to be clear: read the document, not just the summary.

Old documents

If expired versions of a procedure stay in the collection, the assistant may quote them. Decide in advance who adds new documents, who removes old ones and how often the list is checked.

Access rights

An employee should not learn through the assistant what they could not open in the folder: salaries, personnel files, confidential contracts. Access rights are applied at the search, before the passages reach the AI model, not only in the window where the answer appears.

Personal data

Contracts, personnel files and customer correspondence usually contain personal data, so the GDPR applies. Put into the assistant only what its purpose needs: the regulation requires data to be limited to what is necessary (Article 5(1)(c)). A provider that processes the data on behalf of your company, for example the service that runs the AI model, must provide sufficient guarantees, and the relationship with it is governed by a contract (Article 28).

In our projects, the data is stored on servers in the European Union. If a service used in the project, such as an AI model, processes data outside the EU, we ask for your approval first. What the European law on AI requires of a company that uses tools like this is covered in the guide on the AI Act for small businesses.

How to check it before launch

Before the team relies on the assistant, prepare a set of questions to which you already know the correct answer and the document it comes from. The examples below are fictional.

  1. Questions with a known answer: “What are the payment terms in the standard supplier contract?” The answer must be correct and point to the right contract and clause.
  2. Questions with no answer in the documents: “Do we have a procedure for working from home?”, when there is none. The assistant must say it did not find the information, without inventing a rule.
  3. Questions asked without access rights: a colleague in sales asks about salaries in another department. The answer must reveal nothing from documents that colleague cannot access.
  4. Questions about a changed rule: after you replace a procedure, check that the assistant quotes the new one.

Keep the set and run it again after every significant change to the documents. If an answer comes out wrong, check the documents and the way they were prepared first, then the assistant.

What it costs

Smart document search is part of our AI integration service, whose package starts from 999 lei with us. The final price depends on how many documents there are and what state they are in, on access rights and on the systems the assistant has to connect to. After a conversation you get a firm quote in writing.

On top of the build, the AI model that writes the answers is usually paid by usage: the more questions the team asks, the higher the usage. We explain how the monthly costs add up in the guide on what an AI chatbot costs.

What to prepare for the first conversation

  • the list of documents the assistant should use and who is responsible for each one;
  • the questions asked most often, with the correct answers;
  • who is allowed to see each document;
  • what the assistant must never do.

With these, write to us on the Contact page.

Contact

Let’s talk about your project

Tell us what you need. We’ll get back to you with a clear proposal, including timelines and costs.

What are you interested in?
Approximate budget (optional)

We use your details only to reply to you. Privacy policy

WhatsApp: Message us now (opens in a new tab)