In a company with hundreds of people, the procedures exist but few can find them. They sit in PDFs on a server, on the intranet, in old emails. An internal AI assistant is a program the employee asks in their own words (“how long do I have to submit my travel expenses claim?”) and which gives back the answer taken from the document currently in force.
It is not an all-knowing colleague. It answers only from what it has been given, shows the source, takes account of the access rights of the person asking and says openly when it finds nothing.
The procedure exists, but nobody knows where
HR, admin and IT support answer the same questions every day: how to request leave, who approves a business trip, where the expenses form is. The answers are written down somewhere, but it is easier to ask a colleague than to look.
The problem gets worse when documents exist in several versions: the employee finds the old procedure and follows it. In factories and warehouses there is a further obstacle: many people do not work at a desk and have only their phone to hand.
From policy to answer: document search and a visible source
The method is called retrieval-augmented generation (RAG). In plain terms: the assistant does not answer from memory. It first searches the company’s documents for the passages related to the question and composes the answer solely on the basis of those. The search is semantic, meaning it finds passages by meaning, not just by identical words.
Below each answer the source appears: the name of the document and the section. If the passages found do not answer the question, the assistant says it has found nothing and points to the department responsible.
A language model can word things wrongly even with the right passage in front of it, for example by mixing up two similar rules. The source is displayed precisely so that the answer can be checked. For decisions with consequences, it is the document that counts, not the summary of it.
Internal AI assistant with access rights: who is allowed to find out what
Not every document is for everyone. The assistant takes access rights from the systems you already have: the employee signs in once, with their company account, and the search covers only the documents they could open themselves anyway.
Filtering is applied before the search, not after. A passage the person has no access to never reaches the language model at all, so it cannot appear in an answer, not even paraphrased. The condition is that the rights in your systems are correct and kept up to date.
- Single sign-on: people sign in with their company account, with no separate passwords.
- Versions: only the document in force is indexed; withdrawn ones drop out of the search.
- Log: what was asked and which sources were used, within the limits agreed with the data protection officer.
- Channels: internal chat, the intranet or the phone, for those who do not work at a desk.
Pay, sick leave, a conflict with the boss: questions for a person
The assistant can explain the general rule: how to request leave, which documents to submit, who approves. It does not answer about an individual’s situation: how much they are owed, why a request of theirs was turned down. It does not have the data for such questions, nor should it.
When it recognises a personal or sensitive subject (pay, health, a complaint), it does not attempt an approximate answer. It says who the question should go to. We write the list of these subjects with the HR team and the company’s lawyer.
To be settled from the start: who sees what employees have asked. If people suspect their questions are being read by their managers, they will stop asking. Retention rules are set in line with GDPR and communicated openly.
The order of work: documents first, then the assistant
Most of the work is in the documents. An assistant that reads three contradictory versions of the same procedure will give three answers. The first result is often a list of documents that are out of date or contradict one another.
The documents stay where they are (file server, intranet, cloud), and the assistant re-indexes them when they change. The language model can run with a provider that processes data in the European Union or on the company’s own infrastructure, depending on how sensitive the documents are.
- Inventory: what documents exist, who is responsible for each, which version is in force.
- The real questions: we collect what the departments are asked most often and use the list as a test set.
- Pilot: a single area and a small group of employees, who mark the answers as right or wrong.
- Maintenance: every new or changed document has an owner who also publishes it to the assistant’s source.
Frequently asked questions
What happens when the assistant cannot find the answer?
It says it has found nothing in the documents it has access to and points to the department that can help. Unanswered questions are collected in a list, which shows which procedures are missing.
Do the company’s documents reach the AI model provider?
The passages used for an answer are sent to the language model, so the choice of provider matters. You can work with services that, by contract, do not use the data for training, or with a model run on the company’s own servers.
Can it answer by phone too, not just in writing?
Yes, the same assistant can sit behind an internal number, with speech recognition and speech synthesis. Because the caller is harder to identify, phone access is usually limited to the documents open to all employees.
This is a typical project description: it shows how we usually approach this kind of work and does not present a project carried out for a particular client. Every real project starts from your company’s situation, and the stages, timescales and price are agreed after the initial discussion.