Brussels consultancies carry a kind of knowledge that is hard to search and risky to expose. Proposal language sits in slide decks. Methodology notes live in shared drives. Client context is split across email, CRM records, workshop notes, delivery documents, and partner conversations. A consultant can usually find the right answer after enough digging, but the search depends on memory, access, and who happens to be online.
A private AI knowledge base can help, but only if it is built as an operating system for internal knowledge rather than a chatbot connected to every file. For a consultancy, the goal is not to make the model sound confident. The goal is to help a team retrieve approved knowledge, preserve client confidentiality, respect permissions, and show where every answer came from. That is a different design problem from buying a public AI assistant and uploading documents.
This guide is for Brussels consulting, advisory, and professional-services teams that want retrieval-augmented generation for internal work: proposal support, methodology reuse, delivery guidance, onboarding, research summaries, and account context. It focuses on the workflow, data boundary, controls, and measurement plan needed before a private AI knowledge base is useful in production.
Start with the knowledge workflow
Do not begin by indexing the whole drive. Begin with one recurring question set. A consulting team might start with "What have we already written about AI readiness for Belgian SMEs?", "Which delivery checklist applies to a data-mapping workshop?", or "What should a new account manager know before a client renewal call?" Each question set should have a named user, a known source of truth, and a decision the answer supports.
Map the current path from question to answer. List who asks, where they search, what documents they trust, which folders are off limits, who approves reusable wording, and what happens when the answer is uncertain. This map often reveals that the technical retrieval problem is smaller than the governance problem. Teams do not only need semantic search. They need document ownership, version control, permission cleanup, citation rules, and a feedback loop for wrong answers.
For Brussels consultancies working with Belgian or EU clients, this workflow map should include language and jurisdiction. Some material may be client-confidential, some may be internal methodology, and some may be public source material. French, Dutch, and English documents may describe the same service in different words. A useful knowledge base needs metadata that preserves those distinctions instead of flattening every document into one undifferentiated index.
Choose the first corpus carefully
The first corpus should be valuable, bounded, and maintainable. Good candidates include approved methodology documents, delivery checklists, service descriptions, policy templates, anonymised project retrospectives, public research notes, and internal FAQs. Weak candidates include raw email archives, unrestricted client folders, old proposals with outdated pricing, and documents with unclear ownership.
Use a simple classification before ingestion. Mark each document as public, internal, client-confidential, personal-data-bearing, or excluded. Record the owner, update cadence, language, intended audience, and retention expectation. If the source is SharePoint, Google Drive, Notion, Confluence, a CRM, or a file server, keep that system as the source of truth. The AI index should reflect governed source content, not become an unmanaged copy that slowly drifts away from it.
EU data-protection rules make this step practical rather than academic. The European Commission's data-protection guidance describes the legal framework around personal data, rights, and responsibilities in the EU. For a consultancy knowledge base, that means the team should decide whether personal data belongs in the first corpus at all, what legal basis applies, how access is controlled, and how deletion or correction in the source system propagates to the index.
Design permissions before prompts
A private knowledge base is only private if permissions survive retrieval. It is not enough for the application login to know who the user is. The retrieval layer must respect what that user can access in the source systems or in a maintained entitlement map. A junior consultant should not receive board-level account notes because those notes shared a keyword with a harmless methodology page.
The minimum access model has four parts. First, authenticate the user through the company's identity provider. Second, attach user, team, client, and role metadata to every request. Third, filter retrieval results before the model sees them. Fourth, log the question, retrieved document IDs, answer, user, and feedback without storing unnecessary sensitive content. This creates an audit trail without turning the log into another uncontrolled knowledge store.
NIST's AI Risk Management Framework is useful because it frames AI work around mapping, measuring, managing, and governing risk. In this context, mapping means knowing the knowledge sources and users. Measuring means testing answer quality and access control. Managing means adding refusal, escalation, and correction paths. Governing means assigning ownership for the corpus, the model configuration, and the review rhythm.
Build the answer pattern
For consultancy work, the safest answer pattern is concise, cited, and bounded. The assistant should answer from retrieved sources, name the documents it used, state uncertainty, and refuse when the corpus does not support the answer. It should avoid inventing client results, pricing, legal conclusions, or commitments. When the user asks for a proposal paragraph or workshop plan, the system can draft, but the consultant remains responsible for review before anything reaches a client.
The prompt is only one layer. The application should also enforce source thresholds, citation requirements, protected-action rules, and handoff paths. If the top retrieved documents are weak, stale, or inaccessible, the system should say so. If a question touches legal advice, HR decisions, financial commitments, or client-confidential material, it should route the user to the right owner or require human approval before reuse.
OWASP's large-language-model application work is a useful reminder that these systems introduce application risks as well as model risks. Prompt injection, insecure output handling, excessive agency, sensitive-information disclosure, and weak supply-chain controls all matter when the assistant can read internal content or trigger workflow actions. A knowledge base should therefore start read-only, then earn any ability to create tasks, update CRM fields, or send documents.
Implementation path for a Brussels team
- Pick one use case. Choose a question set used weekly by consultants, delivery leads, or account managers. Avoid company-wide search as the first milestone.
- Clean the source corpus. Remove obsolete files, assign owners, classify sensitivity, and confirm that source-system permissions reflect the intended audience.
- Create the retrieval index. Chunk documents by section, keep metadata with every chunk, and store source URLs, titles, owners, language, and last-updated dates.
- Add the control layer. Enforce permissions before retrieval, require citations, block unsupported answers, and log enough detail to review mistakes.
- Run an evaluation set. Test real questions with known answers, cross-language phrasing, stale-document traps, and access-control checks before launch.
- Launch to one team. Start with read-only answers inside Slack, Teams, or a lightweight web interface. Collect feedback and correction requests for each answer.
This is the kind of implementation Intyb usually places under enterprise knowledge systems rather than generic workflow automation. The same retrieval layer can later support custom AI applications, but the first value comes from making existing knowledge findable with clear controls.
Where it works and where it does not
A private AI knowledge base works well when the content is mostly text, the source documents are owned, the questions are recurring, and the answer can cite evidence. It is especially useful for delivery playbooks, proposal reuse, onboarding, internal policy lookup, technical notes, and preparation for client conversations.
It works poorly when the company wants the assistant to guess strategy, replace expert judgment, or search material that has not been cleaned. It also struggles when critical knowledge is trapped in personal inboxes, undocumented calls, or client files with unclear confidentiality boundaries. In those cases, the right first project may be process design and document governance, not model deployment. The same principle appears in Intyb's Brussels AI workflow readiness scorecard: the process must be ready before the AI layer can be trusted.
The EU AI Act also pushes teams to think about role, purpose, transparency, and oversight. Many internal knowledge assistants will not be high-risk systems, but consultancies should still document the intended use, user groups, limitations, and human review points. That record is useful for clients, employees, and future audits, and it should sit beside the firm's wider AI compliance approach.
Measurement plan
Measure the knowledge base as a workflow, not as a novelty. Before launch, capture baseline search time, repeated questions, onboarding friction, proposal-reuse effort, and the number of cases where staff cannot identify the current source of truth. After launch, review answer usefulness, citation accuracy, unsupported-answer rate, permission failures, stale-source reports, and correction turnaround.
A practical first target is not "the AI answered everything." It is "consultants found approved material faster and could verify the source." Review a sample of answers every week during the pilot. Keep a queue of documents to add, update, or remove. If users keep asking unsupported questions, either expand the corpus deliberately or make the assistant's boundary clearer.
For Brussels consultancies, the strongest implementation path is controlled and incremental: one team, one corpus, one answer pattern, and one review rhythm. Teams that want help choosing that first slice can speak with Intyb's Brussels implementation team. For the cautionary side of the same decision, read why bad process automation costs more.