Plugging AI into your shared drive finds stale answers faster

This is a slightly contrarian post, so here is the claim up front. Connecting your AI assistant to your shared drive, your email and your chat history is useful. It is also, on its own, a fast way to get confident answers from out-of-date documents.

Why connecting everything is so appealing

The pitch is obvious and partly right. Your AI does not know your company. Your company's knowledge is in documents. So connect the AI to the documents.

The vendors have moved quickly here. In late 2025 OpenAI launched company knowledge in ChatGPT, connecting it to tools such as Slack, SharePoint, Google Drive and GitHub so it can answer with citations from your own files. Microsoft has been building the same thing into Copilot across Microsoft 365. Enterprise search products have done it for years.

And the problem they are aimed at is real. Atlassian's 2025 developer experience report named finding information as the biggest time-waster developers face. If that is true for software teams with good tooling, it is truer for an operations team searching a shared drive.

What the connector actually finds

Think about what is in a typical company drive. The 2023 price list and the 2025 price list, both called "Price List". Three versions of the standard contract, one with a clause legal asked you to remove. A proposal to a client that offered terms you never offered anyone again. Meeting notes where an idea was discussed and then rejected the following week, with the rejection recorded nowhere.

A search-based assistant treats all of that as knowledge. It will find the document that best matches your question and quote it, with a citation, which makes the answer look more trustworthy, not less. The citation tells you where the answer came from. It does not tell you whether that source is still true.

The connector cannot know that the 2023 price list was replaced, unless someone deleted it or marked it. It cannot know that the idea in the meeting notes was rejected, unless someone wrote that down. It cannot know which of three contracts is current. The information that tells you which document to trust usually lives in people's heads, not in the documents.

What a memory adds that search does not

Retrieval finds documents. A company memory holds the things documents do not:

• Decisions. "We dropped the annual plan in April." Search finds the old annual plan brochure. Memory knows it is gone.
• Status. Which version is current, which is retired, which is a draft.
• Dates and owners. When a fact was last confirmed, and who to ask.
• Lessons. "We tried offering free installation in 2024 and it lost money." No document says this, because nobody writes a document about a failure unless asked to.
• Rules. What the AI must never say, never promise, never send without approval.

In practice the two work together. At Provecta, every AI working session reads the company's written context, standards and recent decisions before it does anything, and only then goes looking in documents for detail. The memory tells it what is current and what has been decided. The documents fill in specifics.

Questions to ask before you buy a connector

If someone is selling you an AI tool that "knows your company", a few questions separate search from memory quickly.

Can I mark a document as superseded, so the AI stops quoting it? If not, every old version stays live.

Does each answer show the date of the source, and does the tool prefer newer sources when two disagree? If not, the 2023 price list has the same weight as the current one.

Where do decisions go that are not in any document? If the answer is "put them in a document", ask who writes that document and who keeps it current, because that is the actual work.

Can I see everything the tool treats as company knowledge, in one place, and edit it? If you cannot read it, you cannot correct it.

None of these questions rule a tool out. They tell you how much curation you will still be doing yourself.

What memory will not fix

A memory is curated by people, which means it is smaller than your drive and always slightly behind it. It will not hold every detail, and it should not try to. If a fact matters enough to quote to a client, it belongs in the memory. If it is a detail inside one project, search is the right tool.

A memory also goes stale if nobody owns it. The difference is that staleness in a memory is visible, because every line carries a date. Staleness in a drive is invisible, because every file looks equally official.

Before you connect anything

Pick one folder your team searches often, probably sales or operations. Go through it for an hour with one question per file: is this current, superseded or a draft?

Move superseded files into an archive folder, or rename them with "SUPERSEDED" at the front. Write a one-page note at the top of the folder listing the current version of each key document. Then, if you connect an AI tool, point it at that note first.

Check how ready your firm is to put AI on top of its documents at pgco.world/tools/ai-readiness. Book 20 minutes at pgco.world/book.