Your AI Agents Don't Have a Data Problem. They Have a Knowledge Problem.
At a manufacturing client, there’s a woman who’s worked the assembly line for twenty-two years. When a machine makes a particular sound, she knows exactly what to adjust before it becomes a real problem. Nobody else knows. It isn’t written anywhere. It’s not in the maintenance manual, it’s not in the ERP, it’s not even in her head in a way she could cleanly explain if you asked her to. She just knows.
That woman is the AI agent you will never be able to build. Not because the technology isn’t ready, but because what she knows was never data in the first place.
Knowledge is not data
When an AI project kicks off in an organization, the conversation turns to data within minutes, almost every time. It’s incomplete. It’s poorly structured. It’s messy, scattered across three systems, with two different date formats and a “notes” field that actually holds all the useful information, typed in free text by someone in a hurry.
All of that is true. And all of it leads to the wrong conclusion: waiting for the data to be clean before you start.
Data is only one piece of what feeds an AI agent. Your company’s knowledge is the whole thing: the data, yes, but also the context around it, the processes that produce it, who has the right to touch it, and who has the right to share it.
An agent with access to perfect data but no context on why it exists will still hallucinate. Just with more confidence. What I see most often with clients isn’t a shortage of data. It’s a shortage of everything else.
The problem is what’s tacit
Most of the knowledge that actually runs a company isn’t written down anywhere. It lives in the head of the woman with twenty-two years on the floor, in a team’s undocumented habits, in the “that’s just how we’ve always done it” that gets passed to every new hire during their first six months.
That’s not bad management. It’s what naturally happens in any organization that actually functions: you document what you have time to document, and keep the rest in your head because it’s faster.
Now, you’re probably thinking this is a small-business problem, that a large enterprise has everything documented, governed, and properly controlled. You’d be wrong. Large organizations do put more structure in place, that part is true. But the more systems, committees, and exceptions to the rules pile up, the more tacit knowledge grows around just navigating all of it. The twenty-year employee who knows exactly which form to route around, which approval is cosmetic and which one is real, who to actually talk to for something to move: none of that is in the governance manual. My own two cents on it: a large, well-structured organization doesn’t have less tacit knowledge than a small business, it just has one extra layer of it, the layer that explains how its own bureaucracy actually works.
The problem shows up the day you connect an AI agent to all of that. The agent doesn’t have access to anyone’s head. It has access to what’s written, and nothing more. If half the critical knowledge is tacit, the agent is working with half the file, and it doesn’t know that. It answers anyway.
Contextualized to the moment it actually matters
There’s a second, subtler trap. Even the knowledge that is written down somewhere isn’t fixed. A vacation policy changes. A supplier gets replaced. A security procedure gets revised after an incident. Knowledge has an expiry date, and the date that matters isn’t when the document was written. It’s the moment the agent uses it to make a decision.
An agent that answers with last year’s policy isn’t wrong because it’s a bad model. It’s wrong because nobody told it the truth had changed.
What already exists to help
Here’s the good news: your data doesn’t need to be perfect for you to move forward, as long as your knowledge covers more than data alone: information (data plus context), documented processes, clear access, defined sharing rights.
In the Microsoft environment, a good part of that toolset already exists, and most organizations are using only a fraction of it. SharePoint quickly becomes unavoidable for your Copilot agents the moment you’re talking about office documents: that’s where a good chunk of a company’s written knowledge already lives, and I’ve written before about its role as grounding for Copilot Studio.
What’s changed recently is that Microsoft put a name on the expiry problem I described above. SharePoint Advanced Management now includes a Content Management Assessment that scores a tenant’s “Copilot readiness”: ownerless sites, inactive content, overly broad permissions, with a re-run recommended every 30 days instead of an annual audit. The “Everyone Except External Users” report alone surfaces the top 100 sites shared organization-wide in the last 28 days, which gives a fairly concrete sense of how much oversharing exists before AI even enters the conversation.
The detail that stopped me while researching this piece: when a site moves into Microsoft 365 Archive, the content, permissions, and metadata are preserved, but Copilot is no longer trained on it. Microsoft literally built a mechanism to tell an agent “that one’s stale, ignore it,” instead of leaving that call to whoever still remembers.
That’s where Purview comes in, and not just as a compliance ledger. Sensitivity labels follow the content all the way into the agent’s answer: an encrypted file is only returned to the user if they hold the EXTRACT right in addition to VIEW, regardless of what the agent “read” to build its response. Purview Data Lifecycle Management retention policies apply to the agents’ own prompts and responses, not just the source documents. And the retrieval API Copilot uses to pull content from SharePoint applies security trimming: it only returns what the person asking already has the right to see, regardless of what the index actually contains.
Together, SharePoint Advanced Management and Purview answer the question I raised earlier: is this information still true today, and does the person asking actually have the right to see it?
It’s not the only answer
I’m not claiming SharePoint and Purview cover a company’s entire knowledge base. That would be dishonest, and it wouldn’t survive five minutes with a client whose ERP and CRM live elsewhere. A good chunk of critical knowledge lives in those external systems, with their own access rules and their own lifecycle.
What I’m saying is that the office productivity foundation, the one holding the policies, the procedures, the email threads that explain the why behind the what, deserves the same governance rigor as your ERP. Today, it rarely gets it.
Next time an AI project stalls on “our data isn’t ready,” ask the question that actually matters instead: is our knowledge ready, and do we still know, at the moment it counts, what’s true?