• Upstream Ag Insights
  • Posts
  • What Agribusiness Leaders Need to Know from the MIT State of AI In Business Report

What Agribusiness Leaders Need to Know from the MIT State of AI In Business Report

The real reason most corporate AI efforts flop (and what ag can do differently).

Author

Shane Thomas
November 01, 2025 • Read Time: 12 minutes

Index:

  1. Overview

  2. Who’s winning and what agribusiness can learn

  3. Why 95% have failed and what that means for agriculture

  4. Organizational design determines success

  5. Context is King

  6. Buy vs. build

  7. Other takeaways

  8. Final Thoughts

AI has dominated the corporate conversation over the last three years.

Every major business is talking about and thinking about AI. Some even add it into presentations just to participate in innovation theatre for their board, investors or customers.

A new “State of AI” (released Oct 9) report suggests that over 50% of US based enterprises pay for AI models:

Agribusinesses are likely aligned.

AI comes up on investor calls, town halls, board meetings and more, whether it is AGCO, Corteva, John Deere, Nutrien, Bayer or anyone else — they have announced some level of AI initiative through various methods such as customer engagement tools, predictive maintenance, or digital agronomy.

The real question isn’t whether companies are doing anything, but whether they are getting a return today, or implementing it effectively to differentiate their business and derive a return in the future.

This week, I collaborated with Headstorm to dive into the MIT State of AI in Business 2025 report and unpack the why just 5% of companies are getting results from AI and apply to the modern agribusiness enterprise.

This is part one of a two part series. Part two will pull Headstorm’s cross-industry insights into the conversation to better apply what else agribusiness can do to improve AI strategy.

Who’s winning and what agribusiness can learn

In other industries, the winners share a common pattern: they deploy AI in specific, narrow, repeatable workflows that allow systems to learn from feedback.

Financial institutions use AI to reconcile invoices and contracts, cutting processing time by weeks. Logistics firms automate claims handling, training the model each week based on human corrections. Retailers use generative models to auto-draft customer communications, then track which suggestions perform best.

Agribusinesses can, and are, mirroring this approach:

  • A crop protection manufacturer can apply AI to regulatory dossier assembly, where hundreds of hours are spent formatting and cross-checking documentation. UPL has cited they are doing this today.

  • Ag retailers could use invoice-matching AI to automate credit and rebate reconciliation, areas still dominated by manual spreadsheet work.

  • Equipment dealers, like AGCO has shown, use AI for work order triage, where incoming service tickets are categorized, inventory is predicted and managed, and technician schedules are adjusted in real time.

These are not glamorous initiatives, but they’re measurable, repeatable, and integrate within existing systems. They lay the foundation for where AI turns to an operating advantage.

Why 95% have failed and what that means for agriculture

According to the MIT report, the reason AI projects have been not been seeing success isn’t due to technical limitations. It’s due to structure. In organizational ways (more later) and in structure of the AI tool.

Most enterprise AI systems have not been set up to learn.

They don’t retain context, adapt to user feedback, or integrate naturally into existing processes and workflows.

They are jankily deployed and become static, one-off tools that make a good demo but do not become “default use” tools — much like we see in the farm management software space.

This pattern has been palpable within agriculture throughout the software deployment era and it shows in enterprise with AI.

Organizations stuck on the wrong side continue investing in static tools that can't adapt to their workflows, while those crossing the divide focus on learning-capable systems.

The companies that succeed think differently.

This is where control points, and understanding where they sit, and what they influence is important.

For example, if we apply concepts from Reshuffle: Who Wins When AI Restacks the Knowledge Economy by Sangeet Paul Choudary we can think systematically about what AI implementation influences, and how to build a smart system around the control point that not only fits the work flow, but augments it.

According to Choudary, AI compresses value chains by lowering coordination costs.

He thinks about it in three categories:

  1. Relationship control - Who owns the farmer interface when AI-powered recommendations become standard? If that interface shifts from retailer to digital platform, influence follows.

  2. Workflow control - If an AI model automates procurement workflows, who governs the data standards and integrations? That layer is a control point. What system does that occur in today? What is it connected to? Where does data come from to improve it? What other external areas influence its inputs and what external companies are influenced by the outputs?

  3. Intelligence control - When AI models learn from farm and field data, who controls the training data and feedback loops? That determines who captures compounding advantage over time.

In each case, the AI model or platform is just the enabler. The control point lies in how it coordinates others and who depends on it.

The enterprises that become successful with AI are more likely to identify which category they want to improve and build dynamic systems around that process and control point.

Organizational Design and P&L Impact Determines Success

According to the report, AI success isn’t random. It’s the product of structure.

The report found that many large companies centralize AI in “innovation labs” or disconnected arms, away from the operating area of the business. The same challenge occurs with corporate venture capital, too.

These groups run pilots but rarely influence actual operations. They report “exploration progress,” but not value creation.

By contrast, the most successful firms give ownership to the people who run the process and are accountable for the outcome. The report shared that companies succeed when they decentralize implementation authority but retain accountability.

The people leading procurement, those driving sales and marketing or those in charge of logistics. They’re the ones who feel the friction every day, can measure the improvement and make changes that drive better outcomes.

For agribusiness, this might mean:

  • Assigning a Director of Credit to lead an automation initiative in invoice verification.

  • Putting a Regulatory Affairs lead in charge of AI-enabled compliance documentation.

  • Giving marketing the incentives to build more efficient marketing asset generation processes, or better lead prioritization integration with the sales team.

The report emphasizes that these teams need to be small and tied to outcomes, not locked away in an AI division. Ideally, the outcomes are always tied back to the P&L.

Context is King

Poor organizational structure is a central shortcoming, but a lack of contextual learning, and misalignment with day-to-day operations were also high up on challenges.

Most GenAI systems do not retain feedback, adapt to context, navigate edge cases or improve over time. That means teams have to constantly re-prompt, add information in or get frustrated with the lack of understanding, leading to poor usage or value creation.

Designing for constant and adaptive context was a key driver of success with LLM systems.

In a recent Lenny’s Podcast with Scale AI CEO Jason Droege talked about the shortcomings of nuanced context, too.

For example, each person’s job within an entity needs different context, and that context might change day to day, and the context required from one company to the next changes, which means broad models can only be so accurate in a variety of settings, which means there is even more need for customizing these models and systems to specific use cases, and improving the data acquisition and tagging within an enterprise to ensure there is a growing understanding.

Buy vs. build

When it comes to AI, buying beats building according to the report.

Companies that partner externally see faster deployment, higher adoption, and greater success than those developing tools in-house — something that has seemingly been the case in agriculture software, too.

Strategic partnerships studied achieved a 66% success rate, double that of internal builds.

Internal efforts often failed because they take too long, lack workflow alignment, and become “science projects.” Many exceeded nine months before delivering value, and most never reach the field. I think the reality for agribusinesses is that there is a lack of skillset and culture with AI internally that would hinder success, too.

What we found most notable in the report was that top performing enterprises treated AI vendors like Business Process Outsourcing partner (BPO), not software suppliers. They demanded customization, measured results by operational outcomes, and look at adoption by working with managers. The success hinged on process alignment, not just model quality which was a consistent finding from within the report.

Most internal projects fail because they underestimate integration friction and overestimate control. Without feedback loops and adaptive learning, tools lose user trust quickly. Success comes only from narrow, deeply embedded use cases and not broad platforms.

Additionally, we found the keys for decision making for vendors to be notable with the “deep understanding of workflow” being the second most prominent aspect, conveying the need to understand the industry, the company and the detailed nuance of current practices internally:

Final Thoughts

MIT’s conclusion is blunt:

“Organizations that cross the GenAI divide do three things differently:
They buy rather than build, empower line managers rather than central labs, and choose tools that learn and integrate deeply.”

What this suggests is that success with GenAI in enterprises is about focus, change management and a holistic strategy.

As we think about where AI is going, from expertise and answers, to taking action (agents), we see why change management and workflow understanding will be as important as ever if enterprises are going to be able to derive value from AI in the short term.

Stay tuned for Part 2 where a collaboration with Headstorm digs deeper into frameworks for moving AI forward in agribusiness.

State of AI Report 2025 - Airstreet Capital