We need Donnas, not Amelia Bedelias
Most LLM implementations wait for the human to interact. We need to do better.
Welcome to the June 22, 2025 edition of SFTW. Today's edition will cover two topics.
- Most of LLM implementations today are chatbots, which are typically passive in nature. What we need is active experiences, which anticipate our needs, and know when to interject, just like Donna Paulsen on Suits.
- Some quick but important lessons from AgVenture Alliance's Building Bridges series.
Programming Notes:
- SFTW Convos will return on Wednesday with a conversation with Christine Gould, founder of GIGA, and previously founder of the international community Thought for Food. I am expanding the type of SFTW Convo topics by talking about agriculture policy and regulation with Emma Kovak (Breakthrough Institute), genetics & topics related to higher education with Dr. Channa Prakash (Dean of Tuskegee University), genetic breeding with Dr. Brad Zamft (Heritage Agriculture), agriculture economics with Prof. Terry Griffin (Kansas State University), and AgTech experimentation and adoption with wheat farmer Andrew Nelson.
- The second part of the GenAI in Agriculture white paper is available for free to all members
Scaling Regen Ag Starts with Better Data
Farmers Edge is building the digital backbone for regenerative agriculture—designed to simplify Scope 3 reporting and deliver credible results, fast. With Managed Technology Services built for agribusiness, Farmers Edge connects real-time field data to your sustainability goals across every acre and every grower in your network. From in-season tracking to end-of-year proof, Farmers Edge makes regen ag measurable and manageable. If you're serious about scaling impact and reporting, this is the infrastructure that makes it possible.
We need Donnas, not Amelia Bedelias
Not a day goes by without some company introducing a ChatBot agent on top of their existing data. For example, Bushel recently launched an AI assistant (Buddy) for their CRM system. It helps agribusiness teams ask Buddy questions in a natural language like “What customers haven’t been contacted in the last week?” or “How many customers have past due invoices?”
This is good use of the existing technology and makes agribusiness team members more efficient by making access to information more easily accessible through a natural human language interface.
Today, within and outside of food and ag, most LLM based applications are bolt-ons to existing data, where teams have spent time to get the LLM to ingest, and synthesize proprietary datasets combined with public data sets to deliver a different user experience to users.
None of these interfaces have fundamentally changed how we do business or even have made big changes to workflows or created new workflows which help us unlock value in new and different ways.
This is expected, as we are in the early phases of large language model adoption and understanding for business use cases, even though ChatGPT is the fastest adopted product in the history of the world from a consumer side.
The CPO of Microsoft, Aparna Chennapragada, wrote an interesting LinkedIn post titled, “On Horseless Carriages and Horses with Steering Wheels”
We are clearly in a tweener phase, with this split-screen experience where the UI hasn’t fully committed to the model, and the user hasn’t either.
And this is not a model issue.
What we need are interfaces fully native to AI. where the model is in the loop from the start, not added after the fact. That means rethinking control, context, and interaction from the ground up.
Until then, I suspect most users will keep reaching for the reins.
She also says,
ChatGPT knows a lot about your work, remembers your preferences, and could easily check in on things you care about. But it just sort of sits there, waiting, like a super polite guest who won't speak until spoken to.
I believe this is one of the most interesting design challenges right now, and it is not a model issue. The current LLMs are like a shy teenager who is hesitant to make the first move.
The AIs we have are incredibly capable and they could proactively help us, but they are not designed for it. The human has to make the first move.
We need Donnas, not Amelia Bedelias
I have to admit I am a big fan of the show “Suits.” There are many interesting characters in the show, but I am particularly fond of Donna, played by Sarah Raferty.
Donna is one of the best executive assistants in the world. She anticipates Harvey’s needs, is a strategic thinker, has exceptional memory, is a problem solver, has strong communication skills, shows loyalty and discretion, goes the extra mile, is funny, and most importantly knows when to interject with Harvey.
I used to love reading the Amelia Bedelia books (books from the 1960s) to my daughters when they were younger. Amelia Bedelia is a series of stories from the 1960s.
The stories follow Amelia Bedelia, a maid who repeatedly misunderstands various commands of her employer by taking figures of speech and various terminology literally, causing her to perform incorrect actions with a comical effect.
Imagine if the model knows the history of your interaction with a particular grower. This is feasible today, though it might be scary due to privacy and control of data issues.
But imagine if the model comes back and says,
Hey remember the fungicide strip trial you did with the grower on their Back 40? I have received new data, which changes the main conclusion of the trial completely.
or
Hey, remember last year, the grower had told you that they do not like to work with this product in your phone conversation with them? Your current prescription includes this product.
You might say,”is this just more notifications coming from my device or computer? I am already tired of all the pings from Instagram, and Tiktok.”
But this is not the case of a simple notification, but it is space and time of shared awareness between you and the AI. It is a case of the AI which notices something, understands the context, and smartly decides when to say something, because what it has noticed is going to be material to you.
The challenge is not purely technical in nature and so there is room for many other types of companies to solve this problem, rather than thinking that the big boys of tech can only solve it.
It is interesting how we have a lot of talk about agentic workflows, but ironically most of the agents today lack agency.
The challenge is to understand what matters to the human user “in the moment”. This requires deeply understanding how to build trust, and how to model relevance over time.
At least for many of the applications, what we want the agents to do is to behave like super-start executive assistants. The best EAs don’t just respond to what an exec asks for. They are very good at anticipating, filtering, and knowing when to interject.
This is not purely a technology problem but it is a human centered design, behavioral, social, and psychological problem.
Tools won't just need better interfaces, they'll need better instincts.
The good thing is that nobody knows their growers better than people who work with them closely and have trusted relationships. It is the co-op agronomist, it is the equipment dealers, it is the seed salesperson. Every industry, including agriculture has access to this innate knowledge of their customer needs. Even if the agriculture industry will not be developing many of these AI models, they have the know-how, the data, and most importantly the emotional connection and relationships to make these tools much more powerful than just being a chatbot.
We need our LLMs to be like Donna from Suits, not someone who only responds to queries like Amelia Bedelia.
What is the color of your revenue?
When I was at Mineral (an Alphabet company, which no longer exists), during our regular board meetings, we would present the state of the business and then also provide details on the financial performance of the company.
Revenue (or inflows) had to be categorized in multiple ways, but at a high level we would categorize them into two different categories.