Tami Craig Schilling, (retired) Vice President & Agronomic Digital Innovation Lead, Bayer
From the Land Grant University in Urbana Champaign, Illinois, this is the Closing Market Report. I'm Extension's Todd Gleason away from the office on this November 2025, a Wednesday afternoon hosting the Nutrient Loss Reduction Strategy Annual Partnership Conference here on campus. Coming up this afternoon we'll hear about digital agriculture and the benchmarking system that is being put together by the Center for Digital Agriculture here on campus and a consortium of corporations and organizations from around the planet. We'll hear from Vikram Adave who is the co director of the center along with Tammy Craig Schilling, former vice president and agronomic digital innovation lead at Bayer. I asked Tammy first to explain how it is that she and Bayer became involved in this University of Illinois project.
Tami Craig Schilling: 00:55About three years ago, and really, it's been longer than that. It's probably been the last nine years. I have been focused on finding ways to improve farmers' experience with, bear products and agronomy information. And so the past three years, what we were doing in the prior time was we were doing more social media with our agronomists in the field. We were, continuously talking with farmers about what kind of agronomy information do you want, doing some data mining to see what topics when, and got to the point where we were building some systems to organize those and build the back end data to deliver that.
Tami Craig Schilling: 01:40We started doing this in '21, 2021. Well, 2022 in the fall, ChatGPT came out and was the first generative AI tool. And we said, instead of us building our own systems, why don't we build on top of ChatGPT and do a product and agronomy tool, and try that first? So by summer, we had Microsoft and Ernst and Young, had agreed to fund their part, and we'd do a ninety day POC to see if we could build a better a better chat tool to provide products and agronomy information than an out of the box chat GBT. By October, we proved we could do 40% more accurate, and we determined we're onto something.
Tami Craig Schilling: 02:24So, currently at Bayer, which I am an emeritus retired from, there is a tool, and it's a generative AI tool that about 5,000 of our employees have access to. And eventually, we hope that farmers and ag retailers also have access eventually. But this tool can give you product label information. It can give you disease information. We've gotten a lot of great content there that we have about 20,000 questions have been asked to this to the tool, and it's called Eli, e l y.
Tami Craig Schilling: 02:57We've gotten a couple really nice global awards and some here in The US. And about, oh, probably almost a year ago, I would say, Vikram and I because my appreciation for AI Farms and Center for Digital Ag, and I'd been engaged with John and the team and Vikram, We also had interest in licensing in CropWizard. And so a team of us got together that were building this tool within Bayer and said,
Todd Gleason: 03:29okay.
Tami Craig Schilling: 03:30Here are some topics we're thinking about with GenAI. What are you thinking about? And there were two that really came out, and one was benchmarking was the first one and and kind of that well, I think benchmarking was actually the second one, but that's how I I landed here was this whole idea of standardization and benchmarking across the industry. And instead of a company leading this effort, I felt strongly that we needed an honest third party broker. And in my opinion, you don't find a better one than University of Illinois.
Tami Craig Schilling: 04:03And so
Todd Gleason: 04:04What do you benchmark against? Well Maybe Vikram Vikram can tell me that in
Tami Craig Schilling: 04:10can talk about that, but but I'll give you the format. I'll give you kind of the situation. Lots and lots of companies, startups, consulting groups, even some farmers and and academics are building generative AI tools for products and agronomy. We need to have some way to compare and contrast them, and we need to hold all of us in the ag industry accountable from an accuracy and a trust perspective.
Todd Gleason: 04:38Okay. So I've got a couple of questions for you about the Bayer product to begin with and only the Bayer product, and then I wanna see what happens Okay. As we get to pick them. So when you were building this on top of ChatGPT, is it accessing a limited scope? Are you feeding all of the data into it and saying, these are the documents you may search?
Todd Gleason: 05:01This is where only where you can find information because if you don't, the large language models gonna find a whole bunch of other stuff that is really worthless. Right?
Tami Craig Schilling: 05:13Yes. So there's this there's this little toggle within ChatGPT.
Todd Gleason: 05:18Yep.
Tami Craig Schilling: 05:19That you can be it's basically they call it a I think they call it a creative dial. You can be really creative, which means you can use all the content out there, or you can have it really tight and only use specific content. For most of the things we're working on, it's really tight. So we direct it by doing question and answer pairs. We direct it to the content we'd like it to look at.
Vikram Adve: 05:45Okay.
Tami Craig Schilling: 05:45But it's not perfect. And so that's part of when people say fine tuning. When you fine tune one of these models for the tool we built, you have to fine tune it. You have to constantly check its accuracy. And that's one of the reasons we haven't launched it to the external world is we know it's gotta be darn close to perfect.
Todd Gleason: 06:05Yeah. So it is still accessing information from other parts of the Internet in that case. Or have you locked it Originally, did you lock it down completely?
Tami Craig Schilling: 06:16You can't really completely lock it down. And and because of the way K. So in the beginning, the our IT folks thought we could. It it's possible. But if you're gonna completely lock it down, in some cases, you want it to learn from what else is out there.
Tami Craig Schilling: 06:37Like, some astronomy situation
Todd Gleason: 06:38bringing the center for DigiLabs in was to help was to do that. And I'm gonna make a leap of faith saying that if you wanna go across the industry that they are the hub, you are a spokes Absolutely. As you started the project, but this is the hub that is looking out to Porteva and is looking at Syngenta and is looking at Bayer and saying these are the things. But, Vikram, you're gonna have to tell me whether I'm right or wrong or even close about that. Tell me about the magic.
Vikram Adve: 07:07Oh, where do I start? So maybe I'll take a step back and say that this whole area of GenAI and chatbots are really in their infancy. So to provide some context, in the Center for Digital Ag, we have an AI Institute for Agriculture called AI Farms, which is funded by the USDA. It's a research institute. And within that, we have a research project on GenAI for agriculture called CropWizard, which is basically exploring different ways in which GenAI can be used and play a role in eggs and what kind of benefits can it have.
Vikram Adve: 07:45And part of that is figuring out how to improve the accuracy and how to improve the capabilities of the underlying platforms. And part of that is, can we constrain the answers as well as possible? And so, for example, the underlying software stack underlying crop wizard is called UIUC. Chat. And UIUC.
Vikram Adve: 08:08Chat was originally designed, originally developed to provide virtual TAs for classes. But the faculty really want answers from that service to be limited to the course materials that they uploaded to the website. And so you can, in the prompt, be very specific and very insistent that it only take answers from the documents that you provided. And it will do that almost always. And so with high confidence, you can limit it to a set of particular documents.
Vikram Adve: 08:40Now disconnecting from the internet, just to pick up on something you said earlier, is separate from the pre training. So the underlying large language model was trained on a vast amount of data from the internet. And so it can bring in that pre training even if you disconnect it from the internet. If you do connect on the internet, it can do an additional search for current information.
Todd Gleason: 09:02When Bayer came to you and said, I got this thing, we kinda like to use CropWizard, we need you to be involved. How did you all consider that and what did you come up with as a solution? Well, so
Vikram Adve: 09:17to be clear, what we're really working on most closely is the benchmarking side of this. And so in the CropWizard project, we have been doing research on benchmarking. In fact, we have a couple of papers on how you evaluate the answers from these kinds of systems. We built a benchmarking pipeline with questions, answers, and evaluation methodology. That's not easy because the answers are not short.
Vikram Adve: 09:43It's not like a single number or a multiple choice question. The answer is, as you've probably seen yourself, you get a page or half a page of text. And so even if you think you have the right answer, comparing them is not easy. And so the pipeline actually uses AI to do the evaluation. And so this benchmarking pipeline is what we've been working with Bayer on.
Vikram Adve: 10:02They have done extensive work on benchmarking internally in the ERI project to evaluate their own systems. And so talking to Tammy about this at that time, she pointed out that it could be really important to have a public benchmarking service because there are a number of companies that are creating these kinds of question answering systems, commercial services for agriculture. But as we realized later, farmers and agronomists and many other scientists are already going to the chatbots to get the answers. You don't need new services for that. And so they have no way to evaluate how good those answers are.
Vikram Adve: 10:37There's no public benchmarking system service or any of that or dataset even to really do a systematic evaluation. And so talking about that, we decided that we should create a public benchmarking service, and that's what's called AI Agribench. And so that's an outgrowth of the CropWizard research and the work that's been done at Bayer and some other companies. But now we have this benchmarking consortium we're about to launch. It's led by the Center for Digital Agriculture, but Bayer and John Deere and Microsoft, the Extension Foundation, which is a nonprofit group that brings together many of the extension units across the country, Kissan AI, which is a smaller company, early stage, but has really powerful technology in this area.
Vikram Adve: 11:24These are the founding members. We have a couple of other companies that are interested in joining as well. And one of the things we're trying to do today is to also invite other companies to join this effort. And the goal of this is to actually give people the ability to evaluate how well these question answering systems are working. How accurate are they?
Vikram Adve: 11:45How complete? How relevant? How concise? And that can give people a way to actually understand for ChatGPT or Gemini or CropWizard or the Extension Foundation, the Extension Bot, or Eli, or many other commercial services, how well are they actually doing in answering these kinds of questions?
Todd Gleason: 12:08How would a producer or a user see the benchmarking in the process? I mean, you go into Jim and I or ChatGPT and ask you the question, is there a is the expectation that there will be a pop up along with the answer that says answers from this are rated a six of 10? Or I I I don't quite get the benchmark process and how a producer might see it.
Vikram Adve: 12:36Right, yeah. So the benchmarking is gonna be basically published results where we are coming up with a set of questions and answers that we consider representative of questions a user might ask about agronomy, and we are having these reviewed by a panel of experts. So we have 18 agronomists today and we're adding a few more. But they are manually reviewing these. So this is sort of ground truth questions and answers that we trust, and we evaluate all of these services on this set of questions and answers.
Vikram Adve: 13:12And with our evaluation pipeline, we can give a score for the answers to this set of questions. We don't evaluate live questions because we have no way to know what the right answer is. So that's why we need a curated set of questions and answers that we know the right answers to.
Todd Gleason: 13:30So going back through my Ag Econ day, and Tammy, you may remember this from the PharmDoc team. Before they were the PharmDoc team, the AgMass project that took a look at all of the brokerage services. They entered themselves into the benchmarking system and did not like it when they were in the bottom 100 as opposed to the top 10 of the brokerage services based on their recommendations. Yeah. But I assume that the end product for you will be something similar where there is a place that people can come and look, and it says, Bears Answers, we have high confidence in.
Todd Gleason: 14:09That's right. Snake oil answer over here has a really low confidence. Is that where you're at?
Vikram Adve: 14:15That's exactly right. So there's a website we're creating and that will have the results. But we're trying to be a little bit more detailed than just a single list because we have four different metrics, accuracy, completeness, and also we have subcategories of questions because, for example, a particular service may be focusing or specialized on something like, let's say, pest management, or maybe it's all of crop protection, but a different one may be more interested in nutrition and more general farm management questions. And so people will have the ability to look at just a subset of the data and look at how well you're doing on particular types of questions and then figure out which tools or which services do well on that type of question. And so all of that is going to be built into the website.
Vikram Adve: 15:01So that's called a leaderboard, and this is actually a very popular thing. So in fact, a lot of the progress in AI has been driven by benchmarks. And there are public benchmarks and all of the big labs publish results on these benchmarks. The benchmarks show where the gaps are, and so they actually motivate people to get better. And so they're actually critical in driving progress in all of these kinds of technical research.
Vikram Adve: 15:27It's true in AI, it's true in computer hardware, it's true in many areas where you need to be able to measure progress, you need to figure out where the progress is needed. And so it gives both the motivation and the yardstick for driving progress.
Todd Gleason: 15:42So, Tammy, you are at the cross section between these two worlds, technical, computer, academic within the university system. And I'm here in the commercial side within the industry system. I need you to understand more information and you're wanting to make these benchmarks work. You've had to think long and hard about what that needs to look like. What kinds of things have you shared back and forth between the center for digital ag and yourself?
Tami Craig Schilling: 16:15Yeah. So one of the things people probably don't realize is that companies in the ag industry of all different sizes, startup, midsize, large corporations, we work together on in areas that are what we would call precompetitive. So an area like this of a new technology and figuring out how do you ensure the accuracy, a high level of accuracy of a new technology is a precompetitive discussion. So it's something that we go to conferences together. We attend and are part of organizations.
Tami Craig Schilling: 16:52Like, one of them is AgTech Alchemy. And it it was started by three folks that are in the ag tech space, and we are all members, and we go to their events around the country. And the reason we do this is because many of us in the industry have spent our whole careers in ag, and we know there's a farmer at the end of this that is expecting an answer, and they wanna trust the accuracy of that information. At Bayer, we were doing our own benchmarks, and that was good for us to get better. But my husband farms.
Tami Craig Schilling: 17:32He's looking for information from lots of different companies. He and my son are gonna use more than one chat tool most likely. I want those companies to feel the same pressure to have the right answer. And your corporations I feel pretty good your corporations will because we have big legal departments. But what we were hearing from some of the startups is they don't have enough funding or support to do as much benchmarking.
Tami Craig Schilling: 18:00They don't have as many developers and people to do it. But let's do it together, and we're all gonna be better off. And the ultimate winner is a farmer because they're gonna get the right information, whether it's by using a tool or by whether having maybe a crop specialist or an agronomist or consultant that uses a tool that is trustworthy. And so then there's a second piece of this, though, that I'll mention, and it's also regulatory management. So a lot of governments around the world are trying to figure out how to regulate this new technology.
Tami Craig Schilling: 18:34And we are following it from what so OECD, which is a global economic consort global economic effort, United Nations related people in different countries are working on different levels. And one of the things we know in ag is regulations add cost. And regulations also will limit companies from doing business or bringing new technology and tools into certain markets. Well, that, again, limits a farmer. So let's try to self regulate and and show we are trying to get accuracy and get correctness and build trust so not every single country has to build their own.
Tami Craig Schilling: 19:22So in The US, it's not bad if we have one set of regulations. But in Europe, it's a lot of chaos because small countries the size of many states are building their own structures on what you can do and what you can't do in this generative AI space. And they're doing it for the consumer that lives in their country, but it's causing a lot of havoc. And so what we wanna do is have it trustworthy for farmers, trustworthy and accurate so that policymakers believe it, and competitively, we are all getting better at this, and there's trust in the tools so that farmers feel like they wanna use it.
Todd Gleason: 20:01How do you expect companies now big companies is one thing to be involved. I think that actually that feels like that's easier. Actually, maybe even simple. But a little entrepreneurial company wants to be benchmarked. How are they incorporated?
Todd Gleason: 20:18How is their data checked? How do you end up benchmarking them over time? And is there I I I but it it could also end up being that you have a benchmarking system that only benchmarks the companies that are in it and not those that are are not.
Tami Craig Schilling: 20:35So we view it as we want as many people as possible to join the group. We think it's it really gets better that way. And it also provides support so not everybody's spending the cost to do it. It really provides an efficiency is how we've gone about it. And it's a very interconnected ag tech is a big topic, but it's not a ton of people.
Tami Craig Schilling: 21:02And so we are getting to know those people. I've been to two of the Ag Alchemy events, one in Seattle and one in California, and attending the different conferences. And whether it's Bushell or others, we're constantly talking with them about, would you be interested? Currently, they can be on a mailing list. So that's one of the things.
Tami Craig Schilling: 21:26We said you don't have to be a member to start off. You can start off as on a mailing list and just get information because we want as many people involved as possible, we think, and internationally. We don't want it just to be limited to The US.
Vikram Adve: 21:39If I can add one more thing to that. So I've talked to, I think, three others. So besides the one that she mentioned, Kissan AI, which is a small company, I've talked to three other, two small and one relatively small, but not that much bigger company that are all interested in joining the consortium. And from talking to them, I think the value that they see in it, so the motivation for them is that they think they actually have better technology than some of the larger players. So I suspect that they are the ones who are going be eager to join in greater numbers because they want to be able to show that they have competitive products, leading products potentially.
Vikram Adve: 22:19And your concern about the data, I think, is important, but the way it gets handled act like alchemy. The the way that your concern about the data gets handled is that when we create the benchmark, it includes the data set that we use for the evaluation, the ground truth data set. That's the level at which we are careful about what data we include and what we don't, and to create something that's really unbiased in terms of any particular company's products. So far, we haven't had to face this because we're using completely public documents, in fact, from land grant universities. In the future, and we're going to continue to do that as long as we have that available.
Vikram Adve: 23:02I think that in some cases, we're gonna have to bring in proprietary data or data source from companies, and then we're just gonna have to find a way to basically randomize and make sure that the evaluation is statistically fair across all companies.
Todd Gleason: 23:21Is the process benchmarking sort of the value of the data? Is there also a standardization function within the consortium similar to standardization for tech companies used for white Wi Fi and Bluetooth and USB a and C and those sorts of things?
Vikram Adve: 23:41So we have not tried to do any effort at standardization in this effort. We have talked to other groups that are doing standardization for agricultural data. There's a group that's run by the IEEE that's trying to do this. There's Ag Gateway, which has been doing this for quite a while, but sponsored by commercial companies. So there are other organizations, and I think that's a much bigger but also much different problem and much different goal.
Vikram Adve: 24:11So it's not really within our scope at all. We are trying to basically focus on evaluation.
Todd Gleason: 24:17How much computing power does it take? Is it a Mac mini or is it blue waters?
Vikram Adve: 24:23It's a lot. It's probably a little bit closer to blue waters than a Mac mini. And the reason for that is because we so let's say we have about 500 questions, question answer pairs, which is on the low end of what a benchmark will be. For every one for every answer they give us, we're using three different LLMs to judge the answer they give us, comparing it to their ground truth answer and evaluating four different metrics. And now that's 500 QA pairs, three different LLMs for every service we evaluate.
Vikram Adve: 24:56We want to evaluate there's probably 15 or 20 different public chatbots that we will evaluate, plus any commercial propriety service that is willing to be part of the leaderboard. And so that adds up really fast.
Todd Gleason: 25:10Yeah. For those who are listening or do not know, a Mac Mini is about $500. However, Blue Waters is a building that is next to the State Farm Center. If you look to the west, it's straight across an entire building that is only a computer.
Vikram Adve: 25:27It's For a computer. Yeah. Yeah. Yeah.
Todd Gleason: 25:29It is only for a computer. That's that's there. So those are the two differences that we're trying to that I was trying to
Vikram Adve: 25:34scale. Yeah. That's right. Yeah. Yeah.
Todd Gleason: 25:36Okay. No.
Vikram Adve: 25:37It's it's a good way to
Todd Gleason: 25:38We begin to wrap up this conversation. What are the takeaways for the listeners and for me today as it's related to where crop wizard, where the benchmarking from your side might go well, from what your side used to be might go, and how you're gonna put it together with all of these other companies.
Vikram Adve: 26:00The main takeaway I'd like to leave people with is that this latest generation of AI tools, generative AI, has some really important use cases in agriculture, which even the use cases we know about today can dramatically transform how farms are managed, how products are developed, how software is built, and many other applications of it. And that includes advisory services for farmers. It includes management tools that can reduce a lot of the drudgery in dealing with data, it includes management tools that can provide more sophisticated decision making by taking into account a lot more farm data, public data, weather data, soil data, and all of those things into a single sort of unified model. All of these capabilities exist in the technology and are being developed as applications by companies, by research projects like ours in CropWizard. And over time, are going to be available.
Vikram Adve: 27:05The big challenge, I think, is adoption. I think people need to be able to gain trust in these kinds of capabilities. They need to be able to see the value. They need to be able to quantify the return on the investments. And this spans both small and large farms.
Vikram Adve: 27:22So I think one of the misconceptions about advanced technologies is that they're really meant for large production commercial farms. And I think they certainly are important for those and maybe will be adopted fastest by those in the large scale. But many of the technologies I'm talking about, especially the question answering and simple automation capabilities, I think will be accessible for the smallest farmers all over the country and in other parts of the world as well. And I think that's really important as a way to democratize how technology can enter the farms because AI has the ability to be very low cost as a software as a service kind of product. And at low cost subscriptions, you can get access to these technologies.
Vikram Adve: 28:06And so if you're gonna do that, you need then the benchmarking to be able to evaluate them and gain trust in them as well.
Tami Craig Schilling: 28:12You know, Todd, I I've spent my whole career thinking about how to improve the experience for a farmer in whatever we brought, whether it was an individual product. Well, today, I look at the technology that way. And it it really goes to what Vikram said, and this is where it was so easy for us to come together, that if a farmer doesn't trust a tool to give them good answers, they won't use the tool. And if they believe that the tool or the technology, generative AI, can't be trusted, they won't find they won't use it. And so that's been the impetus.
Tami Craig Schilling: 28:58And, really, since I've retired, I stayed on the group as an independent adviser. And I'm so passionate about this because these tools are gonna be developed by companies, large, small, in between, universities, organizations. They're gonna come. I want them to be right. I want my husband and every other farmer around the world that's going to use them to get right answers.
Tami Craig Schilling: 29:25But I also want farmers to be able to adopt them and feel that they can be trusted and to have a way to look to see if they can be trusted. Just like today, you can go to see if something on the Internet is true or not. If you go to Snopes or if you go to one of the other other checker, you know, we didn't have that. Before the Internet, you didn't need that. Right?
Tami Craig Schilling: 29:50Because you didn't have that environment. So this to me is a natural evolution. The final, our final ask is if you are an agronomist or you have interest in this space and would wanna be involved, go to the website. Go to the Agribench website. You can find all kinds of stuff on LinkedIn about the organization.
Tami Craig Schilling: 30:14Fill out a form. There's a a form online you can fill out. Be added to a list. Get more information. If you're an agronomist and have interest, we want more agronomist reviewers.
Tami Craig Schilling: 30:27If you're an agronomist, come and join. And event you know, if you're a company, come and join. If you're an individual that wants more information, come and join. We believe and I personally believe this. I started my career thirty five years ago and got to hear the story about the future of biotech.
Tami Craig Schilling: 30:46In 1996, I got my first 30 bags of Roundup Ready soybeans to distribute to six farmers, five farmers or six farmers. I know it was five or six bags of farmer, which ain't very much soy ain't ain't very many soybeans for farmers to try it. And today, that technology is in generation two and three, four, and lots of other companies are doing it. These kinds of efforts, precompetitive efforts, happened back in the day to set up regulatory standards, to get it so farmers could trust technologies. And it's just really cool that we're at a place where we can take some of those learnings and apply it to the digital space and do some really neat things.
Tami Craig Schilling: 31:35And the idea is to help to help growers and all those that provide good advice and information to be more effective and to be more efficient in what we're doing because agriculture is moving at the speed of light.
Todd Gleason: 31:50Thank you so much, Tammy, Craig, Schilling, and Vikram. Advi, we appreciate you being with us. Tammy is a former VP and agronomic digital innovation professional for Behr. And Vikram, of course, works for the University of Illinois at the Center for Digital Agriculture.