A company wants to build an internal AI assistant so its team can consult procedures, contracts, technical specifications or customer data using natural language questions. The idea seems straightforward, but a problem emerges before development even begins: the information is scattered across folders, emails, the ERP, the CRM and duplicate documents.
In this context, connecting an artificial intelligence model directly to every source will not resolve this disorder. It may make the problem less visible, but it could also return answers based on outdated documents, combine conflicting versions or show information to people who should not have access to it.
Therefore, before developing an AI system that works with internal documentation, the company needs to make an architectural decision: determine where each type of information lives, which source is authoritative, how it is updated and what kind of environment the people who need to consult it actually require.
The problem is not just finding documents
When someone searches for information manually, they usually apply criteria that a system cannot simply assume. They may recognise the most recent file, know that a procedure is obsolete or ask a colleague whether anything has changed.
An AI assistant needs these rules to be defined. If it finds three versions of the same document, it will not necessarily know which one is current. If a CRM record contradicts a spreadsheet, it needs to know which source takes priority. And if important information only appears in an email thread, the company should first decide whether it belongs in its shared knowledge base.
The most common issues include:
- Duplicate documents and no version control. It is unclear which file is official or who updated it.
- Information spread across different systems. Sales data may be in the CRM, product information in the ERP and procedures in shared folders.
- Poorly defined permissions. Not all documentation should be available to every person or department.
- Content without an owner. No one is assigned the responsibility of reviewing and maintaining it.
- Knowledge that only exists in emails or conversations. This information is difficult to validate, structure and reuse.
This scenario does not invalidate the project, but it indicates that the first phase should not be configuring the AI model. It is necessary to start by analysing the information, the systems and the actual use cases. A digital and technical consultancy before execution allows for defining priorities and avoids building the assistant on an unstable foundation.
Once the problem has been defined, the development of an internal platform and its integrations can turn this architecture into a useful, secure tool that is ready to evolve.
Is it necessary to organise documents, integrate systems or create an intranet?
There is no single answer. These are three different and, in many projects, complementary actions. The decision depends on where the reliable information is located and what the team needs to do with it.
When should you start by organising documentation?
You should start with documents when the company’s primary knowledge resides in PDFs, manuals, protocols, presentations or office files. Before connecting the AI, it is advisable to eliminate duplicates, identify current versions, establish a coherent classification and assign maintenance owners.
This does not mean having to rewrite all the documentation. The goal is to have a controlled repository with basic metadata such as document type, department, review date, version and access level.
An AI can retrieve relevant fragments and draft a response based on this content. This architecture is commonly known as RAG, which stands for Retrieval-Augmented Generation. But the system will continue to depend on the quality, currency and consistency of the sources it consults.
When should the ERP and CRM be integrated with the AI?
Integration is necessary when questions require current, structured data that already resides in management systems. For example, the status of an order, a customer’s commercial conditions, available stock or open incidents should not be periodically copied into documents.
In these cases, an ERP CRM integration with AI can allow the assistant to consult authoritative data via APIs—that is, controlled connections between applications. The assistant interprets the question and presents the information, but the data remains in the system that manages it.
This distinction is important: AI does not replace the ERP or the CRM. Nor should it be used to synchronise fields or execute deterministic rules that can be better resolved with conventional programming, automations or webhooks. The article on architecture and criteria for applying AI agents delves deeper into this separation between models, tools and integrations.
When does an AI-powered intranet make sense?
An intranet is appropriate when, in addition to asking questions, the team needs to navigate, validate sources, access tools and manage processes from a common environment. The assistant becomes a feature within a broader platform, not the entire platform itself.
An AI-powered intranet can bring together a search engine, documents, a directory, notices, forms, procedures and access to other applications. It can also show the sources used in each response so that the user can verify them and open the original document.
The design of this environment is relevant because business knowledge management is not just about storing information. It must be organised, easy to consult and adapted to different profiles. Here, clear, user-centred digital environment design helps prevent the intranet from becoming just another folder that is difficult to navigate.
How do you define a reliable foundation for the assistant?
Before deciding on a specific technology, it is advisable to create a source map. This inventory should indicate what information exists, where it is located, who is responsible for it, how often it changes and which profiles can consult it.
From there, specific decisions can be made:
- Define the source of truth. For each type of data, there must be a primary source: the ERP for the catalogue and stock, the CRM for commercial activity or the document repository for procedures.
- Separate stable content and dynamic data. A manual can be indexed, while the status of an order must be consulted in real time or via controlled synchronisation.
- Establish permissions. The assistant’s access must respect the organisation’s roles. Being able to ask a question does not imply having the right to see the answer.
- Show the sources. Whenever possible, the response should indicate where the information comes from and allow the original document or record to be reviewed.
- Plan for maintenance. It is necessary to define what happens when a document is updated, an API changes or a new tool is incorporated.
It is also advisable to assume that the generated response may require supervision, especially in legal, technical, labour or financial matters. As with AI-generated content that needs reviewing, a well-written response is not necessarily a correct response.
What is the minimum viable architecture?
The best first version is not the one that connects the entire company, but the one that solves a defined use case with reliable sources. You can start, for example, with the procedures of one department, a specific document base or a set of well-defined commercial queries.
This first phase allows you to check whether people find the answers useful, which questions are left uncovered and where quality or permission issues arise. It also helps in deciding whether to expand the repository, integrate another system or develop specific functionalities.
A reasonable architecture can combine a controlled document repository, connections with ERP or CRM for specific data, an authentication system, role-based permissions and an internal interface that shows answers and sources. Not all projects need all these components from the start.
The important decision is not which AI model will be used, but how information will flow and who will guarantee its quality. When this foundation is well-defined, the assistant stops being an added layer over scattered documents and becomes a useful piece within the company’s knowledge system.
If you want to assess how to apply this to your specific case, write to us and we can look at it together.