Asking your company documents answers that come with their source
For whoever has manuals, procedures, data sheets, drawings and non-conformance reports scattered across shared folders and business software: what each department asks its documents, when the assistant you already have is enough and when it is not, how search works, who sees what and what it costs to start.
Can AI answer questions using your company's documents?
Yes, and the answer can come with the document it is taken from. Whoever asks writes the question as they would to a colleague, for example «what tightening torque does the procedure set for assembly 40213?», and the assistant answers citing the procedure, the revision and the passage, with a link to the file. Nobody needs to know which folder the document is in or what it is called.
The technique is called RAG, retrieval-augmented generation. First a search engine finds in your documents the passages that answer the question; then those passages go to the language model (LLM) together with the question, and the model writes the answer using them and citing them. The model does not need to have learned your documents, and it does not learn them: it reads them at the moment, every time, and only those the person asking can open.
That gives the two things that decide whether it works. The first is search: if the engine does not find the right passage, the model gives a good answer to the wrong question. The second is permissions: the assistant sees what the person asking sees, no more and no less. Most of this guide is about those two.
What does each department ask its documents?
Each department has its own documents and asks its own questions, and a useful assistant is built starting from those. This is the map of a manufacturing company; a professional firm or a services company has different departments, but the logic is the same.
| Department | Documents | Typical questions |
|---|---|---|
| Engineering | drawings, bills of materials, specifications, standards | which drawing revision is valid today? which standards does the customer's specification cite? |
| Quality | procedures, non-conformance reports, corrective actions, audits | have we had this defect on this item before, and what did we do? |
| Production and maintenance | machine manuals, setup sheets, service reports | what does the manual say about this alarm? |
| Purchasing | framework contracts, specifications, supplier certificates | does the material have a 3.1 certificate? |
| Sales and quotes | past offers, price lists, terms and conditions | have we made a similar part before, and at what price? |
| Administration | contracts, payment terms, insurance policies | what penalty does the contract with this customer set? |
| Health, safety and environment | risk assessments, contractor risk documents, safety data sheets | what protection does this operation require? |
| People | internal rules, collective agreements, training | how many training hours does this role need? |
Two things already show in the table. One department's documents are often useful to others: non-conformance reports serve purchasing, past offers serve engineering. And not everyone can see everything: contracts, payslips and performance reviews stay with those entitled to them. The assistant is the knowledge that supports the department: it answers with the documents the department really uses, and gives each person only those they can open.
Are ChatGPT or Copilot enough?
Sometimes they are, and it pays to check before a project. The assistants many companies already have read documents in the services they connect to, respect the source's permissions and do not use business data to train models. According to Microsoft's documentation, Copilot only surfaces organizational data the user can at least view, and prompts, responses and data accessed are not used to train the models; according to OpenAI's help center, with ChatGPT Business a member retrieves only information they can already access in the connected source.
| Route | When it is enough | Price on 13 November 2026 |
|---|---|---|
| Microsoft 365 Copilot Business | documents in Microsoft 365 or in sources with a connector | 18.20 € per user per month, 15.60 € in the first year |
| ChatGPT Business | documents in connected apps, in formats it reads | 21 € per user per month |
| Custom system | documents also in business software, file servers or scans; exact codes; permissions by role | feasibility study up to 2,500 € |
Copilot Business is an add-on to a Microsoft 365 Business plan, up to 300 users, with an annual subscription and VAT excluded, on the Italian price list; the 15.60 € price is an offer valid until 31 December 2026 (Copilot Business price list, Microsoft Learn). It reads documents in Microsoft 365 and in external sources linked through more than 100 connectors: check that your documents outside Microsoft 365 have one, and that questions with an exact code find the right document. For large companies Microsoft 365 Copilot costs 26.00 € per user per month paid annually or 27.30 € paid monthly (Microsoft 365 Copilot price list). ChatGPT Business costs 21 € per user per month as OpenAI's page shows it in Italy, in dollars 20 billed annually and 25 monthly, for teams of 2 to 200 people; in SharePoint search it reads .txt, .pdf, .docx, .pptx, .xlsx and .csv up to 100 MiB per file, and may fail to extract text from some files (SharePoint app guide).
The choice depends on where the documents are and what you ask. If they live in SharePoint and Teams, with folders and permissions in order, and questions are in everyday language, the seat-based assistant is the fastest route. A custom system is needed when part of the documents sits in the business software, on a file server or in a scan archive; when questions contain item, standard or drawing codes that must be found exactly; when permissions follow jobs, customers or business software roles rather than folders; when documents cannot leave your servers. Before a project, test the tool you already have with a few dozen real questions from the department: if it answers with the right document, you do not need the project.
Would you rather talk it through?
Book a call with people who work on these topics every day: we will look at your case together.
How does it find the right document?
With a search engine that works in two ways, by words and by meaning, and merges the results. This is the part that decides the quality of the answers, more than the model that writes them.
Keyword search compares the terms of the question with those of the documents. The most widely used algorithm is called BM25: according to the Azure AI Search documentation on BM25 scoring, it favours documents that contain many instances of the search term, and even more so if the term is rare across the index but common within the document. It is the one that finds «UNI EN ISO 9001», «40213 rev. C» or the number of a non-conformance report, because a code is a rare and precise word.
Semantic search turns each passage into a vector, a list of numbers that represents its meaning, computed by an embedding model; the question becomes a vector the same way, and search takes the closest passages. It finds a document even when it does not contain the words of the question: «how do I adjust the mould pressure» finds the page on calibrating the clamping force. For the same reason it can miss an exact code, which in meaning resembles many others.
Hybrid search runs both searches together and merges the two rankings. The Azure AI Search documentation on hybrid search describes it like this: the two searches run in parallel and results are merged with Reciprocal Rank Fusion (RRF), which lifts documents ranked high in both; the same page notes that product codes, specialized jargon, dates and people's names are found better with keyword search.
| Search | What it finds well | Where it fails |
|---|---|---|
| Keyword (BM25) | codes and acronyms: «NC-2025-118» | synonyms, wording unlike the document |
| Semantic (vectors) | everyday questions: «why does the blade break» | exact codes and numbers |
| Hybrid | both: «does procedure 7.3 cover welds?» | one more index to keep updated |
In an industrial archive we made searchable, 225 GB of regulations, HSE, safety and operations documents coming from four different archives, search is hybrid: BM25 for exact terms such as a standard number, vectors for meaning. Next to the indexed passages the system keeps the original files, so the reader opens the whole document and not just the fragment. We design the assistant so that every answer reports the documents it comes from, with a link to the original file, and says it cannot find the answer when the documents do not contain it. Developers will find the comparison with other architectures in the article on GraphRAG and hybrid RAG and an example with code in the one on RAG with Mastra.
Drawings, scans, revisions: where does it fail most often?
Where documents are not clean, orderly text. An assistant answers as well as the archive it reads is legible and orderly, and in manufacturing companies the difficult cases are always the same.
| Case | What happens | What you do |
|---|---|---|
| Scanned documents | the PDF is an image, search finds nothing | the text is read with OCR, optical character recognition |
| Technical drawings | geometry is not text | title block and notes are read: code, revision, material |
| Several revisions of the same document | the answer mixes rev. B and rev. C | each document carries its status, valid or superseded |
| Duplicates in several folders | the same procedure answers three times | duplicates are recognised and one source is kept |
| Tables and spreadsheets | split into passages, they lose their headers | they are indexed row by row, or read as data |
| Counts and totals | search returns the most relevant passages, not all of them | the count is asked of the business software, as in the next section |
Preparation work is often the largest item in a project, and it is estimated in the feasibility study on a sample of your documents, not by eye. An archive is put in order once, with written rules on what is valid and what is superseded, and those rules stay yours even if you change tool.
How does it connect to the business software, and who sees what?
Folders are indexed, the business software is queried. Documents, which change little, are read once and kept in sync when someone adds or edits one; data, which change constantly, are asked of the business software at the moment of the question, through its APIs. The stock of an item or the status of an order should not be copied into an index, where they would already be stale tomorrow.
To bring the two together the assistant becomes an agent: a program that, faced with the question, decides which tools to use, for example the document search, a read from the business software or a count of non-conformance reports, and puts the answer together. The open standard through which an agent calls these tools is called MCP, Model Context Protocol. It is the same one Microsoft uses for Copilot's federated connectors, which retrieve content in real time without indexing it (connectors documentation): a system built on MCP stays connectable to the assistants the company already uses.
The document is searched in the archive, the data is asked of the business software at the moment of the question: that way stock levels or an order status are never last week's. How to build an agent that works on the company's systems is in the guide on AI agents for SMEs; when the goal is not answering but getting work done, from email to business software, the right guide is the one on process automation.
How do you make sure each person sees only their documents?
By applying permissions in the search, before the model sees anything. Each indexed passage carries who can open the document it comes from; when someone asks, search considers only passages from documents that person can open. A passage that does not pass the filter never reaches the model, so it cannot end up in the answer, not even summarized. It is the model Microsoft describes for Copilot and that the Azure AI Search documentation calls document-level access control, from ingestion to query.
Permissions are inherited from where they already are: shared folders, the company directory with its groups, the roles of the business software (RBAC, role-based access control). In the industrial archive above each role sees its own part, and a confidential document does not enter the answers given to someone who cannot consult it. Permission is applied before searching, not after answering: what you cannot open, the assistant does not read for you.
Two consequences are worth knowing before you start. First, an assistant makes findable what was previously only accessible. A folder of contracts left open to everyone by mistake, which nobody went looking for, becomes an answer for anyone who asks about that customer; the permission review comes before switching the assistant on, not after. Second, a permission removed in the source applies to answers when the index is synchronized, as the same Azure documentation says. How often it is synchronized is a question to ask anyone proposing a system, us included.
What do GDPR and the AI Act require?
GDPR comes in when documents contain personal data: CVs, payslips, fitness-for-work certificates, but also non-conformance and service reports with people's names. The assistant processes them as the archive does: you need a legal basis, access limited to those who need it and a contract with the model provider stating what it does with the data. The simplest choice is to keep HR documents out of the index, or reserve them for the role that manages them. Where the data go and with what guarantees is in the answer on how we use company data and in the guide on AI in the cloud or on your servers.
The AI Act, in article 4 as amended by Regulation (EU) 2026/1744, requires those who use AI systems to take measures to support the AI literacy of their staff, without imposing a specific level. For a document assistant this means explaining to its users what it does, where it takes answers from and why the cited document should be opened before deciding. The other obligations of those who use AI are in the guide on the AI Act for deployers.
What does it cost, and where do you start?
The first step has a fixed price: the feasibility study costs up to 2,500 €, takes one to two calendar weeks and is deducted from the pilot if you go ahead, as in the guide on how to start with AI in your company. The study looks at a sample of your documents, the departments' questions, the sources to connect and the permissions, and ends with the quote for the pilot. The cost of the project depends on a few things, which the study measures.
| What | Costs less when | Costs more when |
|---|---|---|
| Documents | a few thousand, in PDF or Word with text | hundreds of thousands, in different formats |
| Scans | the text is already there | scans to read, including handwritten or low quality |
| Sources | one shared folder | folders, business software, email and different archives to keep in sync |
| Permissions | the same per folder | per role, job or customer, inherited from the business software |
| Revisions and duplicates | an archive already in order | versions to recognise and rules to write |
| Where it runs | cloud, with model providers that do not retain data | your own servers, with a dedicated graphics card |
| Languages | one | documents and questions in several languages |
Then there are running costs. The model is paid per use, and for an archive of this kind it costs little: in the example of the guide on AI in the cloud or on your servers, a 60-person company with about 5,000 documents and 2,000 questions a month, our estimate is about 2.50 € once to index the documents and from $3 to $120 a month for model use, depending on the model chosen, at the price lists read on 30 October 2026. Hosting and maintenance come on top. For comparison, seat-based assistants for the same 60 people cost 60 × 18.20 = 1,092 € a month with Copilot Business, 936 € in the first year with the offer, on top of the Microsoft 365 plan, and 60 × 21 = 1,260 € a month with ChatGPT Business, at the Italian price lists of 13 November 2026.
On the software we build for you there is no per-user licence fee: recurring costs are hosting, model use and maintenance, stated in the quote (recurring costs). Where and how models run, with the costs of each route, is in the guide on AI in the cloud or on your servers.
Where do you start?
With one department and a list of questions. You choose the department where searching for documents costs the most time, usually engineering or quality, and the people who work there write the real questions they ask every week, each with the right answer and the document it comes from. That list becomes the measure of the pilot: how many answers cite the right document, how many say they do not know when they should, how many are wrong.
The feasibility study takes one or two weeks and the pilot, on that department's documents from recent years, four to six. At the end you have a result measured on your questions, which tells you whether to extend to other departments and with which documents. In the industrial archive above it took four months to go into production, and in the first phase the system was used by 20 people before opening it to the others. You extend one department at a time, and each department comes in with its own questions and its own measure, not with a promise.
An example: the engineering department of an 80-person company
An 80-person engineering company has an engineering department of six people and a quality team of three. Drawings are PDFs on a file server, with the revision in the file name; ISO 9001 procedures are in a shared folder; non-conformance reports are in the business software, with attachments; supplier data sheets are a bit everywhere. The details are invented, to be replaced with yours.
The feasibility study, on a sample of documents, finds that a third of the older drawings are scanned and that many procedures exist in two versions. The pilot starts with engineering and quality, on the last three years of documents: scans go through OCR, drawings are indexed by code, revision and treatment read from the title block, superseded procedures are marked as such. Permissions come from the company directory groups; non-conformance reports stay in the business software and the agent reads them from there when a count is needed.
The pilot questions are those written by the department: «what heat treatment is set for 40213, and from which revision?», «have we had cracks on this casting before, and what corrective action did we take?». The measure shows where the assistant answers with the right document and where it does not, and on that measure the company decides whether to extend to purchasing and production. How we built an archive of this kind for an industrial company is in the case study on document RAG for industry.
Sources and further reading
- Microsoft, «Microsoft Copilot per le aziende», Microsoft 365 Copilot Business prices for Italymicrosoft.com, in Italian, accessed 13 November 2026
- Microsoft, «L'IA per la produttività aziendale», Microsoft 365 Copilot prices for Italymicrosoft.com, in Italian, accessed 13 November 2026
- Microsoft Learn, «License Options for Microsoft Copilot»learn.microsoft.com, accessed 13 November 2026
- Microsoft Learn, «Data, Privacy, and Security for Microsoft Copilot»learn.microsoft.com, accessed 13 November 2026
- Microsoft Learn, «Microsoft 365 Copilot connectors overview»learn.microsoft.com, accessed 13 November 2026
- OpenAI, ChatGPT Business, pricingopenai.com, accessed 13 November 2026
- OpenAI Help Center, «Company knowledge in ChatGPT»help.openai.com, accessed 13 November 2026
- OpenAI Help Center, the SharePoint app in ChatGPThelp.openai.com, accessed 13 November 2026
- Microsoft Learn, Azure AI Search, hybrid searchlearn.microsoft.com, accessed 13 November 2026
- Microsoft Learn, Azure AI Search, BM25 scoring in full-text searchlearn.microsoft.com, accessed 13 November 2026
- Microsoft Learn, Azure AI Search, document-level access controllearn.microsoft.com, accessed 13 November 2026
- Regulation (EU) 2024/1689, AI Act, article 4, as amended by Regulation (EU) 2026/1744ai-act-service-desk.ec.europa.eu, in Italian, accessed 13 November 2026
Can AI answer using my company's documents?
Yes. With the technique called RAG the assistant searches your documents for the passages that answer the question and writes the answer using only those, with the document and the passage it comes from. The documents are not learned by the model but read every time, and only those the person asking can open.
Is Microsoft 365 Copilot enough to make documents answer?
Often yes, if documents live in Microsoft 365 with permissions in order and questions are in everyday language. On the Italian price list Copilot Business costs 18.20 € per user per month, 15.60 € for the first year with the offer until 31 December 2026, VAT excluded and on top of the Microsoft 365 plan. A custom system is needed when documents also sit in the business software or on a file server, when exact codes matter or when permissions follow jobs and roles.
What is the difference between keyword, semantic and hybrid search?
Keyword search, with the BM25 algorithm, finds exact terms such as an item code or a standard number. Semantic search, on vectors, finds a document even when it uses words different from the question. Hybrid search runs both and merges the results, so it finds both «40213 rev. C» and «how do I adjust the mould pressure».
Does everyone using the assistant see all documents?
No, if the system is built well. Permissions are applied in the search, before the model reads anything, and come from folders, the company directory or the business software roles. That is why permissions should be reviewed before starting, because the assistant also makes findable what was left open by mistake.
Are the documents used to train the models?
No. Microsoft and OpenAI state that business data from Copilot and ChatGPT Business do not train the models, in OpenAI's case unless the company chooses otherwise. In our projects we use services that exclude it by contract, and services that do not retain data, or your own servers, when more is needed.
How much does an assistant over company documents cost?
You start with a feasibility study, which costs up to 2,500 € and takes one or two weeks. The project cost depends on how many documents there are and in what format, on scans to read, on sources to connect and on permissions. Model use costs little, in our estimate $3 to $120 a month for a 60-person company with 2,000 questions a month, plus hosting and maintenance.
Does it work with technical drawings and scanned documents?
Yes, with some preparation. Scans are read with OCR, optical character recognition. From a drawing the title block and notes are read, so the assistant finds it by code, revision and material; dimensions and geometry stay in CAD.
How long does it take to start?
Five to eight weeks for a result measured on one department, one or two of feasibility study and four to six of pilot. Extending to the whole company takes longer, depending on documents and sources; for an industrial archive of 225 GB it took four months to go into production.
Documents scattered across folders and business software?
Tell us how they are organised, in an introductory call, free and with no commitment: we tell you whether the tool you already have is enough or what it would take to make them answer.
- You won't talk to a salesperson
- NDA before we get into the details
- The code stays yours
Articles on the same topics.
GraphRAG or hybrid RAG: when the graph is worth what it costs
The graph wins on questions that connect several documents and ties on point questions, where it costs much more. Benchmarks, costs and tools as of November 2026.
Read
Web Bot Auth: visitors with no eyes
AI agents read your site and convert better than organic, yet they never show up in analytics. How to measure them and verify them with Web Bot Auth.
Read
Generative UI: for humans and machines
Explore how Generative UI and Dual-Interaction UX transform enterprise SaaS by creating adaptive interfaces for human operators and AI agents.
Read