AI assistant for the company: document-based chatbot
An enterprise AI assistant can make it easier to find information in procedures and documents, but requires the right sources, permissions and response control. We explain how this solution works and where its limits are.
In many companies, the answer to a simple question is in one of several dozen documents, but the employee does not know which one. An AI assistant can help you find the right passage and prepare an answer. However, it does not replace current documentation, correct authorizations or human responsibility for the decision.
How the document-based assistant works
A frequently used approach is RAG, i.e. generating responses using sophisticated materials. The system divides documents into fragments, creates an index, searches for content related to the question and passes it to the language model. The answer may include references to sources.
RAG reduces some errors, but does not eliminate them. The system may look for the wrong part, miss an exception, combine conflicting documents, or formulate an answer that goes beyond the source. “I don't know” behaviour needs to be designed and tested; it does not automatically result from the use of RAG.
Where an assistant can help
Reasonable use cases include:
- searching for information in instructions and procedures;
- support for the implementation of new employees;
- preparation of a preliminary document summary;
- finding differences between material versions;
- creating a draft response based on the approved knowledge base;
- directing the question to the appropriate person or process.
An outcome with legal, financial, human resources, medical or safety implications should be reviewed by a person with appropriate competence.
What materials can be processed
Supported formats depend on the selected document reading component. A text PDF is usually easier to process than a scan, a complex table, a handwritten note, or a document with many columns.
Before indexing you need to check:
- whether the document is current and approved;
- who owns it;
- who can see its content;
- whether it contains personal data or trade secrets;
- how tables, footnotes and attachments will be recognised;
- what will happen to the old version after the update.
The index should not become a hidden copy of material that has been removed from the source system.
Permissions cannot exist only on the interface
The assistant should search only in documents available to the inquirer. Merely hiding a link in an interface is not access control. Permissions must be enforced at the application and search layers.
Tests should include attempts to:
- asking about another department's data;
- extract content using a command placed in the document;
- bypass restrictions by changing question wording;
- obtaining a response from the document after revoking privileges;
- plays a confidential fragment based on the conversation history.
External model vs. local model
When using an external API, the question and selected document fragments may be sent to the model provider's infrastructure. You should check the terms and conditions of the specific service, retention, processing location, subcontractors, data control mechanisms and data processing agreement.
The locally-running model reduces dependency on external APIs, but does not fully guarantee confidentiality. Data may leave the environment through telemetry, copies, integrations, incorrect permissions, or a compromised account. Your own installation also transfers responsibility for updates, monitoring, performance and security of servers to the company.
The choice should be based on data classification, quality requirements, cost and maintainability, not on the assumption that one model is always safer.
How to measure quality
A demonstrative answer to a few simple questions is not sufficient to accept the system. You need to prepare a test kit including:
- questions with a clear answer;
- questions requiring a combination of several sources;
- questions about information not available in the database;
- conflicting and outdated documents;
- typos, abbreviations and synonyms;
- attempts to obtain information without permission.
For each case, the correctness of the retrieved fragments, the compliance of the answer with the source, completeness and correctness of the refusal should be assessed. It is worth keeping the results to compare the quality after changing the model, the way documents are divided or the system manual.
Document content protection
The document may contain a command directed at the model, either intentionally or accidentally. The system should not treat text from the source as a superior instruction. Tool constraints, result filtering, permissions checking and prompt injection resistance tests are needed.
If the assistant can send a message, change a record, or run a process, the impact of the error is greater than with a text response alone. Actions should require confirmation, have minimal permissions, a scope limit, and a clear log of operations.
Implementation plan
- Select one use case and specify prohibited uses.
- Organise documents, owners and permissions.
- Map data flow through all components.
- Prepare a set of test questions and acceptance criteria.
- Run a pilot for a small group.
- Collect wrong answers and correct the process, not just the model manual.
- Establish monitoring, costs, retention and shutdown procedure.
- Expand only after quality and safety criteria are met.
Summary
An AI assistant can shorten the path to information if it works with structured materials and respects user rights. You should not promise to answer solely on the basis of documents or always admit that you do not know something until such behaviour has been thoroughly tested.
NexaIT designs AI assistants and automation after analysing data, risk and expected quality. Contact usif you want to prepare a safe pilot.