A large percentage of GenAI based projects fail to get out of the starting blocks due to “incoherent strategic rationale,” sub-par data and infrastructure, and poor execution, says agrifoodtech consultant Rhishi Pethe.
“Imagine [an AI-driven] model recommending a crop protection product that is not a good fit for the farmer’s context, or worse, does not fit the label instructions necessary to apply the product.”
But that doesn’t mean we should throw out the baby with the bathwater, says Pethe, who has penned a white paper commissioned by Bayer Crop Science and supported by AI-powered startups Digital Green, Kissan AI, and Traive to help firms develop successful GenAI projects.
To make GenAI work, he tells AgFunderNews, firms need clarity on strategy and use cases, sufficient maturity in data management and technology, relevant talent with domain expertise, software engineering and data science skills, and process maturity to manage all the governance, trust, legal, and marketing issues that will inevitably arise.
And the first point—strategic rationale—is especially important, he says, as some companies are approaching AI projects as an end in and of themselves rather than a tool to solve a real problem.
“One of my clients came to me and said, ‘We’re getting a lot of pressure from our board that we have to do something in AI. So I said, OK, what problems are you trying to solve? They came back to me with 20 use cases, and for 15 or 16 of them, they just didn’t have the data, or the problem could be solved in a different way.
“People want to use GenAI, but they are not always starting with a problem and saying, is GenAI the right tool for that?”
‘A model that is 80-90% accurate is not good enough’
In agriculture, however, “a lot of the challenges are around knowledge dissemination and getting it to the right person in the right format at the right time,” says Pethe. And in such scenarios, GenAI can be an extremely useful tool, provided the model is validated by experts with domain knowledge to such an extent that it is sufficiently accurate to be trustworthy.
“Let’s say I’m a crop input company and I’m launching a new chemical product,” he says. “I’ve already gone through the expensive process of product development and regulatory approvals, but after that, there’s a whole lot of education that has to happen with your dealer network, your salespeople, and your growers on how to use this new product. And this requires some level of expertise which not always there.”
An AI assistant is ideally placed to go through reams of documentation, CRM data, safety data, regulatory data and other public or private information to answer questions from salespeople, dealers, or end customers in a fraction of the time it would take a human, he notes.
“But if the cost of a bad decision is high [you apply too much of the said chemical and waste money, apply too little and suffer crop failure, or recommend the wrong product altogether], you need a very high level of accuracy. A model that is 80-90% accurate is not good enough. So there will be trust issues.”
But trust can be built, he says. “You need to give the model the right type of data, build the right benchmarks, train it accordingly, and then validate with a domain level expert.”
Faster, higher-quality responses to grower questions
Bayer, for example, “has a product portfolio that is getting more complicated as time goes by,” says Pethe, who presents the firm’s GenAI based tool E.L.Y. in his white paper. “It believes regulation is going to get more stringent or more complicated. And because of that, how you position your product for a given context becomes more difficult to answer with certainty.
“For example, if I’m a grower and you are a sales rep for Bayer or a technical agronomist and I have some questions for you about a crop protection product, you might be able to answer 50% of my questions, but you might not have the information to answer the other 50% at your fingertips. So you might say, ‘Let me just call up some people and get an answer for you,’ and that may take some time.”
In some cases, you might simply need to refer to a crop protection label, rather than a colleague, he said. But this also takes time. “Each of those labels can be 20, 30 pages long, and there are so many different rules around how and when to use the product.”
Another challenge in any business where domain knowledge is important is that you lose expertise whenever a person leaves the team, he added.
“And so the problem statement at Bayer was: Can we make it easier to answer people’s questions with confidence? It has a sales team and a bunch of agronomists out in the field, all of this publicly available information about labels and agronomy data, and hundreds of webinars for its growers and dealers.
“So it was a case of, can I have all of this information in my pocket and make it easily available so that the response time to the grower is better, the quality of the response is better? Some of the people [at Bayer using this to assist growers] are saving 10% of their time, four to five hours a week.”
As for the data itself, he said, it is complex and comes in multiple formats: “A lot of the data is external, EPA regulations for example, and Bayer has taken that data and combined it with its agronomy data, field observations, drone imagery, internal training videos, seed trial data, presentations, and customer support tickets to basically create an agronomist on steroids.”
Bayer then used its team of inhouse domain experts to do initial testing and validation of its model, explained Pethe. The model was then tested over four cycles with a small subset of users from a particular region with expertise in crop protection products in particular and agronomy in general.
The testing cycle was then expanded to a larger subset of a few hundred users across multiple regions, with feedback continuously fed back into the model to improve it, he explained.
Bayer uses the following five criteria to evaluate the performance of its E.L.Y.
1. Accuracy: How accurate the responses compared to the desired response as determined by a team of human experts?
2. Groundedness: Are the responses based on the information contained within the documents and data provided to the GenAI model?
3. Cost: What is the cost to build the model, maintain, and to provide response to user?
4. Response time: What is the response time to queries provided by the end user?
5. Consistency: Are the responses consistent when the same question is asked within the same context to the model?
Every time the model is updated, it is re-evaluated according to these criteria.
Farmer.Chat via Whatsapp
Another firm with a case study featured in Pethe’s white paper is Digital Green, a nonprofit using an AI-powered assistant called Farmer.Chat to provide smallholder farmers in India and Africa with localized, timely, and actionable advice tailored to their needs.
Using voice, text, and image inputs, Farmer.Chat coaches farmers through critical decisions such as crop management, pest control, and market timing, said Pethe. “Digital Green has been around for several years making videos for farmers to help them make their farming operations more profitable.”
Historically, he said, Digital Green has provided its videos to extension agents who go out to villages and set up screenings. By using GenAI, the Farmer.Chat system enables farmers to interrogate Digital Green’s data coupled with publicly available data from universities and other sources on their phones via Whatsapp in their local languages.
“With this type of an approach, Digital Green can support additional languages at a much lower cost and reach far more people,” said Pethe.
Credit risk predictions
Traive, a subsidiary of BASF, meanwhile, has been testing GenAI to see if it can predict when a farmer may fail to pay an invoice after 90 days, explained Pethe.
“Traive used advanced methods such as Graph AI and Deep Learning to accommodate different data types such as satellite imagery, text, financial reports, and historical loan data sourced from nine of the largest financial and supply chain entities in Brazil with delinquency rates varying from 1.05% to 22.27%.
“It used fine-tuned language models using GenAI through a product called ‘Travis’ to let credit analysts ask natural language questions about risk scores and credit memos, which was typically a black box for them, and turn it into a data-driven conversation.”
Fix your data, then get started
Stepping back, said Pethe, “Where there is a lot of unstructured data and tribal knowledge, that’s where there’s a lot of scope for GenAI to improve the user experience. For example, with Digital Green, being able to talk in your local language is a huge improvement.”
He added: “I think a lot of people are feeling fairly positive about using GenAI in their businesses, but in lot of cases, they don’t have the right data structure to get started, so they might need to improve that to take advantage. As an example, a lot of people are still using desktop client software, so there’s all this data sitting on local machines that’s not in the cloud [for an AI model to train on].
“So those type of basic challenges still exist in certain parts of the industry; companies are realizing that they can’t take advantage of these [GenAI] tools until they fix their data.”
