Everyone in AgriFood Wants a GLP-1 for AI. There Isn't One.

There's no GLP-1 for AI. No shot, no shortcut, no pill that melts the effort away.

"AI is like weight loss," a commercial leader told me a while back. "Everyone knows they need to do it, and everyone's looking for the GLP-1." He's right on both counts. Everyone in AgriFood knows they need to get serious about AI, and almost everyone is quietly hoping for a shot that does the work for them. There isn't one. But there is something that works.

I've spent the last few months talking with people all across AgriFood. Leaders and operators, product managers, agronomists, growers, and the investors backing them. Different seats, same conversation. Here's what I keep hearing:

"I thought this was the next generation of leaders' problem."

"We launched but adoption plateaued."

"How do we differentiate if everyone can access the same models?"

"Our people's knowledge value is decreasing because the customer is using GenAI."

"Our systems don't learn, so we repeat the same mistakes."

None of those are the words of someone who's behind. They're the honest sound of an industry in the middle of a transition it didn't ask for and can't opt out of. Nobody in these conversations is lazy. They're smart people running real businesses, handed a technology that changes every few months and told to figure it out on top of a full-time job.

The shape of it is jagged. The models dazzle on one task and flop on the next. One corner of a company is rich with context, the one beside it has none. And adoption is the most uneven thing of all. Here's the tell: two people at the same company get the same AI subscription, and one is vibe-coding and building agents while the other tries it twice and quits. Same building, same tool.

So what separates them? It starts with a willingness to learn.

I'm self-taught on all of this. I sat down in front of these tools with no more preparation than anyone else, and I decided early that I couldn't afford to be intimidated, so I sat in the discomfort until it wore off. That's most of the trick. When a leader tells me they don't even know what an LLM is, I get it, and I know how fast that feeling passes. You don't have to become an engineer. You have to stop flinching, put in a few focused hours, and pay attention to how the tool actually works.

Do that, and you start to see what the people getting real results already figured out. They set the thing up. They give it memory, tune the settings, connect it to the files and tools that matter, and above all they feed it context. A model with no context is a genius with amnesia, brilliant and fast and starting from zero every time you talk to it. The ones pulling ahead aren't running a better model than you. They're feeding it something you're not.

Context: The part that was never written down

Context capture is the process of turning an expert's unwritten judgment, the reasoning behind their decisions, into a document an AI system can use.

The most valuable input AI needs was never in a database. It's the reasoning your best people carry in their heads and have never had a reason to write down.

How a product manager decides what to say about a product, in which market, against which competitor, and why. How a sales leader positions the same offering across different cropping systems. The logic behind the proposals, the pricing calls, the market reads. None of it lives in a system, because no system ever asked for it.

That judgment is the gold. It's sitting uncaptured inside your people, and capturing it is what turns one person's know-how into something the whole company owns.

Capturing it is real work, too, the interesting kind. Your best people usually can't just hand you their reasoning on a page. They earned it over years and never had to explain it, so it comes out as instinct and half-sentences. Drawing it out and shaping it into something the AI can use takes patience and skill. Done right, it treats their expertise as the asset it is, not as something the tool is here to replace.

I watched this get concrete on a project with a commercial team. The leader asked me a question I hear constantly: how do we help sales tell the ROI story on our products? The logic to answer it lived in one expert's head. The product configurations existed. The reasoning that turned them into a recommendation for a specific farmer did not.

At 9 North Group, we sat with the expert, reverse-engineered how they actually make the call, and captured it in a markdown file.

Then we connected that document to AI and built it into a tool. You feed it a farmer's basics. Operation size, crop mix, application timing, financial situation. It hands back a tailored sales script with a ROI estimate built for that specific farm.

What once took a week, now took about ten minutes. Speed can mean reaching a grower in time to matter that season instead of the next. In our business, a missed season can be a full year.

Here's the part that matters more than the time saved. That document didn't disappear when the project ended. It became a knowledge asset the company owns and can refine.

Another leader put the problem to me plainly: "We're churning people. Knowledge is walking out the door, and new people have to relearn everything." A captured asset is the direct answer. It outlasts the expert who's nearing retirement. It ramps a new hire in weeks instead of seasons. And connected to a feedback loop, it becomes the start of a system that actually learns, instead of a tool frozen at whatever it knew on day one.

That's the whole trick. The model didn't get smarter. It got informed, and the company kept what it learned.

AI models don't improve on their own. They improve when a company takes the time to train it and maintain it with the knowledge from its most precious resource: the team.

Why the good stuff stays stuck

If capturing context is that doable, why isn't every company running on it?

Siloes plague company effectiveness, which is only more true when AI is an independent endeavor. Someone builds something useful, and it stays on their screen. The person down the hall never sees it.

It's because nothing around them was built to catch it. No budget to institutionalize and train the rest of the team on how to use it, to maintain it.

Some of the fixes are almost embarrassingly simple. A shared prompt library, so the prompt one person spent a week getting right is one click away for everyone else. A place to share the skills people build, the repeatable setups that turn a recurring task into a template. And sharing agents, so a workflow one person automated can be handed to the whole team instead of rebuilt from scratch. None of that needs new technology. It needs someone deciding that what one person figured out belongs to everyone.

That's not a people problem. It's a setup that was never designed for this. The good news is the structure is buildable.

What the evidence shows actually drives results

I don't know a single company or individual that has it all figured out. But some are trying, and they're seeing results. Here's what that looks like.

First, they are building systems to capture what their people know. Not IT knowledge bases full of policy. Real judgment, written down in a form the AI can use, so the tool draws on your best people instead of the open internet.

They also fund it like they mean it. Not a line borrowed from a team budget for a one-off pilot, but a dedicated budget with room to run real experiments. Then they work in short sprint loops. Ship something small, measure whether it moved a number, learn, and put more behind whatever's showing value. Less a launch than a set of bets, managed toward the ones that are working.

The third piece is checking output to measure performance and target improvements. A model hands you a wrong answer in the same confident voice it uses for a right one, and a bad call rarely announces itself. It gets absorbed as noise.

An AI evaluation is a test set of real questions and expert-verified answers, used to measure whether an AI tool is giving correct, useful responses before and after changes.

Checking it is more concrete than it sounds. It's called an evaluation, and you don't need a data science team to run one.

Write down 50 to 100 real questions your team actually asks on the targeted use case. Have your expert write the right answer to each. That's your gold-standard set. Run the AI against it and see where it misses.

When it's wrong, you fix it, usually by adding the missing knowledge or narrowing what the tool is allowed to say. Then you keep a panel of experts in the loop, scoring new answers and feeding corrections back. The gold set grows. The experts keep it honest. The tool gets a little better every cycle instead of freezing at launch day.

That's the answer to "our systems don't learn." They don't, on their own. The learning is the loop you build around them.

Last, the ones pulling ahead build so they can swap the model underneath. A better or cheaper one shows up every few months, and cost and performance both depend on moving to it fast. The context you captured and the gold-standard set you built are what let you drop in a new model and re-tune in days instead of starting over. Your edge rides on whatever model is best that quarter instead of being chained to one.

None of this needs you to be technical, and none of it needs you to be early. Better still, it compounds. Every piece of context you capture makes the next model that ships work better for you the day it lands.

Questions we hear

What is context capture?

Context capture is the process of turning an expert's unwritten judgment into a document an AI system can use, so the reasoning behind a decision, not just the data, becomes something the company owns.

Is there a "GLP-1" or shortcut for AI adoption in AgriFood?

No. There is no tool or model that removes the need for effort. The companies seeing results are the ones capturing their people's judgment and feeding it into AI systems, not waiting for an easier tool to arrive.

How do you capture institutional knowledge for AI use?

By sitting with an expert, reverse-engineering how they actually make a decision, and writing that reasoning down in a document (often a simple markdown file) that an AI tool can reference.

What is an AI evaluation and why does a company need one?

An AI evaluation is a test set of 50-100 real questions with expert-verified answers, used to check whether an AI tool's responses are accurate and to catch drift over time. Without one, wrong answers get absorbed as noise instead of corrected.

‍