The Steel Industry's AI Moment Has Three Speeds — And Business Teams Are Stuck in First Gear
There is a framework making the rounds in AI startup circles that I think every steel industry leader needs to hear.
Jake Heller, in his video (see below) puts it simply: every AI-enabled process falls into one of three categories. Assist. Replace. Do the Unthinkable.
His point, made to a room full of founders at Y Combinator’s AI Startup School, was about product strategy. Pick the wrong category, and you build something people like but don’t depend on. Pick the right one, and you build something that changes what’s possible.
Inspired by his lecture, I want to take that framework, carry it into the middle of a large integrated steel plant… through the furnaces, the rolling mills, the quality labs, the dispatch yards… and make an uncomfortable argument:
Software vendors are still debating which tasks belong in which bucket. But the real bottleneck isn’t the AI. It’s that business teams haven’t yet decided to show up.
First, Let Me Make the Taxonomy Real for Steel
Abstract frameworks only earn their keep when they survive contact with industrial reality. So let me walk through what Assist, Replace, and Unthinkable actually look like across the eight major functional areas of an integrated steel plant.
Assist: AI Sharpens the Human in the Room
Assist is where every transformation starts. The AI doesn’t replace the expert; it makes the expert faster, better-informed, and harder to surprise.
In Sales, an order feasibility co-pilot checks mill schedules, inventory buffers, and raw material positions the moment a customer asks for a non-standard spec or a tight delivery window… so the sales engineer gives a real answer in the meeting, not two days later after chasing the planning teams. In Customer Service, an AI complaint triage tool reads the incoming claim, pulls the heat and coil genealogy, and presents the service representative with a ranked hypothesis tree before the first customer call even happens.
In Planning, AI synthesizes order book trends, market signals, and consumption patterns into a weekly 90-day demand signal… surfacing the data that helps planners see around corners. In Quality, an SPC monitoring assistant watches hundreds of process parameters simultaneously, spots emerging drift patterns, and presents likely root causes, while the QC engineer decides the response.
In Production, operators get real-time process parameter recommendations when a furnace temperature or mill rolling force starts deviating from the optimal corridor. In Scheduling, AI suggests optimal heat sequences to minimize grade transition losses, ladle turns, and tundish changeovers, and the scheduler approves, adjusts, and owns the decision.
Across Logistics, Warehousing, and Dispatch, Assist looks like smarter load planning, better yard allocation recommendations, and dispatch priority rankings that account for customer priority, crane availability, and transport lane status… all surfaced to a human who still makes the call.
Assist is valuable. It is deployable now. And it is deeply undersold in every vendor conversation I’ve observed.
Why? Because Assist requires domain knowledge to configure well. You can’t build a meaningful order feasibility co-pilot without someone from Sales and Planning sitting together and defining what “feasibility” actually means in your plant’s specific operating context. And that conversation, the business conversation, hasn’t happened meaningfully yet at most plants.
Replace: The Economics Are Hiding in Plain Sight
Replace is where the real cost story lives. These are tasks that don’t require human judgment; they require human time. And they consume enormous quantities of it.
Consider the mill test certificate. Every coil, every heat, every shipment requires one. A quality engineer manually cross-references test results against customer spec sheets, populates the certificate, checks for values near tolerance limits, and releases the document. Multiply this by the daily shipment volume of a large integrated plant. That is not judgment work. That is transcription work, and AI can do it, at any volume, without fatigue, without Monday morning errors or a week-end syndrome.
The same logic applies to rolling schedule calculation in hot strip mills. For every incoming slab, a pass schedule needs to be computed… reductions per pass, speeds, tensions. This was done manually, or with rigid rule-based tables. AI can calculate it dynamically, accounting for actual slab chemistry and thermal state, every time.
In Logistics, standard carrier dispatch… selecting carrier, issuing transport order, generating e-way bill, triggering customer notification… is a defined process over structured data. There is no creative judgment in it. AI can run the entire sequence for standard shipments, with exceptions escalating to humans.
In Warehousing, coil inventory reconciliation currently requires periodic manual stock-reconciliation (it is a real-pain…monthly…) because crane movement logs, RFID reads, and ERP records drift apart. AI can continuously reconcile these streams and maintain an accurate real-time inventory without the stocktake shutdown.
In Planning, the rolling horizon plan refresh... recalculating the two-week production schedule every shift to account for actual yields, machine availability, and order changes, is currently a weekly meeting ritual. AI can run it continuously, publishing an updated plan without waiting for a consensus meeting to convene.
None of these are exotic. None require breakthroughs. Indeed.
The technology to replace these tasks exists today. The gap is not capability. The gap is that no one has sat down with the business team and reverse-engineered the decision tree, the exception conditions, and the quality standards that govern each of these tasks. That work is not a vendor’s job. It is a business team’s job.
Do the Unthinkable: Where the Moats Are Built
This is the category that gets me out of bed in the morning.
The steel value chain has a structural characteristic that makes it uniquely suited for the Unthinkable category: massive, deeply interdependent combinatorial state. The number of variables simultaneously influencing a single cast… ladle chemistry, furnace temperature profile, slab queue length, rolling mill condition, downstream coating line availability, energy cost curve, active order book… exceeds what any human team or conventional optimization software can hold and process together in real time.
This isn't a technology problem that better software will eventually solve. It's a problem of dimensionality. And AI is the first class of technology that can operate meaningfully in that space.
I want to be precise here: some of what follows is already deployed and documented at scale. Some are demonstrated in research and early pilots, moving toward production. I will be clear about which is which, because intellectual honesty is the only currency that matters in thought leadership.
Here is what Unthinkable looks like in each function:
Sales: Proactive order-to-production matching. Today, a sales team may only become aware of available open-order slabs after the planning team completes a batch analysis, often overnight or on a weekly cycle. AI could enable a continuous, live scan of the cast-to-roll pipeline against the open order book, identifying customers whose specification windows could be satisfied by heats already in progress, before those heats are committed elsewhere or downgraded.
This could be viewed as a natural evolution of the slab-to-order matching optimization problem. The underlying combinatorial optimization capabilities already exist. What could be missing in many plants is the organizational integration required to connect the sales order book with the production pipeline in real time, and the decision-making framework that empowers teams to act on those insights.
In that sense, the primary challenge may not be technological. It could be a business design challenge… aligning processes, responsibilities, and incentives so that real-time production intelligence can translate into commercial opportunity.
Customer Service: Contain suspect material before it ships; communicate before customers complain. AI systems deployed in steel plants today can already connect predictive maintenance alerts with the quality lots produced during an equipment degradation window, enabling potentially affected material to be identified and inspected before dispatch. Building on this capability, the next evolution could be the automatic notification of customers before the material ships, transforming what might otherwise become a future complaint into a proactive and collaborative discussion around deviations and mitigation options.
In most plants today, this remains a human-mediated process. AI may surface the signal, but a customer service representative typically decides when and how to communicate it. Over time, organizations could move toward fully automated proactive customer alerts, where validated quality risks trigger timely notifications and recommended actions without waiting for manual intervention.
The technology to generate these insights is increasingly available. What could determine the pace of adoption is organizational readiness… the confidence, governance, and decision-making processes required to act on such signals quickly, consistently, and transparently.
Planning: Simultaneous cross-process campaign synchronization. A blast furnace maintenance window could influence BOF tap-to-tap cycle times, which in turn could shift continuous caster sequences, alter hot rolling mill slab queues, and propagate downstream into cold rolling and coating operations, and beyond. This type of cascading impact is not merely theoretical; it reflects the interconnected nature of integrated steel plants, where a disruption at one stage can potentially affect multiple processes further downstream.
AI-enabled scheduling platforms are increasingly demonstrating the ability to model and manage these interdependencies within a single facility in near real time, helping planners evaluate alternative scenarios and respond more effectively to disruptions. Looking ahead, the next frontier could be simultaneous optimization across multiple plant locations, balancing production constraints, grade transition costs, inventory positions, and delivery commitments across an entire manufacturing network.
Some of these capabilities are already being explored through advanced research programs and early-stage pilot implementations. They represent a likely direction for the evolution of APS and MES platforms. However, truly network-wide, multi-site optimization remains an emerging capability rather than a mature, off-the-shelf solution that can be readily purchased and deployed today.
Scheduling: Continuous slab-to-order dynamic matching. One of the most compelling opportunities could emerge from the slab design and order-matching problem, a well-known combinatorial challenge in steel operations. Traditionally, solving this problem often required overnight batch processing. By the time a matching solution was generated, the production conditions it was based on could already have changed, limiting its practical value.
Advances in constraint programming, optimization algorithms, and AI-enabled solvers are making it increasingly feasible to solve large, real-world matching problems involving hundreds of orders within seconds. Building on these capabilities, AI could enable a continuously running optimization process that evolves alongside plant operations. Every new heat tapped, every incoming order, and every yield deviation at the caster could potentially trigger a fresh optimization cycle, keeping recommendations aligned with the current production reality.
The significance of this shift may extend beyond faster computation. It could represent a transition from a periodic planning model to a continuously adaptive operating model, where decisions are updated dynamically as conditions change. Rather than simply improving the efficiency of existing processes, such capabilities could fundamentally reshape how production planning, order fulfillment, and material allocation are managed across the steel value chain.
Scheduling: Disruption-resilient rescheduling. When a rolling mill stand experiences an unplanned failure, the impact can extend far beyond the maintenance team. A single stand outage could potentially halt hot strip mill production for several hours or even days, creating significant production losses and triggering a chain reaction across planning, logistics, and customer-facing functions. In many plants today, the response may still rely heavily on manual coordination… planners revising schedules, operations teams assessing downstream impacts, and sales teams responding to customer inquiries while critical information is still being assembled.
AI-enabled scheduling systems could fundamentally improve this response process. Advanced scheduling platforms are already capable of generating revised production schedules within minutes of a disruption, evaluating alternative recovery scenarios, and assessing the implications for production commitments and delivery dates. Instead of waiting hours for a coordinated response, decision-makers could gain near real-time visibility into the operational and commercial consequences of the event.
Looking ahead, the greatest opportunity may lie not only in faster schedule recovery but also in proactive stakeholder communication. AI-driven systems could potentially provide continuous updates on recovery progress, revised delivery expectations, and customer impacts while repairs are still underway. The ability to combine rapid operational replanning with timely, transparent customer communication could become a defining characteristic of the most resilient and customer-centric steel producers.
Production: Metallurgical state estimation from process signals. Today, grain size measurement in steel production is typically a destructive and laboratory-based process. Samples are sectioned, prepared, analyzed, and the results are often available only after a delay that may exceed the transit time of material through certain continuous processing lines. As a result, grain size generally cannot be incorporated into real-time process control or in-line quality specifications in most steel manufacturing environments.
Recent advances in AI and machine learning could change this landscape. Research has demonstrated that convolutional neural networks and other machine learning models can classify steel microstructures and estimate grain size from metallurgical images with a high degree of accuracy. These developments suggest the possibility of moving beyond traditional laboratory measurements toward more responsive and data-driven quality assessment approaches.
Looking further ahead, a particularly promising direction could be the inference of metallurgical state directly from process signals generated during rolling, annealing, or other production operations—without requiring destructive sampling. While this remains an active area of research rather than a broadly deployed industrial capability, the concept is technically credible and increasingly attracting attention from both researchers and industry practitioners.
If realized, such capabilities could enable entirely new forms of closed-loop metallurgical control, allowing process parameters to be adjusted in real time based on predicted microstructural outcomes. This would represent more than an incremental improvement in quality assurance; it could create operational possibilities that are difficult or impossible to achieve with current measurement methods. For business and technology leaders, the opportunity may be less about waiting for a finished commercial product and more about monitoring emerging developments that could redefine quality control and process optimization over the coming decade.
Quality: Predictive quality twin and yield loss root cause discovery. One of the most promising applications of AI in steel manufacturing could be its ability to uncover operational and quality improvement opportunities that remain hidden within vast volumes of plant data. Digital twin technologies and advanced analytics have already demonstrated the potential to identify previously unnoticed inefficiencies and process behaviors, revealing opportunities that may not be apparent through conventional analysis methods alone.
Similarly, AI models are increasingly showing the capability to predict downstream quality outcomes earlier in the production process. Rather than waiting for defects to become visible through inspection or testing, future systems could identify emerging quality risks based on process conditions and operating patterns, enabling corrective actions before additional value is added to the material.
Another area of significant potential lies in the analysis of historical operational data. By examining years of production records, laboratory results, maintenance histories, and customer feedback, AI could identify complex interactions among process variables that contribute to quality deviations or performance losses. These relationships may involve combinations of factors that are difficult for even experienced analysts to detect because of the scale, dimensionality, and variability of the data involved.
The value proposition is not that AI already knows where the next breakthrough opportunity exists within a specific plant. Rather, it is that AI provides a practical mechanism for discovering insights that would otherwise remain buried in the data. The underlying principle has been demonstrated across multiple industrial applications; however, the specific opportunities, correlations, and improvement levers that emerge are unique to each operation and can only be revealed through systematic analysis of that plant’s own data. In that sense, the most valuable insights may be the ones that nobody is currently looking for.
Logistics: AI-powered Capable-to-Promise. Traditional Available-to-Promise (ATP) processes in many steel organizations are primarily based on inventory visibility and open order commitments. While effective for certain scenarios, they may not fully account for the broader operational context, such as production status, yard inventory availability, logistics constraints, transportation capacity, and transit times at the moment a customer inquiry is received.
Advances in AI, optimization, and integrated planning systems could enable a more sophisticated Capable-to-Promise (CTP) approach. In such a model, customer commitments could be evaluated against live operational conditions across the value chain. When a customer requests a specific quantity, grade, and delivery date, the response could be generated using real-time information from production, inventory, logistics, and scheduling systems rather than relying solely on periodic planning snapshots.
This vision aligns with the direction in which many advanced planning and scheduling platforms are evolving. The objective is not simply to provide faster answers, but to provide answers that are dynamically aligned with current plant capabilities and constraints. Such a capability could significantly improve confidence in delivery commitments while reducing the need for manual coordination across planning, operations, and logistics teams.
In many plants today, responding to a complex customer request may still require multiple discussions across departments before a reliable commitment can be made. The opportunity lies in narrowing that gap between customer expectation and operational visibility. AI-enabled Capable-to-Promise systems could transform delivery commitments from a largely manual, reactive process into a data-driven, real-time business capability that improves both customer responsiveness and operational performance.
Warehousing and Dispatch: A self-improving delivery system. Sequence-aware yard orchestration could become one of the most impactful applications of AI in steel logistics and material handling. In such a model, the physical location of every coil, slab, or finished product could be continuously optimized based on evolving production schedules, dispatch priorities, customer commitments, and transportation plans. Because the number of possible storage and retrieval decisions changes constantly throughout the day, this represents a highly dynamic optimization challenge that can quickly exceed the practical limits of manual yard management.
AI-enabled systems could evaluate thousands of possible relocation, stacking, and retrieval scenarios in real time, helping ensure that material is positioned not only for current requirements but also for anticipated future movements. The result could be reduced reshuffling, lower handling costs, improved equipment utilization, and faster dispatch performance.
An equally important opportunity lies in closing the feedback loop between planning and execution. Future systems could continuously compare planned delivery commitments against actual shipment, transportation, and customer receipt performance. These outcomes could then be fed back into planning, scheduling, and logistics models, enabling them to learn from operational reality and improve future decisions automatically.
Taken together, these capabilities could transform logistics from a largely reactive function into a continuously learning system. Rather than simply executing plans, the system could adapt based on observed outcomes, refining its recommendations over time. This would represent a significant departure from traditional industrial operating models, where planning, execution, and performance analysis are often conducted as separate activities. AI creates the possibility of connecting these functions into a single adaptive loop, allowing the delivery system itself to evolve and improve with every production cycle and customer shipment.
Some of this is here. Some of this is arriving. None of it is waiting for better AI. What it is waiting for is business teams willing to engage at the level of specificity these problems require.
The Uncomfortable Truth About Who Is Stalling This
Here is where I want to be direct, because I have been in enough rooms on both sides of this conversation to know where the friction lives.
Software vendors are circling these problems. Every major MES, APS, and ERP vendor is running AI pilots, publishing white papers, and announcing roadmap features. Some are building genuinely interesting things. But most of the conversation is still happening at the technology layer: models, integrations, APIs, platforms.
And business teams, by and large, are watching. Many times on the fence.
Not because they are uninterested. Not because they don’t see the potential. But because they are waiting for something that looks finished, proven, and safe to implement… a shrink-wrapped solution they can evaluate and buy the way they’ve always bought software.
Honestly, that is not how this works.
The Assist category requires business teams to define - precisely and specifically, what “assisting” means in their operational context. What data does a sales engineer actually need before a customer call? What does the quality engineer check before releasing a certificate? What does “feasibility” mean for a non-standard order, and who gets to say no? These definitions cannot be extracted from a vendor demo. They live in the heads of experienced practitioners, and they need to be externalized, debated, and translated into logic that AI can act on.
The Replace category requires business teams to do something even harder: map their own processes honestly enough to identify where human judgment is genuinely required and where it is merely habitual. Many tasks that “require a human” only require a human because no one has ever built a system good enough to do them without one. AI changes that, but only if the business team is willing to interrogate the assumption.
The Unthinkable category requires business teams to do something most industrial organizations have never done at all: articulate problems they don’t currently solve, not just problems they want to solve faster. What decisions do you currently avoid because the information required to make them doesn’t arrive in time? What optimizations do you approximate because the real calculation is too complex? What value do you leave on the table because connecting one part of the plant to another part of the order book is beyond current operational bandwidth?
Those questions… the Unthinkable questions, will not be answered in a vendor evaluation. They will be answered in workshops, in cross-functional conversations, in moments when a production planner and a sales engineer and a metallurgist sit in the same room and realize they are each holding a different piece of a problem that AI could solve if only someone thought to combine their perspectives.
What Business Teams Need to Do Differently… Starting Now
This is not a call for wholesale digital transformation programs with five-year timelines and nine-figure budgets. It is a call for a different posture in the conversations that are already happening.
Own the problem definition. When a vendor shows you an AI demo, don’t ask “can it (also) do this?” Don’t settle down with the jazzy neon-lit dashboards. Ask “what would I need to be true about my operations for this to generate real value?” Then go find out whether those conditions exist.
Classify your own work. Take your function: Sales, Planning, Quality, whichever, and spend two hours with your team listing every recurring task. Then sort them: Assist, Replace, or Unthinkable. You will find that the list is longer than you expected and that many Replace candidates have been treated as sacred human tasks for no good reason.
Build bridges across functions. The highest-value AI use cases in steel… the Unthinkable ones… all require data and decisions that currently live in silos. The yield optimization problem connects steelmaking to rolling to quality to customer service. The delivery promise problem connects production to warehousing to logistics to sales. The people who need to co-design these solutions are not all in the same department. Creating the cross-functional conversation is a business leadership responsibility, not a technology one.
Pilot with intent. Don’t run AI pilots to learn whether AI works. Run them to learn what your operations would need to look like to let AI do more. A good pilot generates organizational learning, not just a proof of concept.
Insist on feedback loops. Any AI system your organization deploys should get smarter over time from your operational data. If the vendor cannot explain how the model improves through use, you are buying a static tool with AI branding. The compounding value of AI comes from the learning loop - and that loop requires your data, your feedback, and your commitment to close it.
The Window Is Real, But It Won’t Stay Open
The steel industry is not early to AI. It is not late either. It is, right now, at the inflection point where the gap between organizations that engage seriously and those that watch thoughtfully is about to widen into a chasm.
The vendors will catch up. The tools will mature. The integration patterns will standardize. When that happens, technology will no longer be the differentiator, because everyone will have access to roughly the same technology.
What will differentiate the leaders is organizational capability: the ability to define problems precisely, to configure AI systems for their specific operational context, to build the cross-functional bridges that the highest-value use cases require, and to sustain the learning loops that make AI systems compound over time.
That capability is not something a vendor can install. It is something business teams build: by deciding, now, to be active participants in this transition rather than its audience.
Assist, Replace, Unthinkable. Pick your starting point. Map your function. Find the conversations that haven’t happened yet.
The technology is ready when you are.
Disclaimer: The views expressed in this article are personal in nature, for informational purposes only, and do not constitute professional, operational, or financial advice for specific manufacturing facilities.
#ManufacturingAI #SteelIndustry #AIStrategy #DigitalTransformation #ManufacturingIntelligence #AIMSquare #IndustrialAI #IntegratedSteelPlant #MES #APS