You have a SharePoint full of slides. A deck from last year's strategy offsite, reference material from a European industrials engagement two years back. Somewhere in there is the exact framework you need right now — you just cannot find it.
So you open Claude and ask. Claude can write code and explain quantum mechanics, but it has no idea what is in your archive. It does not know your firm's IP exists.
This post explains why that is, what your options are, and when dedicated knowledge search software makes more sense than a direct integration.
Why Claude cannot see your slides by default
Claude is a language model with no persistent connection to your file storage. When you open a new conversation, it starts fresh — no memory of previous sessions, no access to your SharePoint or OneDrive, no visibility into anything your organisation has produced.
Claude's developers built it this way deliberately, with no ambient access to your files by default. Something has to bridge the gap.
Option 1: Upload slides directly in the conversation
The simplest approach: open Claude.ai, drag in a .pptx file, and ask your question.
Claude can read PowerPoint files natively. It will extract text from slides, understand the structure, and answer questions about the content. For a single deck — reviewing a client proposal before a meeting, for instance — this works well.
The ceiling: You can only upload files you already have open. If you do not know which deck contains the framework you are looking for, you cannot upload it. And you certainly cannot upload 800 decks to find the one relevant slide.
Option 2: Use Claude via MCP with your file system
MCP (Model Context Protocol) is an open standard that lets AI assistants connect to external tools and data sources. Anthropic, the company behind Claude, supports it natively.
With an MCP server configured for your file system or OneDrive, Claude can browse folders, open files, and read their contents within a conversation. You describe what you are looking for, and Claude goes and looks.
How to set this up:
Claude Desktop (the desktop application) supports MCP integrations. Google Drive has a native MCP connector available in Claude's settings. For OneDrive and SharePoint, a community MCP server exists, though it requires some technical setup — you will need to configure authentication against your Microsoft tenant and run the server locally or in your infrastructure.
If you have a developer on hand, the setup is manageable. If not, it is a meaningful project.
Where it breaks down: MCP gives Claude the ability to read files — but not to search across them intelligently. If you point Claude at a folder of 300 decks and ask "find the slide where we benchmarked European logistics operators against service level targets," Claude will read files sequentially until it either finds something or runs out of context. It will not retrieve by meaning. It retrieves by browsing.
For small, well-organised archives, this is fine. For anything resembling a real firm knowledge base, the approach does not scale.
Option 3: Use a retrieval layer designed for this
A different architecture: instead of giving Claude access to your file system, you build (or use) a layer that sits between Claude and your archive and does the hard work of understanding what is in it.
This is what retrieval-augmented generation (RAG) systems do. They index your documents — extracting content, building semantic representations, understanding what each slide is about — so that when a query comes in, the system can surface the five most relevant slides from a corpus of five thousand.
That retrieved context then gets handed to Claude (or another language model), which synthesises an answer grounded in your actual materials.
The practical shape of this:
- Your documents are indexed — processed once, stored in a vector database alongside their metadata.
- A query comes in: "What did we use as a reference for procurement transformation in the public sector?"
- The retrieval system finds the relevant slides, not by browsing, but by semantic similarity across the indexed corpus.
- Claude receives those slides as context and formulates a response.
The result: Claude "accesses" your archive in the sense that its answers are grounded in it. The intelligence of finding the right material sits in the retrieval layer, not in Claude's browsing.
When each option makes sense
| Situation | Recommended approach |
|---|---|
| One deck, one question | Upload directly in Claude.ai |
| Small, structured OneDrive folder, developer available | MCP integration |
| Large archive (100+ decks), multiple users, firm-wide knowledge | Dedicated retrieval + Claude via API |
| Sensitive materials, strict data residency requirements | Dedicated retrieval on your own infrastructure |
The data residency question
If your slides contain client data — and for most professional services firms, they do — you need to know where they go when they are processed.
Uploading to Claude.ai means your content passes through Anthropic's infrastructure. That may be acceptable for non-sensitive materials, but it is unlikely to pass a client confidentiality review for engagement-specific content.
An MCP integration running locally keeps data within your own network, though the files are still read and their content sent to Claude's API for inference.
A self-hosted retrieval layer — indexing run in your own cloud environment, queries resolved there, only the retrieved context sent externally — gives the most control over where client data actually goes.
What Slidesearch does
Slidesearch is a knowledge retrieval product built for professional services firms. It connects to SharePoint and OneDrive, indexes your PowerPoint archive, and surfaces relevant slides — with source, date, and trust status — in response to natural language queries.
Slidesearch is the retrieval layer that makes Claude, or any other assistant, useful against your firm's actual knowledge base. The two work together: Slidesearch finds what is relevant, the language model synthesises and responds.
If you are dealing with an archive of any real size, and the materials contain anything you would not want processed by a third-party consumer product, that architecture is almost certainly what you need.
Slidesearch is built for professional services firms that want AI grounded in their own knowledge. Get in touch if you want to see how it handles your specific setup.