Executive Summary¶
The Thesis¶
Cleveland-Cliffs does not have an AI problem. It has an information flow problem that AI can solve.
Over ten weeks, across four sites and two industries, we conducted 40+ conversations with maintenance leaders, process engineers, plant managers, and frontline technicians. We found the same pattern at every site: 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 systems are different everywhere — three different CMMS platforms, three different ERPs, different historians, different process control systems. But the diagnosis is identical: data exists in islands, and no system connects them. The result is reactive maintenance, quality investigations that start from scratch, procurement bottlenecks, and production scheduling driven by institutional memory.
What We Did¶
| Dimension | Scope |
|---|---|
| Sites visited | Cleveland Works (OH), Middletown Works (OH), Tilden Mine (MI), Burns Harbor (IN) + Indiana Harbor validation |
| Conversations | 40+ structured sessions with 2-4 practitioners per session |
| Initiatives identified | 175 across 4 sites |
| Site projects defined | 61 (grouped from initiatives) |
| Corporate projects scored | 11 (cross-site programs) |
| Systems inventoried | 20+ platforms across 12 categories |
| Industries covered | Steelmaking (3 sites) and Iron Ore Mining (1 site) |
What We Found¶
Cleveland — the system is visibly broken. Maintenance runs 70% reactive. Zero close-the-loop between operations and maintenance. The 2001 ISG restructuring eliminated the process discipline layer and it was never rebuilt. But Phil Thorman's slab optimization proves the thesis: one information loop closed, $3M/month in returns.
Middletown — the system works because people carry it. The best steel shop in the company — near-zero unplanned turnarounds, scrap below 1%. But the knowledge is in people, not systems, and retirements are imminent. Three leaders independently called this "the biggest problem facing the plant."
Tilden — the thesis crosses industries. A $50M/year reagent spend with only 70% iron recovery — below the plant's design benchmark of ~75% and well below the 80% achievable with optimized hematite flotation — because there is no predictive link between what comes out of the ground and how the concentrator responds. Mining maintenance is strong (70% planned), but the information flow gap is in process chemistry.
Burns Harbor — the longest chain amplifies everything. CLF's largest site showed us both the proof (BF department built integrated information systems with zero budget) and the warning (Indiana Harbor: 282 unread PdM emails, 4 radio channels, no shared information).
What We Recommend¶
A 24-month program of 10 projects organized in three progressive horizons, with milestone gates at Month 6 and Month 12. Total investment: $22.6M. Annual returns at full deployment: $103-344M depending on scenario.
Horizon 1 (Months 0-6): Bridge the Gap — $3.5M. Close the fundamental information loops. Ops-maintenance integration at Cleveland. Knowledge capture at Tilden and Burns Harbor. PdM proof of value at Cleveland (active — contract in final approval). Procurement automation at Middletown. Concentrator desliming optimization at Tilden (active — charter submitted, $312K Phase 1). Plate shipping automation at Burns Harbor. Quick wins visible in 4-8 weeks.
Horizon 2 (Months 7-12): Build the Foundation — $11.6M. Scale H1 projects across sites. Launch quality optimization (Middletown Ametek classifier), logistics (Burns Harbor shipping velocity, Palmer's #1 priority), BF process intelligence, and caster chemistry optimization. Cross-site data flowing.
Horizon 3 (Months 13-24): Expand and Optimize — $7.5M. Cross-site platforms, corporate dashboards, advanced optimization. Resource-intensive but built on proven foundation, established team, and existing data pipes from H1 and H2.
10 projects qualify for the roadmap. They are sequenced by evidence: validated at multiple sites, endorsed by executive leadership, grounded in specific operational metrics, and structured with milestone gates and off-ramp criteria.
Three principles govern every project:
- Solve pain first, build foundation as a byproduct. We are not asking CLF to invest in infrastructure and wait for returns. Every project solves an immediate operational problem. The data foundation emerges from the connections each project makes. The Databricks platform proceeds in parallel — our architecture runs on it from Day 1.
- Operators decide when to trust. Every project follows the same maturity path: advisory, then semi-automated, then closed-loop. The gate between stages is operator acceptance, not model accuracy. A model that scores 95% precision but that operators do not trust stays in advisory mode.
- Backtested before deployed. Every model is validated against incidents CLF already knows about before it touches production. Every recommendation exposes its reasoning so operators can verify before they trust. Vooban handles the AI; IE handles plant-floor integration and change management.
What It Costs and What It Returns¶
| Conservative | Base | Optimistic | |
|---|---|---|---|
| Total investment (24 months) | $22.6M | $22.6M | $22.6M |
| Annual returns (full deployment) | $103M/yr | $224M/yr | $344M/yr |
| ROI multiple | 4.6x | 9.9x | 15.2x |
| Payback period | 10-14 months | 7-10 months | 5-7 months |
The investment is deterministic — bottom-up from team composition, anchored to two charter actuals ($312K Tilden, $236K Cleveland PdM). The value ranges reflect three scenarios: conservative (30-50% capture rates, slow adoption), base (50-70% capture, normal execution), and optimistic (70-90% capture, full executive backing).
The program self-funds: Horizon 1 quick wins generate returns by Month 3-6 that offset Horizon 2 investment. In the base scenario, cumulative returns exceed cumulative investment by Month 8-9. At each milestone gate (Month 6 and Month 12), corporate evaluates evidence and decides to continue, redirect, pause, or expand. You are committing to six months at a time. Once you're through Gate 2, the program runs on its own momentum.
The maximum at-risk amount at any point is the current phase investment, not the full $22.6M.
Already Underway¶
Two projects have been scoped, chartered, and priced — direct responses to Chad Asgaard's directives:
- Cleveland PdM PoV (PRJ-03): Charter submitted, contract in final approval. Eight-week multi-asset PoV at 1SP covering BOF bag house, scrubbing system, and Crane 300, plus a data readiness assessment across all reachable 1SP equipment. $236K investment. Separate SOW.
- Tilden Concentrator Optimization (PRJ-10): Charter v3 submitted, Phase 1 priced at $312K firm. Seven-week Proof of Value proving the core thesis: do existing process signals correlate with recovery outcomes strongly enough to justify the full $1.3-1.56M system build? Four-component AI stack around the desliming circuit. Erico Lemos managing as project lead.
Combined: $548K already committed, with $1.3-1.56M indicative for PRJ-10 Phase 2 pending Phase 1 results.
The remaining Horizon 1 projects launch at program approval:
- Ops-Maint Integration (PRJ-01): $540K — misattribution analysis at Cleveland in 4-6 weeks
- Knowledge Capture (PRJ-09): $684K — HPGR knowledge base at Tilden + coke plant capture at Burns Harbor
- Procurement Automation (PRJ-06): $612K — fast-track rules + inventory deduplication at Middletown
- Logistics (PRJ-07): $684K — plate hit list at Burns Harbor + coil planning at Middletown
- Coke Plant (PRJ-11): $390K — data consolidation + tribal knowledge capture before retirements
H1 total: $3.5M for 7 working proof-of-value projects, each producing tangible results within 8-10 weeks.
The business case, project charters, and resource requirements for each are detailed in the chapters that follow. The pre-read package provides full project definitions (Chapter 5), methodology transparency (Chapter 6), and portfolio economics (Chapter 7). Site-specific roadmaps are provided as a companion volume.
The constraint is throughput. The lever is reliability. This program makes information flow where decisions get made.