The Information Flow Thesis¶
The Pattern¶
Cleveland-Cliffs does not have an AI problem. It has an information flow problem that AI can solve.
Four sites, two industries, 40+ conversations, one pattern: the information people need to make good decisions either does not exist in digital form, does not flow to where decisions are made, or arrives too late to act on. The result is reactive maintenance cultures, procurement bottlenecks, quality investigations that start from scratch, and production scheduling driven by institutional memory rather than integrated data.
This is not a technology gap. Your plants are instrumented. You have historians with a decade of process data, CMMS systems with asset hierarchies, quality inspection systems collecting defect images, and fleet telematics tracking every haul truck. The data exists. What does not exist is the connective tissue — the information loops that turn data into decisions and decisions into learning.
Every AI project we recommend closes one of these loops. The roadmap is not "deploy AI across technology clusters." It is: make each plant smarter by making information flow, starting where the pain is greatest, building a progressive data foundation, and layering predictive capability on top as the data matures.
Cleveland: The System Is Broken¶
Cleveland showed us the problem at its most visible.
The 1SP continuous caster — the most capacity-constrained asset in the company, booked at 100% utilization through May — loses production every day to unplanned downtime. Maintenance runs 70% reactive, 30% planned. When equipment fails, a technician gets a radio call with minimal context, climbs to the unit, diagnoses from decades of experience, fixes it, and moves on. Nothing is documented. Months later, the same failure occurs and the diagnostic starts from scratch.
The root cause is structural. The 2001 ISG restructuring eliminated inspector roles and combined the planner, reliability engineer, and area manager into a single position. The process discipline that kept reactive maintenance in check was never rebuilt. Today, paper PMs are crumpled and discarded. Requisitions are silently cancelled after 60 days. Operations tracks delays in one system using one naming convention; maintenance tracks work orders in another system using a different one. There is no cross-reference between them.
The constraint chain is clean:
- Production throughput is constrained by the 1SP (28 heats/day target)
- Missed production comes from unplanned downtime
- Unplanned downtime comes from a heavily reactive maintenance culture
- Reactive maintenance persists because information loops are broken
When Paul Aaron Dash, your maintenance leader with decades of institutional knowledge, summarized the situation, he was precise: "We're not closing the loop. We go out, fix it, and move on. We never learn from what we just did."
And yet the proof that the solution works is already on the floor. Phil Thorman's slab cut optimization — connecting Pi historian data to scheduling decisions — delivers $3 million per month. One person, one information loop closed, measurable results.
Middletown: The System Works — But It Is Fragile¶
Middletown proved the thesis by being the exception. Your Middletown steel shop is the best-performing in the company: near-zero unplanned turnarounds while other sites average 5-12 per week, scrap below 1% when others run above 10%. R&D described the quality difference as "an order of magnitude."
But it works because of people, not systems. When those people retire — and many are close — the information they carry retires with them. Three separate leaders independently converged on the same diagnosis: knowledge silos are "the biggest problem facing the plant."
The information flow problem at Middletown manifests differently from Cleveland. Here it is not that the system is broken — it is that the system is invisible, carried in the heads of experienced practitioners. The CMMS (Teams/SWAMI) is a legacy Armco-era system with no vendor support. Quality investigations rely on institutional memory of defect patterns. Maintenance planning lives in local files and personal notebooks.
Meanwhile, Middletown has the longest finishing chain in CLF — from the vacuum degasser all the way through to electrogalvanizing, aluminizing, and enameling. A chemistry issue at the degasser may not surface until six process steps later at coating inspection. When it does, the quality team starts the investigation from scratch every time because there is no through-process data linkage.
The Ametek surface inspection cameras are installed on four lines but classifying at only 60% accuracy on the defect types that matter most. That is fixable — bounded ML, existing hardware, measurable improvement. And your R&D team has already started building AI models on their own initiative, choosing Middletown's BOF as the test site because it has the best existing model to beat. That kind of organic momentum is rare and worth accelerating.
Cleveland showed us the broken system. Middletown showed us the fragile one — and the urgency of capturing what's in people's heads before it walks out the door.
Tilden: The Thesis Crosses Industries¶
When we walked into Tilden, we expected to find the same pattern we had seen at two steel mills. What we found was different — and even more compelling.
Tilden does not have a maintenance problem. The fleet runs 70% planned, 30% reactive — the inverse of Cleveland. There are reliability engineers, SKF analysts, and real-time tire monitoring. Mining maintenance is well-resourced because the consequences of a $300,000 wheel motor failure on a $12 million haul truck are visceral and immediate.
But Tilden does have an information flow problem — it is just in a different place. The concentrator, designed in 1974, processes ore that has fundamentally changed over fifty years of open-pit mining. Clay content is increasing. Reagent effectiveness is declining. The plant spends $50 million per year on chemical reagents and recovers approximately 70% of the iron content — below the plant's design benchmark of ~75% and well below the 80% achievable with optimized hematite flotation.
The data exists on both sides of the gap. Every drill hole is sampled at approximately 10-foot spacing — the mine has rich compositional data about what is coming out of the ground. The concentrator has a DCS and a G2 fuzzy logic control system that has been managing grinding operations for years. But there is no predictive link between what is coming out of the ground and how the concentrator should respond. Operators adjust reagents reactively, wait days for laboratory results, and adjust again.
This is the same thesis — information does not flow where decisions get made — expressed as a process chemistry optimization problem worth tens of millions annually. A 5% improvement in reagent efficiency alone is $2.5 million per year. Closing the recovery gap from 70% toward the 75-80% range achievable with optimized hematite flotation adds throughput on a fully-booked pellet plant where every additional ton of pellets has immediate revenue value.
The Tilden readout was the strongest of the engagement. Ryan Korpela, the operations manager, called the work "really impressive" — twice — and validated all four proposed projects without raising a single budget objection. His ranking was clear: concentrator feed-forward optimization first, then pit flow intelligence, then logistics. The grassroots champions at Tilden — Adam Bingham already using Copilot to build knowledge bases, Pete Austin modeling fleet lifecycles in Excel, Jeff Domann conceptualizing drill-and-blast optimization for years — represent the kind of organic technical appetite that makes adoption succeed.
Four sites, two industries, one pattern. The information flow thesis is not a steel problem — it is a Cleveland-Cliffs problem.
Burns Harbor: The Full Chain — and What's at Stake¶
Burns Harbor is the largest integrated steel mill in the portfolio — five million tons per year, four thousand people, the only site running the full chain from coal to plate. It gave us two things no other site did: a proof of what integration looks like, and a warning of what happens without it.
The BF department at Burns Harbor is the best-integrated operation we observed across four sites. One process engineer and one PA coder spent five years building shared HMI screens, data-driven daily meetings, and GPS-verified material tracking. They did it with zero budget and no corporate mandate. That is what intentional information integration looks like — and it was built by two people who decided the status quo was unacceptable.
But that integration is person-dependent, not systemic. And fifty miles away at Indiana Harbor, we found the opposite extreme: four production areas on different radio channels, no Wi-Fi, turn call repairs with no work order, 282 unread predictive maintenance emails in a single month. One maintenance engineer uploading binders from the 1970s into a personal AI tool because no institutional knowledge capture system exists. A hundred people turned over since 2019.
The coke plant is the most urgent knowledge capture case we encountered across all four sites. Three of five section managers are at retirement age. The electrical manager has 54 years of tenure. The PhDs who built the coal blend optimization models were let go approximately 18 months ago — nobody knows where the models are.
Burns Harbor's longest process chain amplifies every upstream information gap. Off-chemistry heats at 5% incidence, each representing a million-dollar exposure on a 300-ton heat. Caster plugging averaging 25 events year-to-date, each costing 80 minutes of production plus $40,000 in tundish replacement. Quality disposition delays where flags go unreviewed until the next day. And at the end of the chain, a six-person shipping team managing 220,000+ tons per month on a 1980s IMS system — the GM's number one priority.
Dave Holter may be the best-positioned champion we met across all four sites. Twenty years of chemistry data, in-house predictive models, and a clear ask: "Start with something super simple." Five percent of heats are off-chemistry, and the PA team confirmed SQL data going back to 2001 that is, in their words, "pretty kind of ready to feed into an LLM."
Cleveland showed us the broken system. Middletown showed us the fragile one. Tilden proved the thesis crosses industries. Burns Harbor showed us the full chain — and what is at stake if the organization moves too slowly.
One Pattern, One Architecture¶
The structural insight from four sites and 175 initiatives is this: the systems are different at every plant, but the information flow failures are identical. Three different CMMS platforms, three different ERPs, different historians, different process control systems — and yet every site independently surfaced the same diagnosis.
Andrew Mullen, CLF's AI and Innovation coordinator, confirmed it at the corporate level after visiting multiple sites: "It doesn't matter what CMMS — it's not getting done because it's too cumbersome." The problem is not which system a site uses. The problem is that no system connects to the others.
This is why the roadmap works the same way everywhere. Every project we recommend closes an information loop. The first projects close the most fundamental loops — connecting operations to maintenance, capturing knowledge before it retires, linking quality data across the process chain. As those connections are made, a progressive data foundation emerges. Not because we asked anyone to fund a data lake, but because each project contributed its piece.
| Site | Thesis Manifestation | Unique Asset |
|---|---|---|
| Cleveland | System is visibly broken — reactive maintenance, zero close-the-loop | 1SP bottleneck, 10-year Pi historian, slab optimization proof point |
| Middletown | People ARE the system — best performance, most fragile | Longest finishing chain, R&D momentum, Ametek classifier opportunity |
| Tilden | Data exists but does not flow — mine-to-concentrator gap | Only mine, $50M reagent spend, grassroots champions, G2 control layer |
| Burns Harbor | Longest chain amplifies everything — shipping velocity is the symptom | BF integration proof, coke plant knowledge cliff, plate mill uniqueness |
Throughput is the constraint. Reliability is the lever. Information flow is the unlock. AI is the mechanism.