The short answer
An AI assistant on company documents answers by searching the company's manuals, procedures, quotes and certificates, and cites where each answer comes from: the technique is called RAG and does not retrain the model. It works on three conditions: readable documents in a single version; archive permissions in order, because it shows each person what they can already see; a quality measure based on real questions, before you extend it.
- No retraining: the assistant searches the authorised documents and answers by citing document and page.
- Three conditions: readable documents in a single version, archive permissions in order, a way to measure the answers.
- Search "by meaning" misses exact codes: error codes, alloy designations and certificate numbers must also be searched for exactly as written.
- Tables, drawings and scans are not read automatically: they need OCR and additional reading tools, and a check.
- Projects and notebooks share everything with whoever opens them; an archive with different permissions needs an index that respects them.
- It can make things up even with the documents in front of it: measure it with real questions and expected answers before the pilot.
Note. Informational content updated as of 12 October 2026; the providers' pages were reread on 17 November 2026 and, like the product names, change often: before deciding, reread the sources at the bottom of the page and, for personal data and employment relationships, consult your advisers. The national rules described are those of Italy; other EU member states have their own.
Can I ask my documents questions?
Yes. An assistant on company documents is an artificial intelligence (AI) system that, for each question, searches the company’s manuals, procedures, quotes and certificates for the relevant passages and answers by citing where the information comes from. It is not “chat with a PDF”, one file at a time, and it is not a model retrained on the documents: it is a technique, called RAG (retrieval-augmented generation, that is, generation supported by information retrieval), which leaves the model as it is and puts the right documents in front of it at the right time.
In Italy, the AI technology most widely used by the companies that use AI is already the one that reads documents. Among Italian companies that use artificial intelligence, the most common technology is “the extraction of knowledge and information from text documents”, at 70.8% (ISTAT, Italy’s national statistics institute, December 2025). Across all Italian companies with at least ten employees, analysis of written language stands at 11.6%, in line with the European average; in manufacturing Italy is above the average, in basic metals and fabricated metal products below it (Eurostat, 2025 data). The same ISTAT report says that for 45.2% of the companies that considered AI without adopting it, the obstacle is missing or poor-quality data: for an assistant on documents, the data are the documents.
There are three conditions for it to work, and they are the thread of this guide: readable documents, with a single version in force and someone who keeps them up to date; archive permissions in order, because the assistant shows each person what they can already see; a measure of quality based on real questions, before you extend it to everyone.
How it works, without jargon
Five steps, which all the major providers describe in the same way.
- The index. The authorised documents are read and split into chunks of a few hundred words, where possible following headings and paragraphs: Microsoft suggests starting with chunks of about 512 tokens with a 25% overlap, and Amazon uses chunks of about 300 tokens by default. Each chunk keeps track of which document and which page it comes from. The set of authorised documents is the assistant’s knowledge base.
- The embedding. Each chunk is turned into a series of numbers that represents its meaning (an embedding) and saved in a store built for searching “by similarity”.
- The search. The question is transformed in the same way and compared with the chunks: the ones closest in meaning come out, even if they have no words in common with it. But search by meaning alone misses exact codes: Anthropic gives the example of an error code, “TS-999”, which an embedding model may fail to find because it “understands” error codes in general. In a manufacturing company this is the normal case: item codes, drawing numbers, alloy designations, certificate numbers. This is why serious systems pair semantic search with a search for exact words: the combination is called hybrid search.
- The answer. The chunks found are put in front of the model together with the question and an instruction along the lines of “answer only with the retrieved documents, cite every statement, and if the documents are not enough, say so”. Source citation, with document and page, is part of the technique, not an add-on.
- The update. Documents change and the index does not update itself: after every change it has to be resynchronised, incrementally.
One case where RAG is not needed: when the whole archive fits in the model’s context window. Anthropic, in a technical article from 2024, put the threshold at about 200,000 tokens, that is, about 500 pages; today some models have larger windows, but the criterion stays the same: the manual for a press or the dossier for a tender can be read in full; the archive of ten years of quotes cannot.
The everyday example: an operator asks, “What does alarm E-47 on the press mean, and how do I reset it?” The assistant finds the code, exactly as written, in the manual, reads the paragraph, answers with the procedure and gives the page. If the procedure refers to a figure, it says that it has not read the figure.
What the company needs: the documents
In a manufacturing SME the documents that matter are almost always the same: operating and maintenance manuals for the machines; quality system procedures and instructions, with the related forms; quotes and contracts from recent years; customers’ specifications and drawings; inspection, material and calibration certificates; safety data sheets; price lists and terms of supply; minutes; tables from the ERP, the company’s management software. For an assistant to use them well, they must be:
- real text, not images: scanned PDFs must first go through OCR, and how well it copes with stamps, forms filled in by hand and photocopies has to be tested on the company’s documents, not just claimed;
- in a single version, with a date: two versions of the same procedure give two different answers, and the assistant does not know which one applies;
- with headings and paragraphs: splitting along the structure of the document improves the answers, as Microsoft says in the documentation for its search engine;
- with tables and figures handled separately: the basic reading tool extracts only the text; figures, charts and tables need an additional reading tool, with a cost per page and a check; dense PDFs, with many pages in small type or complex tables, fill the context window before the page limit is reached;
- without encryption or passwords: Microsoft warns that the agent does not read protected files even when the user can open them;
- complete with the documents they refer to: links inside a document are not followed; the procedure that refers to form M-07 must be uploaded together with the form;
- with codes and designations written consistently, because exact search matters as much as search by meaning.
You also need a person who can say which documents are up to date, and a rule for synchronisation. The questions the assistant cannot answer become the list of documents to be written: this is one of the most useful results of the pilot.
Permissions, confidentiality, personal data
All the providers apply the same rule: the assistant shows each person what that person can already see in the source archive, and the index neither widens nor narrows access. These are inherited permissions. The consequence is that the problem is not the assistant but internal sharing that is too broad: a folder opened to the whole company for convenience, containing purchase price lists or a customer’s drawings, becomes visible to everyone in the answers too. Microsoft devotes an entire guide to putting this sharing in order before switching Copilot on. ChatGPT and Claude projects and Gemini notebooks work the other way round: anyone with access to the project sees all of its knowledge. They are fine for a homogeneous dossier, not for the archive.
- Trade secrets. In Italy, the Industrial Property Code (Codice della proprietà industriale) protects secret information with economic value only if it is subject to “measures to be considered reasonably adequate to keep it secret” (Article 98). The assistant’s permissions are also one of these measures.
- Personal data. Internal documents contain the names, signatures and notes of employees, customers and suppliers, and the index copies them. The GDPR’s data minimisation principle applies: CVs in personnel folders, disciplinary records and workplace accident files stay out of the index, or go into an index with tight permissions; if the assistant belongs to an external provider you need a data processing agreement, and you have to decide where the documents reside. If you systematically index the data of many people, the data protection impact assessment is a question to put to your privacy adviser.
- Employees’ email is a case apart. Indexing mailboxes in an assistant is not “one more document”: it is a tool that can lead to remote monitoring of employees, which in Italy brings in the rules of Article 4 of the Italian Workers’ Statute, and it processes third parties’ data. The Garante (Italy’s data protection authority) has a guidance document of June 2024 specifically on cloud email programs that collect employees’ metadata. This is decided with your employment adviser and your privacy adviser, not with the IT person.
- The information notice to workers referred to in Law No. 132/2025, Italy’s AI law (Article 1-bis of Legislative Decree No. 152/1997), concerns fully automated decision-making or monitoring systems: an assistant that answers on manuals is not normally one; it would become one if used to evaluate or monitor people.
The ready-made options
| Option | Where documents and index are kept | Permissions | Declared limits | Requirements |
|---|---|---|---|---|
| Microsoft Copilot (Microsoft 365 Copilot until autumn 2026) and agents in SharePoint (from mid-November 2026 created with Copilot Agent Builder; the old .agent files work until March 2027) | In the Microsoft 365 tenant; for European customers within the EU Data Boundary, Anthropic models excluded | SharePoint’s, per user; the index does not change them | Only content the user can see; encrypted files not read; links not followed; broad sharing to be put in order first | Additional Copilot licence, or pay-as-you-go use for agents |
| Copilot Studio (custom agent, with little code) | Chosen sources: SharePoint sites, lists, connectors | User authentication: only what the user can see | At least read permission needed on the sources; citations cannot be reused by other tools | Copilot Studio |
| ChatGPT Business or Enterprise with the Company Knowledge plugin | In the connected apps (SharePoint, Drive, Slack), queried live: synchronisation into OpenAI’s indexes is no longer offered | Those of the connected apps, per user | Plugin to install, apps to authorise | ChatGPT Business, Enterprise or Edu |
| ChatGPT projects and GPTs with files | Files uploaded to OpenAI’s space | Whoever can see the project sees the files | 40 files per project on business plans, 20 per GPT (custom GPTs are being retired on 11 December 2026), 512 MB and 2 million tokens per file | A plan that includes projects |
| Claude projects, with automatic RAG | Files uploaded to Anthropic, storage in the United States | Sharing per project, Team and Enterprise only | When the knowledge exceeds the window, Claude switches to RAG, “up to 10 times” the capacity; new version of projects in testing | Paid plans; sharing on Team and Enterprise only |
| Claude with an MCP connector to the archive | In the company’s archive, queried by Anthropic’s servers | In the connector, with each user’s own credential | The connector has to be built; a shared credential gives everyone the same access | Team or Enterprise to add it to the organisation |
| Gemini Notebook (NotebookLM until July 2026) | Sources uploaded to Google, not used for training | Per notebook, with explicit sharing | 500,000 words or 200 MB per source, 50 sources for free users, text only from web pages | Google account; higher limits with an eligible Workspace plan |
| Gemini Notebook Enterprise | In the company’s Google Cloud project, with data residency | Per notebook, viewer or editor | 300 sources per notebook, 500,000 words or 500 MB per source, with administration in the Cloud project | Google Cloud |
Two notes on reading the table. Names change: Microsoft 365 Copilot has been called Microsoft Copilot since autumn 2026, NotebookLM has become Gemini Notebook, and ChatGPT’s connectors are called connected apps. And the providers’ pages change every few weeks: the table is valid as of the date in the note at the top of this guide. For a comparison of the assistants in general, covering training, retention and contracts, see the guide ChatGPT, Copilot and Claude: using AI at work.
The custom options
When the archive has different permissions for each department, sources that change and an ERP to connect, you build an index of your own, inside the company’s cloud or on its servers, and the model answers through the APIs under a business contract.
- Azure AI Search with Microsoft Foundry: permissions at the level of the individual document, from indexing through to the answer, with four mechanisms including Microsoft Entra groups and sensitivity labels; citations that lead back to the source documents; OCR and reading of tables and images through additional services.
- Amazon Bedrock Knowledge Bases: ready-made RAG, with ingestion, chunks, embeddings and index; answers with citations taken directly from the sources; files up to 50 MB; chunks of about 300 tokens, or by structure or by meaning; a separate advanced reading tool for figures and tables; incremental synchronisation; the data in the chosen region.
- Google Agent Search and Gemini Enterprise: access control per data source, tied to the company’s identity provider; sentence-by-sentence grounding check; the product names are changing.
- Open-weight models in-house: there are open-weight models and libraries with which to build RAG on your own servers, with the limits of hardware, maintenance and model capability described in the guide AI in the cloud or on-premise: how to choose. The mixed choice, with the index and the documents in the company and the model in the cloud under contract, is often the right one: only the passages retrieved for each question reach the provider’s servers, not the archive, and the contract says how they are handled.
How you pay, without figures: pay-as-you-go for the tokens read and generated, pay-as-you-go for OCR and for the tools that read figures, per seat for the licences of the ready-made products; plus the work of preparing the documents, which is the part no provider does in the company’s place. What the cost depends on is covered in the guide What an AI project costs: what it depends on.
How to measure whether it answers well
Before the pilot you write thirty or fifty real questions, the ones people actually ask, each with the expected answer and the document that contains it. This is the “ground truth” against which the assistant is measured, with four criteria that Microsoft uses in its evaluation tools: the answer is grounded in the documents, it is relevant to the question, it is complete compared with the expected answer, and the documents retrieved are the right ones. Google checks grounding sentence by sentence, with a score from zero to one, and treats a sentence that is only partly supported as not grounded; one wrong date is enough. The evaluations are run again at every change of model or documents: OpenAI calls them an “essential component” of a reliable application.
The rest is the fixed-scope pilot: one department, one group of documents, a few weeks, with the unanswered questions noted down as documents to be written. If the time saved and people’s satisfaction do not show in the pilot, they will not show afterwards either.
Where it goes wrong
- It makes things up, even with the documents in front of it. NIST calls this “confabulation”: wrong content presented with confidence. With RAG the risk is reduced but remains: an answer that mixes two versions of a procedure, a figure read from the wrong table, a “yes” to a question the documents do not cover. The instruction “answer only with the documents and cite them” and measurement are the defences (see hallucination in the glossary).
- It reads old documents. The index does not update itself.
- It does not see tables, figures and drawings without an additional reading tool, and citations from images are not yet offered (Anthropic states this).
- It does not read scans without OCR, and OCR makes mistakes on stamps and photocopies.
- It skips protected files and links inside documents.
- It searches “by meaning” for a code that should have been searched for exactly as written.
- It loses the context of the chunks: “tolerance ±0.05 mm” in a chunk that does not say which drawing and which revision it belongs to.
- It can be deceived. OWASP lists, among the ten risks for applications built on language models, instructions hidden in content and the poisoning of indexed data: an uploaded document can contain instructions for the assistant, and the index is an archive to be protected like the original. Permissions, the log and review of the sources serve this purpose too.
When it is not worth it (yet)
- When the archive is untidy, with no versions and nobody in charge of it: the assistant would hand the disorder back with more confidence.
- When the people who would ask questions are few and the documents fit in five hundred pages: the context window of a business AI assistant is enough.
- When the question calls for calculations or actions on the ERP: that is an agent, described in the guide AI agents: what they are and what they can do.
- When the documents are full of personal data and there is not yet a rule on who can see what.
- When the idea is to use it to evaluate or monitor people: it changes category and becomes a high-risk AI system under the AI Act (Annex III, point 4), with obligations from 2 December 2027 and, in Italy, with decisions on the employment relationship reserved for a person since 23 October 2026 (Legislative Decree No. 179/2026). For an assistant that finds information, what remains are the measures for staff AI literacy (Article 4) and, if it talks to customers, the obligation to make it clear that they are talking to a machine (Article 50).
How we work
We build assistants on company know-how by connecting the company’s internal sources and, when needed, FT Mercati’s market data: prices, indexes and exchange rates in the same answer, alongside a clause or a price list. The status of our systems is declared: the CRM with agents that read orders in PDF is in production; FT CEO is a prototype. The path is that of our method: a defined scope of sources, real questions with the expected answer, a pilot lasting a few weeks, and the plan that stays with the company.
Sources
- AWS, What is Retrieval-Augmented Generation?
- Anthropic, Introducing Contextual Retrieval (the 200,000-token threshold, chunks, exact search)
- Anthropic, glossary: RAG
- Microsoft Learn, Chunk large documents for vector search solutions in Azure AI Search
- Microsoft Learn, Chunk and vectorize with the Document Layout skill
- OpenAI, platform Retrieval guide (semantic and hybrid search)
- Google Cloud, What is Retrieval-Augmented Generation (RAG)?
- NIST AI 600-1, Generative Artificial Intelligence Profile (July 2024)
- AWS, Sync your data with your Amazon Bedrock knowledge base
- AWS, Parsing options for your data source (figures, charts, tables)
- AWS, Supported document formats and limits for Bedrock knowledge bases
- AWS, Content chunking and parsing for Bedrock knowledge bases (chunks of about 300 tokens)
- AWS, How Amazon Bedrock knowledge bases work
- Microsoft Learn, What is Azure AI Document Intelligence? (OCR)
- Google Cloud, Document AI (OCR)
- Anthropic, PDF support (limits of dense PDFs)
- Anthropic, Citations (citations with page number, text only)
- Microsoft Learn, Use SharePoint as a knowledge source in Copilot Studio (encrypted files, links)
- Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot (permissions, training, EU Data Boundary)
- Microsoft Learn, What is Foundry IQ? (knowledge bases, citations, permissions at query time; it has replaced Azure OpenAI On Your Data, retired on 14 October 2026)
- OWASP GenAI Security Project, 2025 Top 10 Risk & Mitigations for LLMs and Gen AI Apps
- Anthropic, Define success criteria and build evaluations
- Microsoft Learn, Retrieval-Augmented Generation (RAG) evaluators (Microsoft Foundry)
- Google Cloud, Check grounding with RAG (grounding score)
- OpenAI, platform Evaluations guide
- Microsoft Learn, Semantic index for Microsoft Copilot
- Microsoft Learn, Manage access to agents in SharePoint
- Microsoft Learn, Secure and govern Copilot: foundational deployment guidance (internal sharing)
- Microsoft Learn, Get started with Copilot in SharePoint (25 September 2026)
- Microsoft Learn, Knowledge sources summary (Copilot Studio)
- Microsoft Learn, Microsoft 365 Copilot connectors overview
- Microsoft Learn, Microsoft Copilot licensing (requirements, no figures)
- OpenAI Help Center, Company knowledge in ChatGPT
- OpenAI Help Center, Projects in ChatGPT
- OpenAI Help Center, Uploading files and audio to ChatGPT (limits per file and per project)
- OpenAI Help Center, Connected apps in ChatGPT
- OpenAI, platform File search guide (vector stores and citations)
- Anthropic Help Center, What are projects?
- Anthropic Help Center, Retrieval augmented generation (RAG) for projects
- Anthropic Help Center, Get started with custom connectors using remote MCP
- Model Context Protocol, Introduction
- Google Workspace, Gemini Notebook (formerly NotebookLM)
- Google, Gemini Notebook: plans and frequently asked questions (since July 2026 NotebookLM has been called Gemini Notebook)
- Google, Gemini Notebook Help: Add or discover new sources (limits per source)
- Google Cloud, What is Gemini Notebook Enterprise?
- Microsoft Learn, Document-level access control in Azure AI Search
- Microsoft Learn, RAG and generative AI: Azure AI Search
- Google Cloud, Set up data source access control (Agent Search)
- AI Act, Annex III, point 4: employment and workers' management (AI Act Service Desk)
- AI Act, Article 6: classification rules for high-risk AI systems
- AI Act, Article 4: AI literacy
- AI Act, Article 50: transparency obligations
- Italy's Industrial Property Code (Legislative Decree No. 30/2005), Article 98: trade secrets, on Normattiva (in Italian)
- Regulation (EU) 2016/679 (GDPR), text on EUR-Lex
- Italy's Law No. 300 of 20 May 1970 (the Workers' Statute), Article 4, on Normattiva (in Italian)
- Garante per la protezione dei dati personali (Italy's data protection authority), guidance document of 6 June 2024 on email programs and metadata (in Italian)
- Italy's Law No. 132 of 23 September 2025, Article 11, on Normattiva (in Italian)
- Italy's Legislative Decree No. 179 of 7 October 2026, authorities and penalties, Gazzetta Ufficiale No. 234 of 8 October 2026 (in Italian)
- ISTAT (Italy's national statistics institute), Imprese e ICT (Enterprises and ICT), year 2025, report of 15 December 2025 (PDF, in Italian)
- Eurostat, Artificial intelligence by size class of enterprise (isoc_eb_ai), update of 15 June 2026
- Eurostat, Artificial intelligence by economic activity (isoc_eb_ain2): manufacturing and metals
- Eurostat, Statistics Explained: Use of artificial intelligence in enterprises (December 2025)
- Microsoft Learn, Upgrade agents in SharePoint and OneDrive (retirement of .agent files)
- Google Cloud, What is Gemini Notebook Enterprise? (usage limits)
- OpenAI Help Center, Custom GPT retirement and migration FAQ
- OpenAI Help Center, Admin controls, security and compliance for plugins and apps (apps queried live)
Checked on 17 November 2026. Rules and deadlines change: always check the official source before you decide.