From Toys to Transformational Tech: A Skeptic's guide AI in Agtech

In a recent talk for Croplife Asia , I opened up with a bad dad joke pretending to be confused about AI, thinking that I had been asked to speak about Artificial Insemination and livestock genetics.

As cringeworthy as that sounds, and noting that I’ve been in plenty of agriculture rooms where AI can mean ‘Artificial Insemination’ or ’Active Ingredient’, the other AI - Artificial Intelligence- is a field that’s creating confusion for many people.

My view, having experienced multiple waves of both transformational and over-hyped technologies in and outside of agriculture, is that there is no better way to assess the possibilities until you start engaging.?

For me, that means experimenting with, thinking-out-loud about , and paying attention to the latest in AI - including by unpacking lessons learned by AI practitioners in large and small agriculture organizations. Our recent podcast conversation with guest Feroz Sheikh, CIO at Syngenta Group , got me thinking about where and why I have skepticism vs. optimism about AI in agriculture, and the approaches we can take as investors and operators to navigate this fast-moving space.??

The hard thing about hard things

One big reason I continue to hold skepticism about AI is that building good AI-driven solutions is hard. It is hard for many reasons, not least of which is that we're still very early in the journey and things are moving very quickly.

The current set of AI-tools allows people to make rapid progress. I was able to learn the basics of prompt engineering in a weekend. And I’ve seen companies make it a long way in building systems just by crafting well designed prompts.

It is far more likely to be the case though, that building complex, sophisticated, and safe systems, especially ones that are specific to an industry like agriculture, will take far more than just prompt engineering .

Beyond just the technology, I also have questions about the business models for AI. Even outside of ag, building AI systems is very expensive, and there just aren't many companies at the moment that are making money from AI. In fact, many are losing money in a hurry

As noted by Feroz Sheikh, AI, in the end, is a data-oriented business model. There really aren't any agtech companies where data analytics is the core offering that are wildly profitable.

A mile wide and a mile deep

When you are seeking expert advice, in agriculture or otherwise, humans tend to be a mile wide and an inch deep or an inch wide and a mile deep. The exciting part of AI is that no such constraints apply to an LLM. What if we could remove some fundamental constraints on key activities in agriculture?

Feroz Sheikh gave a great example with Syngenta’s release of their CropWise solution. Rather than attempt to replace agronomists and provide AI directly to farmers, it increases the amount of acres or growers that an agronomist can advise.

“So, where in the past if there was one advisor who was advising 10 growers, the same advisor will now be able to advise 50 growers or 100 growers.”

What if we no longer have to forgo depth to get breadth? We could soon be able to move away from farming by averages, and farm with unparalleled spatial, temporal,? species level precision available to the widest audience.

Running important races at twice the pace

I’m also excited about how AI could help accelerate the development of solutions for climate adaptation and resilience. As a collaborator of ours, Rhishi Pethe, recently pointed out , AI has already been shown to significantly reduce the time required for the discovery and development of new herbicidal compounds, sometimes by up to 50%. For example, Enko Chem claims AI helped them identify a novel herbicide candidate in 18 months compared to the average 11 years.

Feroz also highlighted that Syngenta is applying AI to this important race. They have harnessed their vast stores of data to build models that will accelerate the discovery of new compounds and active ingredients.

Users of, builders on, and builders of

As investors evaluating an increasing number of solutions touting AI, we’ve been asking ourselves who is likely to have the capabilities to build productive, efficient, sustainable and equitable solutions.?

Another collaborator of ours, Kendra Vant , who spent time working in AI solutions for accounting software company Xero, recently proposed a very helpful framework on this. Kendra suggests that organizations need to know if they are builders of AI technology (very unlikely), builders on AI technology (perhaps), or users of AI technology (most likely).

Feroz related a relevant example of this distinction. Initially, Syngenta’s working hypothesis was that a proprietary foundation model, built specifically for a domain, would out perform a generalized foundation model that was fine-tuned for that domain.?

In a rare moment of corporate candor, Feroz pointed out that their experiments showed just the opposite. After appropriate training, the widely available generalized foundation model outperformed their custom model. This is a key insight, and lends weight to the idea that very few organizations will be successful builders of AI fundamentals.?

Toys to tools to transformational tech?

There is no doubt that access to data is a major factor that will determine who benefits from AI, and who can build sustainable advantage in AI-based systems. The podcast episode highlighted how Syngenta approached this, and the degree to which it has focused on both its current data repositories, but also in upskilling the entire organization to be aware of how important AI-aligned data collection practices will be to the future of the company.

But making changes like this will be about far more than the technology itself. Building AI capability within an organization must be done through engagement and education. Feroz shared Syngenta’s catchy framework for this: Toys, Tools, and Transformational Technologies.

This is a powerful approach because it pays attention to the fundamentals of human psychology - people need incentives to change. Starting with toys, fun and simple ways to experiment with AI, provides a low-risk way to build awareness. Moving on to tools, organizations can find ways to add efficiencies into the daily patterns of people across the organization, helping them see further into the future. These early interventions build awareness and buy time for the organization to experiment with and develop the transformational technologies that will pay real dividends, and will vary greatly from one company to the next.

The toys to tools journey is one I’ve been on. As a small venture firm, personal productivity and efficiency is something we’re continuously chasing. Readers and podcast listeners will notice that we use AI generated images to accompany articles. I have played around with many toys in this area. More recently we have had projects and larger talks that needed a longer series of images to support them, and it became a chore to do each image one-off. Moreover, keeping the same style across images of varying subjects, and efficiently generating 20 or 50 images, started to need a tool. I asked ChatGPT for help.?

It suggested I could enter image specifications into a spreadsheet, then run a batch of images using the API of an AI image generator. A few more prompt tweaks, and I now have a Python-based command line tool that reads a spreadsheet and creates consistently styled images, of any set of subjects, with color and camera angle control. A very useful tool, that started as a toy.

You can’t start until you get started

There are a variety of (sometimes extreme ) opinions about how transformative AI will be for most organizations. For us as investors looking at an increasing number of AI-ridden pitch decks, avoiding the hype seems like table stakes. To catalyze systems change, I believe we must also build a deeper understanding of what AI really is and could be, including by experimenting ourselves and ensuring that conversations pay attention to the vital and myriad nuances.?

Syngenta’s approach provides an instructive example of thinking holistically about introducing AI to an organization. Not building an AI strategy, but deeply investigating if there are meaningful ways that an organization can harness AI-enabled solutions that will deliver on core strategic initiatives.

Hopefully, amidst the confusion and hype, we neither fail to heed warnings nor underestimate the promise.