The internal assistant answers based on the company’s SOPs, manuals and catalogue, and links to the source document under every answer. New hires get up to speed faster, and experienced staff stop spending time on the same questions.
SOPs are written and manuals sit in a folder, yet employees still ask their colleagues. The internal AI assistant takes over these questions. It first searches for the relevant document passages, then writes an answer based only on them. This approach is called RAG: “find first, then answer.” If the answer isn’t in the documents, the assistant says “I don’t know” and forwards the question to the person in charge.
Every answer shows which document it relies on, so the employee can check.
The employee asks a question by text or voice. The bot identifies their role.
The assistant searches for relevant passages only in documents available to that role.
A short answer is written, followed by the document title, section and link.
The question goes to the “unanswered” list and is sent to the knowledge owner as a task.
The owner adds the new answer to the document, and the assistant uses it for future questions.
Without these conditions, the assistant answers fast, but the answers are risky.
One current document per topic. Old versions are archived; otherwise the assistant will give two different answers.
Finance, payroll and HR documents are open only to the relevant roles. The assistant never quotes a document the user has no access to.
Every answer cites a source. If no source is found, the assistant says “I don’t know” and refers the question to the person in charge.
Each document group has an employee responsible for updates, who also receives the unanswered questions.
The assistant doesn’t copy documents: they stay where they are, and the assistant indexes them.
The test set is 50–100 real questions with approved answers.
| Metric | Calculation | What it means |
|---|---|---|
| Answer accuracy | Share of correct answers on the test set | Knowledge base and search quality |
| Sourced answers | Share of answers with a link | Verifiability |
| Unanswered questions | Share of “I don’t know” answers and their topics | Knowledge gaps |
| Adoption | Weekly active employees and number of questions | Level of trust |
| New hire onboarding | Days until working independently | Business result |
If SOPs haven’t been written yet, AI won’t invent them. We first put processes and job descriptions in order, then connect the assistant.
Each employee gets answers only from documents open to their role. The most sensitive documents (payroll, personnel files) can be left out of the knowledge base entirely.
The assistant finds the answer in the Russian document and replies in Uzbek. The source link leads to the original document.
In Odoo 19 and 20, an agent can be given PDFs, links, files from the Documents app and Knowledge articles as sources. If all your knowledge lives in Odoo, that is often enough. If it is spread across several places or questions come in via Telegram, we build an external assistant.
Each document shows its last update date and owner. Owners get reminders about documents that haven’t been updated for a long time, and archived documents are excluded from search.
Relevance matters more than volume. We usually start with 20–30 core documents and expand based on unanswered questions.
We review your processes and recommend a specific task, KPI and plan for your first agent.