Another Copilot seat won’t 10x your productivity. Understanding how to link your company’s expertise might.

The current approach to AI is making this worse. At the top, many C-suite leaders — most of whom came up before this technology existed — are assuming AI will let them run their businesses with fewer people. They act on the assumption, and things break: logistics, customer response, institutional knowledge walking out the door. Further down, the same organisations enable Copilot, run prompt engineering workshops, and tell every employee to start using AI. The people who would actually benefit from automation are too busy with the work that’s wearing them out to learn a new tool. The ones who aren’t busy don’t know what to do with it. Nobody sat down and asked: what are the actual problems here, and whose time is being wasted? Instead, the organisation got a chatbot and a marketing statement that it does “all kinds of things.” That’s less than half the picture.

The Insight

What AI actually does has three parts.

First, it removes the coordination and retyping overhead that is currently landing on experts at every level — the email triage, the meeting prep, the bill of lading that has to be retyped from one system into another before it can be sent to a partner. An agent extracts the structured data, runs it through the mechanical pass/fail tests the business already has (plain regex, business rules, no model involved), and surfaces one approval request instead of making a human read forty pages. The expert approves it or sends it back. What the expert was doing — reading the forty pages — wasn’t expertise. It was transcription with stakes.

Second, over time, the agent learns how things actually get done in the organisation. Who delivers on schedule. Which vendors require two reminders. What stakeholders actually care about when they say they care about something. The human nuance that doesn’t exist in SAP or Microsoft or your email or any dashboard. An accountant wants to understand what IT built into the dashboard, what marketing thinks is actually driving the business, and what the stakeholders said at lunch. Today she does that cross-referencing manually, because no system holds all of it. Over time, the agent does — not because the data was structured for it, but because the agent has been watching how expertise gets used and has captured the nuance around it.

Third — and this is what “AI makes us more human” actually means when it isn’t a slogan — the combined effect of removing the coordination overhead and capturing the nuance is that experts at every level get their time back for the conversations that move a business. The phone call where one person says “that’ll be a lot of fun when we do it next week” and the other hears the hesitation in it and understands the project is not going to be as simple as either of them thought. The grunt at the end of a sentence. The laugh that means “I’ll handle it.” Two humans at full bandwidth, communicating information that does not exist in any written form and cannot be reduced to one. AI cannot have that conversation, and so long as humans run organisations and make decisions partly for reasons that aren’t logical — right or wrong — it will not need to. Its job is to make sure the conversation has room to happen, and that the people in it have better information than they otherwise would.

Der Bauernhof

Switzerland is a country of farmers. You’re surrounded by cows, the best cheese in the world, the best produce in the world. Everyone understands the Bauernhof. Everyone knows a Bauer.

A farmer’s real expertise is walking the fields and looking at his land and seeing what needs to be done. That knowledge is built over decades. It’s irreplaceable. But this farmer is also dealing with drones scanning his fields, tractors that generate telemetry, weather information, vendor pricing for seed and fertiliser, regulatory paperwork for Direktzahlungen — all of these data points that he’s capturing in his head while doing the actual physical work of farming. He spends his evenings on a computer instead of resting, because the administration never stops.

You cannot replace thirty years of walking this parcel, touching this soil, getting rained on in the same fields every season. No dataset captures that. I live in Kaiserstuhl, under MeteoSwiss, which is one of the best forecasting services in Europe. For my particular village, the forecast is wrong almost every week, because there is a river here and the model doesn’t know. The model knows the region. The farmer knows the parcel.

What MeteoSwiss sees. What the farmer sees.

Now imagine this farmer has a personal AI agent running on a machine in his farmhouse. It’s been given the same data he has — the drone feeds, the weather, the vendor catalogues, the regulatory calendars. And it’s been seeded with his knowledge: his decision rules, his seasonal patterns, what he cares about, what he ignores. That knowledge was captured over a few weeks of conversations. Someone sat with him and wrote it down — not in a report, but in structured text files the agent can read.

The agent runs 24 hours a day. It processes the drone imagery overnight. It cross-references weather with his planting schedule. It compiles vendor pricing. It pre-fills documentation. The farmer doesn’t interact with a chatbot. The work is just done. And when he wants to ask a question — “what did the drone show on Parzelle 3 last week?” — it’s there. Over time, the agent stores his corrections, his instructions, his patterns. It evolves with him. After a season, it understands this particular farm and this particular farmer in a way no generic software could.

The farmer’s job doesn’t change. He still walks the fields. He still makes every important decision. The difference is that everything around his expertise is handled. He farms.

Die Baustelle

If you’ve walked through Zürich recently, you’ve walked past a Baustelle. The work is deeply human — someone has to pick up the cement, mix it, and lay the cobblestones. Someone has to look at the site and know, from experience, that the ground conditions aren’t what the plans assumed. No technology changes that.

But a Baustelle generates enormous amounts of data. GPS instruments record the exact height of a new sidewalk. They mark where somebody’s property line ends. The rate at which cement is being consumed is tracked. Delivery schedules for materials exist. Traffic and logistics affect the project timeline. The Bauleiter — the site manager — is constantly taking in all of these data points, putting them together in his head, and making decisions. He’s looking at how the current delivery situation or the weather is affecting the overall trajectory of the project, all while managing a crew on site.

Nothing in a dashboard tells the Bauleiter that the driver of the cement truck stops for breakfast at the same Raststätte every morning, so he will arrive forty minutes after his ETA says. Nothing tells him that the stone is coming from a supplier in Ticino, and the suppliers there keep a different schedule than the ones in the north — things are just done differently down there, and the roads through the Gotthard don’t move in summer the way they move in November. The dashboard knows the scheduled arrival. The Bauleiter knows why it is always wrong.

Dashboards for all of this already exist. Every project management platform has them. The problem is that a dashboard is passive — it shows you everything, but it doesn’t know what matters to this particular Bauleiter on this particular morning. What the agent does is make that dashboard active. It stops showing him forty metrics and starts telling him the three things he actually needs to know before the crew arrives — the delivery that’s running late, the grade discrepancy on the east side, the weather window that closes at noon.

The concept of integrating siloed data like this was pioneered by Palantir. Their insight was that companies like Airbus had massive databases that didn’t talk to each other and required people to manually bridge the gap. Palantir consolidated it. What’s changed is that you no longer need Palantir to do this. The same integration capability is becoming available through open-source agent platforms running locally, on your own hardware, without the enterprise price tag. Palantir proved the concept. Now it’s accessible to a Swiss Bauunternehmen.

No new hires. Same five people — connected.

Where This Is Going

Right now, agents work with digital data — spreadsheets, emails, APIs, databases. The expert’s physical perception — what they see, hear, and feel on site — is still inaccessible to the agent. That gap is closing.

Think of a mountain biker with a GoPro. They’re capturing their ride in real time so they can go back and understand how it went, and so others can experience what they experienced. Police officers wear bodycams so there’s an objective record of what actually happened, because human memory is unreliable — events get mixed up in people’s heads, but the camera just records what happened.

Now imagine the farmer walking his fields with a drone overhead. The agent is walking with him. The farmer stops for a second, looks at a patch of crops, and moves on — he noticed something, but he had fourteen other things on his mind and didn’t consciously process it. The agent saw it too. From the air, it saw that this section of the field looks different. And later, at the right time and the right place, it says: “Did you notice that spot you paused at? I saw it from the air. It looks like this area might be lacking nitrogen. You noticed it intuitively, but you might not have thought about it further. I checked — the automated fertiliser delivery may have inadvertently changed the application in this zone three weeks ago. Do you want me to flag a soil sample from this area?”

The farmer didn’t miss anything. His eyes caught it. His brain registered it for half a second. But his attention was full. The agent doesn’t have that limitation. It doesn’t forget, it doesn’t get distracted, and it brings up the right information at the right moment — not in a notification blast, but when the farmer is ready to hear it.

What I Actually Do Right Now

The personal agent infrastructure is being built as we speak. Peter Steinberger’s OpenClaw, and NVIDIA’s enterprise wrapper NemoClaw on top of it, are what Jensen Huang called “the operating system for personal AI” at GTC 2026. Stephen Witt’s biography The Thinking Machine documents the chip-design history that got Huang and NVIDIA to this moment. The technology is early but moving fast. Strip away the branding and the enterprise wrappers and at the core of all of it are just markdown files. Text files on your machine. Rules, memories, decision logic written in plain language. No subscription required. You own them. They’ll work with whatever platform comes next.

Strip away the branding. This is what’s left.

Most businesses already document their expert knowledge — in SOPs, meeting notes, email threads. But it’s documented for other humans to read, not for agents. The formatting is wrong, the implicit knowledge isn’t made explicit, the decision logic is buried in narrative. And the nuance that actually determines how things get done — who’s reliable, which clients pay on time, what stakeholders really care about — doesn’t get written down at all, because the people who have it are too busy triaging email to stop and say it out loud.

That’s what I do. I sit with people at every level of an organisation — not just the senior experts, but the ones everyone defers to for their specific piece of the work. I document what each of them knows, and especially the nuance that doesn’t live in any system, in structured files that an agent can read. I separate the work that requires their judgment from the work that’s routine. I build the automation that handles the routine part today, and the captured knowledge becomes the seed for an agent that eventually goes out and does things behind the scenes when nobody is awake, contextualises what it finds, and gets sent back when it’s wrong. Organisations that start this work now will be ready when the technology is. Those that wait will be starting from scratch while their competitors have agents that already understand how their business works.

• • •

The question for any organisation is not “should we adopt AI?” Most already have. The question is: who are your experts, and what are they doing instead of their actual job?