Est.

No Tech Strategy No Problem How Extractive Industry Sites Fix Operations With AI Anyway

Start with one painful process, then redesign it around AI agents.

· 9 min read
Cover illustration for “No Tech Strategy No Problem How Extractive Industry Sites Fix Operations With AI Anyway”
AI-First Process Redesign · November 7, 2026 · 9 min read · 1,962 words

Extractive industry sites can use AI to fix operations without first buying a platform or writing an enterprise technology strategy. Start with one painful process, audit it as it actually runs, and redesign it with AI agents inside the decision loop.

Why extractive industry sites are deploying AI without a formal strategy in place

Shrinking margins, supply chains that can change without warning, changes in ore composition, and a shortage of skilled workers put extraction sites under pressure before a digital roadmap is ready. BCG’s January 2026 mining and metals perspective identifies these pressures, together with high capital intensity and more frequent commodity cycles, as reasons AI adoption has become a necessity. A site cannot defer an equipment availability problem until its enterprise architecture is settled. The work still has to run. That urgency explains why AI is moving from an innovation initiative into core operations rather than remaining the preserve of digital leaders.

Skepticism about that shift is nevertheless justified. BCG describes mining and metals leaders who have watched earlier technology efforts overpromise and underdeliver. Their concern is not simply resistance to change, because an AI deployment can reproduce the same disappointing pattern when the underlying work remains untouched. Fragmented systems, paper-and-spreadsheet workflows, and inconsistent data do not become coherent when a new tool reads them. The tool can instead accelerate the dysfunction.

The error, then, is acting before understanding the process. A site under operational pressure has a sound reason to move quickly, but speed should sharpen the choice of what needs fixing while preserving diagnosis. The site must identify what needs fixing before AI becomes part of the work.

How do you start AI without a technology strategy? Choose one painful process

The instinct to write a strategy first is rational. Asset-heavy operations have safety obligations, legacy systems, and capital constraints that make disconnected technology decisions costly. A strategy appears to offer order before commitment. However, making it the entry requirement shifts attention from a concrete operational loss to questions about platforms, architecture, and a future transformation roadmap. Those questions can consume the time in which a narrowly defined process could already be improving.

BCG recommends starting with real business problems where changes can deliver value, with operations, finance, and HR leading the work. Roughly seventy percent depends on people and processes, twenty percent on data, and ten percent on technology. Technology is the smallest lever in that account. Even a technically sound platform cannot make up for an unclear decision, a handoff that serves no purpose, or a workflow that fails to fit conditions in the field.

Start with a process that hurts enough for its improvement to have clear operational value. Maintenance scheduling is a more useful starting subject than an ambition to become an AI-enabled mine, because the team can examine how scheduling works and where it fails. The sequence is to identify the pain, audit the current work, and decide what needs redesigning. Tool selection follows that diagnosis. An enterprise AI strategy can then develop from the requirements of real operations.

What "redesign from scratch" actually means in a mine or extraction site

Redesigning from scratch means changing the workflow around the decision while retaining established technology at the site. Predictive maintenance, process optimization, and computer-vision-based safety systems are already part of mining practice. They are a base to build on. Redesign how information from the operating environment becomes a diagnosis, a proposed response, and an action within defined authority. Adding another report after those decisions have been made keeps the same decision loop in place.

BCG’s agentic maintenance example makes the distinction concrete. It describes a process built around three agents, one that diagnoses issues, one that finds the root cause, and one that makes recommendations. The agents operate within a new closed-loop maintenance process rather than serving as a reporting layer above the old one. This changes the relationship between analysis and the maintenance decision. The work is reorganized so that diagnosis and recommendation are part of the process itself.

Triage agents filter and route exceptions, while planner agents propose action sequences for human approval. Narrator agents surface the operating state across disconnected systems and take over both data entry and the work people do to coordinate decisions. In extraction operations, that distinction directs attention toward how maintenance and planning teams resolve exceptions.

Under a graduated autonomy model, a maintenance agent can be authorized to adjust soft parameters but never hard safety limits. Decisions outside its defined thresholds are escalated. Human approval is therefore part of the design where the proposed action requires it. Redesign places agents inside operational work without giving them unrestricted authority over it.

Where extraction site processes actually break — and why that is where to start

The most valuable starting point is not always the most visible failure. A problem described as scheduling or procurement may originate in the coordination between people and systems. The diagnostic task is to locate that fracture. For a site choosing its first process, the relevant distinction is between a task that takes too long and a decision that cannot be made reliably because the necessary information does not reach it.

If fleets are getting older and need more maintenance, which reduces equipment availability, maintenance scheduling is a good place to start the redesign. The examination should cover what appears on a maintenance dashboard and how equipment information enters operational technology systems and planning workflows. A dashboard can describe the problem without changing the scheduling decision. The process to redesign is the path between the equipment’s condition and the work planned for it. That path includes the work people do to coordinate a maintenance response based on the information.

If the difficulty is disruption spreading across demand, supply, logistics, and inventory, operational planning is the more appropriate starting point. BCG points to AI-enabled planning as a way to deal with resource nationalism, tariffs, and unstable supply chains. Here, the audit should examine where a change in one part of the plan has to be reconciled manually with the others. Manual coordination struggles when disruptions cascade. Treating each affected function separately can leave the connecting work untouched.

A harder finding is a mismatch between a declared schedule or maintenance log and actual field conditions. If records are kept mainly to meet reporting requirements, the workflow that creates and uses them should be redesigned first. Remote-site connectivity, legacy systems, cybersecurity, and workforce readiness can complicate that work, so they belong in the diagnosis. But the central question remains operational. Does the information describe the state of the site well enough to support the next decision?

Why data quality is the real constraint — and why it is not a reason to wait

Poor data is a genuine constraint. A process cannot make reliable decisions from records that are inconsistent or stripped of their operating context. Extraction sites also depend on knowledge held by experienced workers that never enters a system at all. A clean pilot dataset does not resolve those gaps. Using the data at the site tests whether it means what the process assumes it means.

BCG’s answer is that data foundations must be “just good enough.” Its guidance is to start with usable data, deliver quick impact, and improve quality iteratively. This changes the scope of the first deployment from solving the site’s entire data problem to finding a process where available information can support a better decision than the current manual step. That is a narrower and more operationally useful test.

The reverse approach commits capital to data engineering before proving that a live process benefits. When results are then demanded, the organization has an infrastructure program but no operating improvement to demonstrate. More data is not necessarily the missing ingredient. Existing information must be consistent and understood in context to support the decision being redesigned. The first agent’s scope should reflect those limitations rather than assume that a broader data foundation will eventually remove them.

Where a decision depends on an experienced worker’s understanding, an agent can be built to capture and structure that knowledge as a byproduct of the workflow. This addresses a gap that a conventional data cleanup cannot reach, because the information was never recorded. Deployment and data improvement therefore proceed together. The output of the first agent can improve the data available for the next process in scope, making each deployment a basis for the next iteration.

How fast a redesigned process can reach production — and what slow timelines reveal

A well-scoped first process redesign should target production in weeks. That is a scoping discipline. The first process must be small enough to use the data on hand and work within set permissions and limits on what needs approval. Production is more than a demonstration. The process has to function as part of the site’s work, including the limits that make it safe to deploy.

BCG reports on a mining company whose large-scale AI program changed how it handled mine-to-port end-to-end scheduling and achieved ROI in three months. Other sites’ outcomes may differ from this reported case. The payback timeline nevertheless requires useful operating work to have begun before the return was achieved. It offers a concrete counterpoint to the assumption that operational value must wait for a prolonged platform program.

If a program reaches month six without a live process, the team needs to review how it is going about the work. The scope may have expanded into a platform strategy, data foundations may have become a prerequisite, or the team may be automating work it has not audited. More elapsed time does not correct those choices. BCG includes focused minimum viable products as a step in the transformation, which supports starting with a small process that proves its value before the work expands.

The process audit as the actual starting point — what it looks like in an extraction context

The process audit begins the redesign rather than preparing for a later technology exercise. Its subject is one painful workflow as the operating team actually performs it, including the differences between documented procedure and field practice. A maintenance log, for example, is not sufficient evidence of how maintenance decisions happen. The audit must establish where the record reflects conditions and where it serves a reporting requirement instead. That distinction determines what an agent would be acting on.

The same examination traces manual coordination between systems that do not communicate and decisions that depend on knowledge held by one or two people. These design inputs show which steps need a different information flow and which require knowledge to be captured within the work. BCG places identifying clear value points before strengthening foundations, developing minimum viable products, and scaling. The audit gives that sequence operational substance by connecting a specific failure to a specific redesign.

It also establishes the first agent’s authority. The audit defines what the agent can execute autonomously, what it must escalate, and what logs it must produce. Those boundaries put governance in place from the start so the agent can be deployed where safety is critical. BCG quotes a mining and metals CEO who describes how AI has had a modest impact where ore quality and mining conditions vary. Such variability is a reason to understand the operating process before choosing the tool.

The concrete first move is to bring in an operational partner rather than begin with a software vendor or a consulting firm. That partner’s role is to sit with the team running the painful workflow and map the work as it happens. The map should expose reporting gaps, manual handoffs, and decisions dependent on unrecorded knowledge. Before anything is built, the team and partner should identify the first process to redesign and define the agent’s decision boundaries within it.

Sources

  1. Executive Perspectives January 2026 Mining and Metals

More in AI-First Process Redesign