For general management, HR, quality and area managers who see their team waste time every day searching for information in an internal documentation scattered across SharePoint, network folders, Teams, Outlook and Wikis. A conversational assistant that answers in seconds with a citation to the source document, embedded in Microsoft Teams and accessible from a browser.
At most mid-market companies, internal documentation exists, but it is scattered. Product manuals in SharePoint, quality procedures in a shared server folder, supplier contracts in the operations director folder, project precedents in two-year-old Teams threads, undocumented FAQs that only live in three people's heads, decisions taken six months ago discussed by email and nobody remembers where. The information is there, but it cannot be found.
New employees do the reasonable thing: they ask the seniors. Each senior spends between thirty and sixty minutes a day answering questions whose answer is already documented somewhere in the corporate corpus. That is expensive people repeatedly answering the same thing several times a week. Frustrating for the senior; slow for the newcomer; invisible to management until it is measured.
The structural problem is the dependency on specific people. If an area manager leaves, part of the operations leaves with them. If the veteran technician goes on holiday, some questions do not get answered until they return. The operational knowledge of the company is, without intent, hosted in heads, not in consultable documents. That turns every onboarding into a slow and costly race, and every departure into a silent risk.
The fourth front is uniform answer quality. When the query depends on the available senior, each answer varies: judgement changes, memory of precedents changes, daily patience changes. Management cannot guarantee that two employees asking the same question to different people get equivalent answers. That noise, multiplied across operations, has a real cost in compliance, training and quality of service to the end customer.
The core of the case is a conversational RAG (Retrieval-Augmented Generation) assistant trained on the customer corporate documentation. The employee asks in natural language ("What is the procedure to onboard a new supplier?", "What did the contract with supplier Y say about warranties?", "How do we handle progress retentions in Andalusia?") and gets an immediate response, drafted in the same language as the question, with explicit citations to the source documents.
Technically, each document in the corporate corpus is chunked, vectorised with Azure OpenAI embeddings and indexed in Azure AI Search. When a question arrives, the engine retrieves the most relevant passages by semantic similarity (not just keywords: it understands synonyms, paraphrasing and context), passes them to a language model along with the question, and the model generates a synthetic answer based exclusively on what the documentation says. If the information is not in the corpus, the assistant says so: it does not make things up.
Access to the assistant happens where the user is. The main integration is as a Microsoft Teams app: the employee asks from the usual chat, with no new tools to open, no prior training. It is also available from a web browser (for guests or employees without Teams) and, optionally, from Outlook as an add-in to resolve doubts while drafting an email or preparing a customer response.
Each assistant response includes citations of source documents, with a direct link to the original file in SharePoint. That is key for trust: the employee can verify the answer before applying it, especially on critical queries (contracts, legal procedures, internal regulations). And for management it is traceability: it is logged what was asked, what was answered and which sources were used, without storing sensitive data outside the tenant.
Optionally, we connect the assistant to transactional data from Business Central through the AI Studio native integration. In that extended mode, the assistant not only answers about static documentation ("How do we handle X?") but can also query live ERP data ("What is the outstanding balance of customer Y?", "What open projects does the Seville branch have?"). It is optional because it adds complexity and is recommended only after the documentary RAG is stable.
| Process aspect | Before (no assistant) | After (with RAG AI) |
|---|---|---|
| Response time to internal question | The employee asks a senior, waits hours or days depending on availability. Sometimes does not ask and resolves badly. | Response in seconds from Teams or browser, with citation of the source document to verify. |
| Dependency on key people | If the area manager or the veteran technician leaves, part of the knowledge goes with them. | Knowledge stays in the indexed corpus. People remain important; they are no longer unique. |
| Onboarding new members | Two or three months of constant questions to seniors to understand procedures and precedents. | The newcomer asks the assistant basics from day 1. Seniors only handle complex cases. |
| Uniform answer quality | Each senior answers based on their judgement, memory and patience level on the day. | Same procedure, same answer. Human interpretation is reserved for edge cases. |
| Query traceability | No log of what was asked, by whom, with what response. The same question is repeated a thousand times. | Each conversation is logged with source citations. It detects which documents are not understood. |
| Structured knowledge | Documents scattered across SharePoint, folders, Teams, Outlook, Wikis. Nobody knows where anything is. | A single conversational entry point. The physical file structure no longer matters. |
| Senior workload | Between 30 and 60 minutes a day answering questions already documented elsewhere. | That time recovered for substantive work. Seniors only handle the escalated question. |
Audit of documentary sources (SharePoint, network folders, Teams, Wikis), map of the most frequent query areas, identification of duplicates and obsolete documentation, proposal of minimum viable curation and sizing of the initial corpus.
Deployment of Azure AI Search and Azure OpenAI in the tenant, initial indexing of the pilot corpus, agent configuration with Copilot Studio, integration as a Teams app, training of a pilot group, prompt and threshold tuning with real feedback.
Expansion to more areas and more documentation, optional integration with live Business Central data, Outlook add-in, response quality monitoring, incremental retraining with user feedback.
Some scenarios do not pay back or are not viable. We say it directly.
Sector-specific RAG for construction: certifications, retentions, critical work items on site.
AI caseA monthly note drafted by AI on financial performance, with explained variances.
AI caseTechnical copilot for your IT team diagnosing BC incidents and proposing customisations.
BC extensionDocumentary viewer for BC. Useful when the RAG needs to connect with documentation managed in the ERP.
AI sectorSpecific AI cases for wealth managers: close, anomalies, document query (in Spanish).
HubFull catalogue of cases, sectors, method and discovery of the AI Studio (in Spanish).
It is not a blocker, but it conditions the scope. The RAG engine learns to search in what exists; if the documentation is scattered, duplicated or outdated, the answers inherit that noise. In the discovery we audit the sources (SharePoint, network folders, Teams, Wikis), detect overlaps and propose a minimum viable curation before the pilot: marking the current version, removing drafts, tagging by area. It is not rewriting everything: it is sorting just enough for the assistant to be reliable from day one.
No. Deployment runs on Azure OpenAI Service and Azure AI Search inside your Azure subscription, with vector indexes hosted in your tenant. Neither the documents nor the embeddings are sent to public OpenAI, nor are they used to train third-party models. The assistant responses are generated using passages retrieved from your documents and returned with an explicit citation to the source file, so the user can verify.
The initial indexing of a typical mid-market corpus (a few hundred to a few thousand documents across SharePoint and network folders) takes between one and three days. Each time a new document is uploaded or an existing one is modified, the index updates automatically in minutes. There is no model training: the RAG engine does not learn the documents by heart, it queries them live through the vector index. That means any change in the documentation impacts the responses immediately.
Yes, those are the two main native integrations. SharePoint acts as the primary document source (automatic indexing of selected libraries, respecting the querying user permissions). Teams hosts the assistant as a conversational app, accessible from the employee usual chat, with nothing to install. It is also available from a web browser and, optionally, as an Outlook add-in to resolve doubts while drafting an email.
The assistant understands and responds in Spanish, Catalan, Galician, Basque, English, French, German, Portuguese and Italian with no additional configuration. The documentation can be in any language supported by Azure OpenAI: the user asks in their language, the engine retrieves in the source language of the document and translates the answer. Useful when there are legacy procedures in English or contracts in several languages.
We frame the assistant within your current Microsoft tenant relationship and the documentation we already manage. Your usual advisor coordinates with the AI Studio.
Talk to the team →We start with the 5-day discovery. We audit the real state of your documentation, size the pilot corpus and deliver a minimum viable curation plan.
Request AI discovery →