We were SharePoint consultants. That sounds narrow, but it wasn’t — to do the job properly you had to speak PowerShell, Azure, M365, Microsoft Graph, HTTP, frontend, backend, DevOps. You were a specialist in SharePoint, but you couldn’t survive without being a generalist underneath it.
A client came to us with a broken webpart. Built by another firm, SPFx, pulling org data via the Microsoft Graph. The list of people it needed to display had grown far beyond what anyone planned for. Hundreds of API calls, sequential, one after another. The page took forever to load. The previous team had solved it like SharePoint developers — they’d fetched what they needed, the way they knew how.
We solved it like people who’d also spent time thinking about APIs.
Microsoft Graph supports batch requests — up to 20 calls in a single round trip, resolved in parallel. We collapsed 200-300 sequential calls down to around 10-15 batched ones. The page loaded. Problem closed.
The fix wasn’t in the SharePoint documentation. It was in knowing that HTTP has patterns, that APIs have conventions, and that the solution to a performance problem isn’t always on the same shelf as the problem itself.
That’s the generalist move. Not knowing everything. Knowing enough about enough things that you can see across the room.
AI is eating specialization.
Not just the repetitive stuff — the deep expertise too. Legal research, medical imaging, code review, contract drafting. AI agents are extraordinarily good at executing within a defined problem space. Fast, cheap, increasingly accurate. The specialist who spent a decade mastering a narrow domain is now competing with a system that can be fine-tuned to match them in weeks.
This makes most people nervous. It should. But the conversation is aimed at the wrong target.
The question isn’t whether AI can out-specialist the specialist. It already can, in more domains than we’re comfortable admitting. The real question is: where does the problem space come from?
Every agent needs a brief. Someone decided what to solve. Someone framed the question, defined the boundaries, chose which corner of reality to point the model at. That decision doesn’t come from inside the problem — it comes from someone standing outside it, with enough context across enough domains to see the shape of what’s actually going on.
That’s the generalist’s job. And it’s the one AI is worst at.
Specialists fail gracefully within their domain and badly outside it.
A senior SharePoint developer looks at a slow webpart and thinks: caching, query optimization, pagination. All valid. All within scope. The fix we found wasn’t outside their competence — it was outside their frame. They didn’t reach for HTTP patterns because they weren’t thinking in HTTP. Why would they be? They were hired to think in SharePoint.
The generalist’s value isn’t a longer list of skills. It’s a wider set of frames. When you’ve spent time in enough different rooms, you start to notice when the furniture in this room looks like a problem you’ve seen somewhere else entirely.
This is what makes the generalist hard to replace. It’s not breadth of knowledge — AI has more breadth than any human ever will. It’s the ability to make a connection that nobody briefed you to make. To walk into a room mid-conversation and say: “Wait. Is this actually a SharePoint problem? Or is this an API design problem wearing SharePoint clothes?”
AI follows the chain it’s given. It doesn’t stop and ask whether the chain leads anywhere worth going.
The generalist’s move has a name: abductive reasoning.
Not deductive — “given these facts, this must follow.” Not inductive — “I’ve seen this pattern a hundred times.” Abductive — “I don’t have the facts yet, but the shape of this problem smells like something deeper.”
Sherlock Holmes doesn’t wait for a confession. He reads the mud on your boots and the callus on your hand and makes the leap. He’s wrong sometimes. But he gets to the right question faster than anyone who insists on waiting for evidence.
In architecture work, this is what the early gates are for. Before the problem is well-defined enough for specialists to execute, someone has to sit in the room and ask: what are we actually solving? That question looks simple. It isn’t. It requires standing in enough domains simultaneously to notice when the stated problem and the real problem are different things.
AI can help you solve the problem you gave it. It cannot tell you that you gave it the wrong problem.
There’s a concept in knowledge engineering called an ontology — a formal map of how things relate to each other across a domain. Not the facts themselves, but the schema. The structure that says: these nodes exist, these edges connect them, this is how meaning is organised.
A specialist’s knowledge is a deep vertical slice of that graph. An AI agent is extraordinarily good at traversing it — fast, exhaustive, accurate within the schema it’s given.
What neither the specialist nor the agent does well is extend the ontology. Notice that two nodes should be connected but the edge doesn’t exist yet. That’s the move that happens before the problem is solvable — someone has to add the relationship to the map before anyone else can navigate it.
That’s the generalist’s job. Not traversing the graph. Drawing the edges.
(That idea deserves its own piece — but that’s a rabbit hole for another day.)
The implication is uncomfortable if you’re mid-career in a specialty.
The hollowing out won’t hit the top of the expertise curve first — the true masters, the people who’ve spent decades at the frontier of a domain, still have a while. It’ll hit the middle. The competent specialist who knows enough to execute but not enough to lead. That’s where AI is most directly competitive.
The generalist, paradoxically, becomes rarer and more valuable. Not because they know more — they don’t. But because the skill they have is precisely the one that’s hardest to automate: standing at the intersection of multiple domains and making the call about which direction to walk.
I’m currently renovating the exterior of my house.
On weekends, on a scaffold I assembled myself, with a nail gun I wasn’t trained to use. I’m not a carpenter. I will never be a carpenter. There are people on my street who could do this work faster, cleaner, and with fewer trips to YouTube.
But I understand enough — the principles, the sequence, the logic — to do most of it myself, and more importantly, to know exactly when to call someone who knows more than I do.
That’s been true my whole career. There have always been better coders, better architects, better specialists in every room I’ve been in. What I’ve been able to do is move between the rooms — carry something from one conversation into another, make a connection nobody asked me to make, and occasionally walk into a slow-loading webpart and think: this isn’t a SharePoint problem.
In a world where AI can staff every room with a specialist, that might be the most useful thing a human can still do.
Know enough. See across. Ask the question nobody briefed you to ask. Be The Last Generalist.