Your organization spent $4 million on an AI initiative last year. You evaluated vendors, selected a platform, signed a license, and deployed it. The demo was impressive. The pilot showed promise.

However, the initiative failed to deliver meaningful results.

This experience is common. MIT found that 95% of enterprise AI pilots produced no measurable financial impact. Corporate AI investment reached $252 billion in 2024 and grew rapidly in 2025. However, S&P Global reported that 42% of companies discontinued most of their AI initiatives, up from 17% the previous year.

The common explanation is that AI is difficult to implement. While this is accurate, it does not address the underlying issues. The real question is what specifically is causing these challenges, and the answer is often overlooked.

The 93/7 Problem

Here's the stat that should be on every executive's desk:

93% of enterprise AI budgets go to technology. 7% goes to people and processes.

Meanwhile, BCG research shows that 70% of AI project success depends on the organizational layer - training, change management, governance, and workflow redesign.

This means that the majority of investment is allocated to the area that contributes least to project success.

This is not a technology issue, but a procurement one. Organizations are approaching AI purchases as they would traditional software: evaluating features, selecting vendors, signing contracts, and deploying. While this approach is effective for SaaS tools, it is ineffective for AI.

This is evident with enterprise AI subscriptions. Organizations using Microsoft's ecosystem often assumed they could add Copilot licenses as easily as Office 365 seats. While this approach was convenient, it led to critical questions being overlooked, such as which workflows, users, success metrics, and support models were needed. Months later, many organizations reported poor outcomes, not due to technology limitations, but because success criteria were never defined. They purchased an AI subscription when they required a comprehensive AI strategy.

This pattern is widespread. The procurement model assumes AI is a finished product, but deployment is only the beginning of the process.

Software is a product: it is purchased, configured, and performs its intended function. AI, by contrast, is a capability that requires ongoing development, staffing, and continuous improvement. The distinction is significant; it is comparable to the difference between purchasing a pump and commissioning an entire plant.

You Bought a Pump. You Need a Plant.

Professionals in energy or heavy industry are familiar with this distinction, though it may not yet have been applied to AI.

Purchasing software is similar to buying a pump: it is specified, procured, installed, and maintained by the vendor. The pump does not adapt to your process or improve over time; it simply performs its function.

Developing AI capability is comparable to commissioning a new unit. This involves designing processes, tuning control systems, training operators, managing unique edge cases, and making ongoing adjustments as conditions evolve. Vendors cannot perform this work, as they lack knowledge of your specific processes. The true value lies in the operational expertise that enables the equipment to function effectively for your application.

When organizations purchase AI as they would a pump, the result is often underutilized technology that has not been tailored to their processes.

Consider another example: hard hats can be purchased as products, but a safety culture must be developed through ongoing training, incident reporting, accountability, and sustained leadership commitment. This is well understood in industrial operations.

AI operates similarly. While a model can be purchased, organizational capability must be developed to realize value. This requires domain experts, feedback systems that incorporate learning, measurement frameworks to track progress, and leadership that treats AI as an operational discipline rather than an IT expense.

I Learned This the Expensive Way

I've spent years training AI models to read engineering drawings - P&IDs dense with ISA symbols, decades of redlines, and company-specific conventions that no published standard covers.

Here's what the work actually looked like:

The model training script consisted of 50 lines, and each training run took several hours. However, the model represented only 20% of the total effort. The remaining 80% involved tasks often overlooked by procurement models:

  • Building seven custom review interfaces so domain experts could efficiently audit AI predictions
  • Designing data curation pipelines that caught a single mislabeled source, corrupting 5,000 training samples
  • Creating cross-validation frameworks after discovering that two test sets of different difficulty made a worse model appear 6% better
  • Iterating through nine training rounds, where each round's human corrections became the next round's training data

None of these activities were technology expenditures. They represented investments in capability: people interpreting data, processes systematically identifying errors, and institutional knowledge being integrated with each iteration.

As a result, detection accuracy improved from 39% with a generic API to over 93% with fine-tuned local models. Tasks that previously required months of manual effort were completed in weeks. These outcomes were achieved not by selecting the right model, but by developing organizational capability to curate data, evaluate results, and drive continuous improvement.

Each expert correction accelerated subsequent deployments, and every identified edge case enhanced the system's robustness. The model processing the twentieth facility was significantly improved compared to the first, not due to architectural changes, but because organizational capability had compounded.

This exemplifies what it means to treat AI as a capability, which cannot be acquired through a software license.

What the 6% Do Differently

McKinsey found that only 6% of companies are considered true AI high performers. Their success is attributed not to technology, but to their operating model.

The successful ones follow what researchers call the 10/20/70 rule:

  • 10% of AI investment goes to algorithms and models
  • 20% goes to technology and data infrastructure
  • 70% goes to people and processes

This contrasts sharply with the industry average allocation of 93% to technology and 7% to people and processes.

Consistent evidence shows that projects with sustained CEO involvement achieve a 68% success rate, compared to 11% for those lacking sponsorship. Additionally, 31% of workers actively resist AI initiatives - refusing tools, providing poor data, or delaying projects - when they are not included in the process.

Notably, 61% of enterprise AI projects were approved based on projected value that was never formally measured after deployment. Organizations are purchasing AI without assessing its effectiveness, then continuing to invest.

In oil and gas, about two-thirds of executives are still running pilots without achieving the outcomes they aspire to. Only about a quarter have progressed to scaling with measurable impact.

"But Will AI Replace Me?"

It is important to address a common concern among professionals.

The fear is understandable. Headlines scream about mass displacement. Entry-level job postings have declined sharply since 2023, with companies increasingly citing AI as the reason. The WEF projects 92 million jobs will be eliminated by 2030.

But here's what those headlines leave out: the same WEF report projects 170 million new roles created - a net gain of 78 million jobs. And the most striking finding comes from the Federal Reserve Bank of Dallas in February 2026: AI substitutes for workers with codified, textbook knowledge - but augments workers with tacit, experiential knowledge. Wages are actually rising in AI-exposed occupations that value hands-on experience.

For those with decades of experience, it is important to note that while AI can replicate textbook knowledge, it cannot replace the insights gained from years of commissioning, troubleshooting, and operating real facilities. Tacit knowledge, such as understanding the significance of a specific valve configuration, is not captured in training datasets or accessible to advanced models.

This expertise is essential for AI to deliver value. The models I trained achieved 93% accuracy, not due to superior algorithms, but because domain experts with extensive P&ID experience reviewed errors, provided corrections, and integrated those improvements into the system. Without this human expertise, the model remained at 39% accuracy and was ineffective.

It is understandable that experienced workers may be skeptical of the term "augmentation," as it is often associated with workforce reductions. The key issue is not whether AI can theoretically augment expertise, but whether organizations design AI systems that rely on, reward, and preserve this expertise before it is lost.

AI does not need to replace experienced engineers. When designed effectively, it enables them to take on higher-value roles, such as guiding the system toward optimal performance.

Workforce data reflects this trend. In oil and gas, 48% of professionals are aged 45 or older, while only 19% are 25-34. The industry faces a projected shortage of 40,000 workers, and 62% of younger talent find these careers unappealing. The primary challenge is not AI replacing experienced workers, but rather the impending retirement of this workforce and the limited window to capture their institutional knowledge.

In Alberta, this is already being taken seriously. Amii is leading a federally funded initiative to train nearly 5,000 energy workers in AI and ML skills. SAIT and the University of Calgary are building the infrastructure to bridge the gap between energy domain expertise and AI capability.

As BCG stated in 2026: "AI will reshape more jobs than it replaces." In heavy industry, this means experienced operators, engineers, and technicians become increasingly valuable. They are essential for identifying AI errors and providing corrections that lead to improved systems. This capability cannot be purchased from a vendor.

The Playbook: Five Shifts from Product to Capability

For executives, treating AI as a capability requires the following actions:

  1. Co-Own the Budget with Operations. If the AI budget sits entirely in IT, procurement will apply IT rules: vendor evaluation, license negotiation, and deployment. For AI that touches engineering or operations workflows, the domain leader needs to co-own the budget, success metrics, and governance. IT brings infrastructure and security. Operations bring the context that determines whether the system actually works.

  2. Staff for the 70%, not the 10%. For every ML engineer, you need domain experts who can audit results, a data team that can curate training sets, and process designers who can integrate AI outputs into real workflows. The model is the easy part. The organizational infrastructure is the hard part.

  3. Measure Capability, Not Product. Software either works or it doesn't. AI gets better with use - if you build the feedback loop. Track how accuracy improves across deployments. Track how human review time decreases per batch. Track how many edge cases the system handles autonomously this quarter versus last quarter. These are capability metrics, not product metrics.

  4. Pair every AI system with a domain owner. Not an IT owner. A domain expert who understands the work the AI is doing and can evaluate whether its outputs are trustworthy. If the model identifies a valve as a transmitter, someone in your organization needs to know that's wrong and feed that correction back into the system.

  5. Engineer the Feedback Loop. The 6% don't deploy perfect models. They deploy good-enough models with excellent feedback systems. Every human correction becomes training data. Every edge case becomes a test case. The system improves because the capability includes the process of improvement, not just the model.

The Category Error

The technology is finally good enough to reveal the real bottleneck: organizational capability. Frontier models are more capable than ever. Open-source fine-tuning tools are more accessible than ever. But accessibility should not be confused with production readiness.

The gap between a model that runs and a system that delivers is not technical. It's the people, processes, data discipline, and institutional commitment that turn a technology experiment into an operational advantage. Software is a product. AI is a capability. That distinction changes everything.

Organizations that get this right don't just deploy AI. They compound it. Each project makes the next one faster, cheaper, and more accurate. The advantage isn't the model - it's the muscle memory.

Organizations that get this wrong will keep spending billions on technology that sits unused, measuring success by deployment count instead of business impact, and wondering why AI isn't working - while assuming the next vendor, the next model, the next platform will be the one that finally delivers.

This will not change until organizations shift from purchasing AI as software to developing it as a core capability.

Where does your organization fall on the 93/7 to 10/20/70 spectrum? I am interested to know if experiences in energy and heavy industry align with those in other sectors, such as oil and gas, manufacturing, utilities, or any field where AI impacts operational workflows.

If you have experienced an AI rollout that either stalled or succeeded, I would appreciate your insights on what factors made the difference: whether it was the model, data, workflow, ownership, or feedback loop.