Evan Scronce / Enterprise AI
Enterprise AI strategy and design leadership at John Deere.
A personal inquiry into intelligence at organizational scale.
How do organizations turn machine intelligence into decisions they can justify, actions they can coordinate, and knowledge they can revise? This is the question behind my work on AI, design, and the enterprise.
Read the thesisFigure 01 / The decision loop
From prediction to accountable action
Purpose defines objectives and constraints
↺ Evidence returns to representation.
Changing processes or rights requires separate authority.
01The central claim
A capable model can produce a good prediction while the organization makes a poor decision. The objective may be contested. The relevant knowledge may be incomplete. The person receiving the answer may lack the authority to act.
This shifts the question from what a model can do to what an organization can reliably decide, accomplish, and learn. The essays below develop that argument, identify where it may fail, and propose ways to test it.
A working definition / Enterprise intelligence
The capacity to turn incomplete knowledge into justified, coordinated action, then revise decisions and assumptions in response to their consequences.
Which outcomes matter, to whom, and under what constraints?
Objectives and conflicts that require an accountable decision.
What do we believe, on what evidence, and with what uncertainty?
A decision-specific account of the relevant state of the world.
Who may choose, commit, intervene, and revise?
Permissions and responsibilities that prediction alone cannot supply.
What commitment changes the world?
Execution with consequences, dependencies, and recovery limits.
What happened relative to what we expected?
Signals that can distinguish an acceptable result from a failure or an open question.
What should change, and who can authorize that change?
Revisions to beliefs, procedures, objectives, or delegated authority.
02Three propositions
A theory becomes useful when it tells us what to expect, what to observe, and where the argument could fail.
When unresolved decision rights or coordination failures constrain a workflow, a more capable model may improve local outputs without improving the accepted outcome.
Prediction. Removing the identified organizational constraint should produce a larger gain in end-to-end results than improving model capability alone.
Boundary condition. This predicts little when model accuracy is already the dominant constraint.
A handoff should retain the distinctions that could change the next decision: evidence, assumptions, scope, uncertainty, and permission.
Prediction. Preserving those distinctions should reduce decisions based on stale or overextended claims, at a measurable coordination cost.
Boundary condition. Extra structure can be wasteful in stable, low-consequence tasks.
An estimate of correctness cannot determine who is entitled to make a commitment. Delegation also needs a purpose, scope, and account of correction.
Prediction. For disturbances visible before consequences become irreversible, timely feedback and a usable intervention path should improve recovery over confidence-only permission rules.
Boundary condition. Some important harms are delayed or irreversible; feedback alone cannot justify the action.
03Working theory
Definitions, mechanisms, counterexamples, and proposed experiments. A developing body of thought on AI in complex organizations.
A model can become more capable while an organization remains unable to decide. A useful theory of enterprise AI has to explain that gap.
Organizational intelligence · 5 min readUseful context must preserve the difference between what was observed, what is believed, what is required, and what is permitted.
Representation & uncertainty · 5 min readDelegation should account for what can be observed, how quickly someone can intervene, and which consequences cannot be undone.
Control & decision rights · 5 min read04Intellectual foundations
The argument draws on bounded rationality, decision-making under partial observability, and feedback control. These traditions motivate questions here; they do not validate this enterprise framework.
Herbert A. Simon on decision procedures under limited information and computational capacity.
A foundation for asking where a constraint sits, not proof that organizational redesign always beats model improvement.
Kaelbling, Littman, and Cassandra on distinguishing the world’s state from beliefs about it.
A conceptual distinction, not a claim that a narrative enterprise record satisfies a formal decision model.
Åström and Murray on the relationship between observations, actions, and system behavior.
A source of questions about intervention and delay, not an organizational stability proof.
These are proposed distinctions and hypotheses, not peer-reviewed findings or reports of John Deere implementations. The tests have not been run as part of this site. Their value depends on whether they help explain and improve particular decisions.
A conceptual lens for AI-enabled organizational decisions. It does not claim that a company is a single mind or that enterprise intelligence can be reduced to an IQ score.
05About Evan
My role at John Deere spans UX, human-centered AI strategy, global manufacturing, and internal business. Previously, I led UX for Global John Deere Financial.
That context brings me to questions at the boundary of intelligence and institutions: how knowledge becomes a decision, how a decision becomes a commitment, and how the organization learns when the result differs from what it expected.
About my workAn open question
For a rigorous conversation about AI strategy, organizational decisions, or the assumptions behind this work, get in touch.