DATTS

The Most Frequent Decision in a Steel Plant and How AI Learns From It

Back to series

Every steel plant runs on plans. Yet every steel plant is actually governed by decisions made after the plan stops being valid.

The paradox is simple:

The most frequently taken decision in steel operations is also the least documented— and it is the richest source of intelligence for AI.

That decision is: How planners override the plan when reality changes.

The Invisible Decision That Runs Steel Operations

Across raw material yards, steel melt shops, casting machines, rolling mills, and dispatch areas, planners and supervisors continuously decide:

“Given current constraints, which job, heat, slab, or coil should move next—and why?”

This decision happens:

  • Every shift
  • Under time pressure
  • With incomplete information
  • Across tightly coupled processes


It is not a single decision — but thousands of micro-decisions.

Yet in most plants:

  • The original plan is stored
  • The final execution is recorded
  • The decision logic in between disappears

Why This Decision Is Rarely Documented

1. It Is Contextual

Each decision depends on a unique mix of:

  • Metallurgy
  • Equipment condition
  • WIP state
  • Delivery pressure
  • Human experience

Static systems struggle to capture this context.

2. It Is Time-Critical

When a caster stalls or chemistry drifts, documentation is not the priority—survival is.

3. It Lives in Human Judgment

Experienced planners rely on intuition built over years:

  • What usually works
  • What often fails
  • Which risks are acceptable today

This knowledge is rarely written down.

4. Systems Capture “What,” Not “Why”

Most APS and MES systems record:

  • What was planned
  • What was executed

They do not capture:

  • Why the plan was overridden
  • Which alternatives were considered
  • What risk trade-off was made

Planner Overrides: Not Exceptions, but Intelligence

When planners override APS recommendations, they are not “breaking the system.” They are compensating for uncertainty.

Each override answers three unspoken questions:

  1. Why was the system’s recommendation insufficient?
  2. Which alternative was chosen instead?
  3. What outcome was expected?


These answers are never explicitly stated—but they can be learned.

How AI Learns From Planner Overrides

AI does not ask planners to explain themselves. It learns by observing decisions in context.

Step 1: Capture the Decision Context

At the moment of override, AI captures:

  • Plant status
  • Resource constraints
  • Material condition
  • Order priorities
  • Time pressure

This creates a decision snapshot.

Step 2: Reconstruct the Option Space

AI identifies:

  • What APS recommended
  • What other feasible options existed
  • Which constraints were binding

This avoids hindsight bias.

Step 3: Classify the Override

Overrides are grouped into patterns such as:

  • Protect quality over delivery
  • Protect yield over speed
  • Avoid equipment risk
  • Preserve sequence stability

This abstraction allows learning across many cases.

Step 4: Link Decisions to Outcomes

Using MES execution and quality data, AI evaluates:

  • Actual delays
  • Quality results
  • Yield and rework
  • Downstream disruptions

Crucially, outcomes are traced back to the override decision, not just the plan.

Step 5: Learn Patterns Over Time

Across hundreds or thousands of overrides, AI learns:

  • Which decisions consistently work
  • Which depend heavily on context
  • Which of those overrides creates hidden downstream problems

Human judgment becomes statistical knowledge.

To summarize, these are the decisions AI can learn over time:

What AI can observe and learn

Why MES Is Central to This Learning Loop

MES sits at the intersection of:

  • Planning (APS)
  • Execution
  • Quality
  • Genealogy

Traditional MES implementations focus on:

  • What was planned
  • What was executed

MES is the natural system to:

  • Observe decisions
  • Preserve context
  • Close the feedback loop

This transforms overrides into decision events, not anomalies.

What Changes When AI Learns From Overrides

APS evolves from:

This is the optimal plan.


to:


In similar situations, planners chose this—and it worked.


Benefits include:

  • Fewer manual firefights
  • Higher planner trust in recommendations
  • Faster onboarding of new planners
  • Reduced dependency on individuals
  • Continuous improvement without rigid rules
How AI learns

What AI Does Not Do

To maintain trust:

  • AI does not blindly copy every override
  • AI does not remove human authority
  • AI does not eliminate planning accountability

It supports, challenges, and learns.

From Tribal Knowledge to Digital Memory

The most valuable knowledge in a steel plant is not stored in procedures—it lives in repeated human decisions under pressure.

By learning from planner overrides:

  • Experience becomes data
  • Judgment becomes scalable
  • Planning becomes adaptive

In a Nutshell

Before and After AI

The most frequent – and least documented – decision in steel operations is how planners override the plan when reality changes; by learning from these overrides, AI turns invisible human judgment into adaptive planning intelligence without taking control away from people

Disclaimer: The opinions expressed in this article are personal.