Company document search: why nobody can find anything and how to fix it

A salesperson needs the current bulk price list, a technician the calibration procedure, a new colleague the travel expenses policy. They type two words into SharePoint search and get forty results, or none. After five minutes they give up and message a colleague on Teams. Company document search barely works anywhere — and it is not because people cannot search.

There are four reasons, and they repeat: classic search looks for words, not meaning; the same document exists in five versions; documents are spread across three or four storage locations; and a large part of the know-how is not written down at all. The fix depends on which of these hurts you most. Sometimes it is enough to tidy up the structure and configure the search you already have. Sometimes Microsoft Copilot is enough. And when sources are scattered and answers need to be verifiable, it makes sense to have your own AI assistant over your documents that shows the source for every answer.

Below is how to tell which one you need — and when it is not worth investing in at all.

Why company document search fails

It looks for words, not meaning. An employee types “how much holiday does a new starter get”; the policy talks about “annual leave entitlement” and “accrual in the first year of employment”. Not a single word in common. In languages with heavy inflection, such as Slovak or Czech, it gets worse still — not every search engine understands that different endings of the same word mean the same thing.

Versions. PriceList_2025_final.xlsx, PriceList_2025_final_v2.xlsx, PriceList_2025_NEW_valid.xlsx. Search finds all three and does not tell you which one applies. Anyone who has once sent a customer an old price list asks a colleague from then on.

Scattered storage. Policies in SharePoint, project documentation in Teams channels, contracts on a network drive, quotes in Google Drive, important agreements buried in email. Each location has its own search, and nobody knows where to start.

Knowledge in people’s heads. “Jane in purchasing knows that.” It works while Jane is at her desk. When she is on holiday or leaves, the company finds out how much was never written anywhere.

On top of that come access permissions: one person cannot see a document they need, while another can see payroll because a folder was shared with “everyone in the organisation” years ago.

What it costs the company

Nobody measures it precisely, but a model estimate is simple. If 20 people lose 15 minutes a day searching or waiting for a colleague’s answer, that is around 100 hours a month. Add the time of the people everyone asks — usually the most experienced and most expensive people in the company. And onboarding, where a new hire spends the first weeks not knowing where anything is.

Four ways to find information in company documents

1. Tidy up the structure and assign owners

The cheapest option, and often underrated. Each area (price lists, policies, technical documentation) has one owner and one place where the valid version lives. Old versions go to an archive, not the folder next door. You agree on file naming.

When it is enough: a smaller company, hundreds of documents rather than tens of thousands, and the main problem is mess. Where it breaks: order without an owner lasts a few months, and it does not write down what is in people’s heads.

2. Make better use of the search you already have

Microsoft Search in SharePoint and Teams, and search in Google Drive, can do more than most people use: filters by type and author, metadata (department, validity, document type), pinned recommended results for the most common queries. A well-configured internal document search solves a lot for a fraction of the cost of an AI project.

When it is enough: people look for a specific document (“the contract with supplier X”), not an answer to a question. Where it breaks: when someone asks “can I invoice this job in euros with an advance payment?” — the answer sits across three documents and none of them is called that.

3. Microsoft 365 Copilot or Gemini in Google Workspace

An assistant that answers questions from your documents, emails and meetings and points to the source. It respects the permissions you already have in Microsoft 365 or Google Workspace. You pay a licence per user per month, so with dozens of people it is not a trivial line item.

When it is enough: everything important lives in one ecosystem, documents are in reasonable shape and permissions are set up properly. For many companies this is the right choice. Where it breaks: sources outside Microsoft or Google (network drives, the ERP, Confluence, databases), or when you need a hard rule of “no source, no answer”, want to measure quality on your own questions, or need data to stay in the EU or entirely on your premises.

4. Your own AI assistant over company documents

This is where we get to what the industry calls RAG — the system first retrieves the passages in your documents that relate to the question, and the model answers only from those, which is why it can show a source for every sentence.

In practice: an employee asks a question in their own words in Teams, the assistant searches SharePoint, Google Drive, the network drive and other sources at once, answers, and attaches the document, page and excerpt. When there is no source, it says “I don’t know”. Someone without access to a document does not get an answer built from it either. What you end up with is a company knowledge base that nobody had to write from scratch — it stands on the documents you already have.

We describe the whole approach, what the system handles and what it costs on the page company document search.

What is enough, when

These routes are not mutually exclusive. Even before building your own assistant it pays to tidy up — and the first phase of any such project is a document inventory anyway.

Risks to watch for with AI search in SharePoint

Making things up. A language model that cannot find the answer in the material will often write one anyway — convincingly, with numbers. The defence is a citation on every answer, a “no source, no answer” rule and a test set that includes questions with no answer. We cover this in detail in RAG without illusions: why your company chatbot lies.

Access permissions. AI will also find whatever was once accidentally shared with the whole company. With Copilot and with a custom solution alike, you need to review who actually has access to what before launch. In a custom solution, permissions come from Microsoft Entra ID or Google Workspace and filtering happens at retrieval time, not when the answer is displayed.

Bad content stays bad content. If two valid policies contradict each other, AI will not fix that. A good system at least flags the conflict and prefers the newer document — and the content owner can see where material is missing.

A tool nobody uses. A separate app with another login never takes off. The assistant has to live where people already work — in Teams, Slack or on the intranet.

When it is not worth it

• A company of up to 10–15 people with documents in one place. A few hours of tidying and an agreement on file naming will do more than any AI project.
• There are few questions. If people ask a few times a week and get an answer within a minute, there is nothing to save.
• The information is not in documents but in a system. Order status, stock, open invoices — that is a job for an ERP integration, not for document search.
• Documents are out of date and nobody wants to own them. An AI assistant over invalid policies will answer wrongly, quickly and convincingly. Content first, technology second.
• Everything is in Microsoft 365 and Copilot is enough for you. Then there is nothing to build — and we will tell you so ourselves.

What it costs

Indicatively, excluding VAT, for a small or medium company: source inventory and solution design €1,200–2,400, a pilot over one document source in Teams or Slack €4,000–8,000, production rollout with access permissions and audit from €9,000, operations and regular updates from €450 a month. We cover the technical side under the enterprise AI knowledge base service.

For a company with scattered documentation we plan for 2–4 hours saved per week per person who currently searches or asks colleagues. The biggest effect is usually in onboarding: a new hire gets an answer with a source straight away, not after a week.

Summary

• Company document search fails because it looks for words, not meaning, and runs into versions, scattered storage and unwritten knowledge.
• Mess in one location: tidy up and assign owners.
• People look for specific files: configure the search you already have properly.
• Everything in Microsoft 365, permissions in order: Copilot.
• Several sources, questions rather than files, citations and audit: your own AI assistant over documents.
• In every case: check content and access permissions first, AI second.

Do you know what your people most often ask their colleagues? That is a good place to start. We are happy to go through it with you — whether tidying up, Copilot or your own assistant makes sense — and if it does not pay off, we will tell you.