"Eight weeks." was the answer I received from a senior Instrumentation engineer when I was managing a pilot plant build. We needed a rough IO count so we could order the equipment on the critical path. It didn't need to be perfect, it just needed to be good enough to place orders on the critical path without causing schedule slipping.
I took the request to our instrumentation engineer. He was the most senior person on that part of the project and the Qualified Person who would eventually sign off the design. He thought about it carefully and came back with his estimate: eight weeks of full-time work.
Eight weeks we didn't have.
I pushed back. I had a different approach in mind, and I walked him through it: the instruments were already in the CAD P&ID files, so pull them out, use Excel to group them by type and match each type to its IO requirements. I described other ways of getting there too. None of it landed. No matter how I explained it, he couldn't see how it could be done any quicker, and he wouldn't take it on. In his mind, an IO count meant going through the design instrument by instrument, the way he had always done it, to the standard he had always held himself to. He wasn't being difficult. He was being honest, and that was the problem.
So I went to one of our junior engineers and gave her exactly the method he had turned down. I told her to pull the instrument list straight out of the CAD files, put it into Excel tables and group the instruments by type. Then she was to match each type to its IO requirements (analogue or digital, input or output, signal type) and add up the totals.
We had our estimate in less than a day.
I've thought about that day a lot since, because I now see the same thing happening in almost every conversation I have about AI.
The Same Story, at Scale
Earlier this year, PwC asked 4,454 CEOs in 95 countries what AI had actually done for their business. 56% said they had seen no significant financial benefit so far. Only one in eight said AI had both cut costs and grown revenue. When PwC checked in again in August, the picture had barely changed.
It would be easy to dismiss this as one survey, but it isn't. McKinsey's 2026 State of AI found that only 37% of organizations see any positive effect on operating profit (EBIT) from AI, and just 6% see significant value. BCG puts the share of companies getting AI value at scale at 5%. IBM found that only 37% of AI initiatives delivered what leaders expected. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027.
Mining is no different. In EY's survey of 500 senior mining and metals leaders, AI tops the investment agenda, yet returns so far have been limited by siloed data and a poor fit with business needs.
One finding stood out to me most. In the same McKinsey study, 80% of people said AI made them personally more productive. So the tools work for individuals, but the business isn't earning more. Everyone is moving a little faster through a process that hasn't changed.
That's my eight-week estimate. The engineer would happily have used a faster tool to do it his way. What he couldn't do was accept a different way of doing it, even when I spelled it out for him.
To be fair, not every study is this gloomy. Wharton's 2025 AI Adoption Report found that three in four leaders who measure their returns report a positive one. In mining, McKinsey's preliminary figures show 15 of 19 major miners now reporting financial gains from AI. The value is real. It's just going to a minority of companies, and they're doing something the rest aren't.
What the Winners Do Differently
The research is unusually consistent on this. The companies that see returns don't add AI to their existing processes. They redesign the work.
McKinsey found that high performers "fundamentally redesign workflows that are enabled by AI rather than insert AI into existing ones." Nearly three-quarters of them have done this, compared with a quarter of everyone else. Bain says it more bluntly: "AI doesn't fix workflow debt; it locks it in, speeds it up, and makes it vastly more expensive to unwind." BCG's rule of thumb is that only 10% of AI value comes from the algorithms and 20% from technology and data. The other 70% comes from people and processes. As McKinsey's mining team put it recently, "A pile of AI initiatives is not a transformation."
The method I gave my junior engineer didn't use any new technology. Excel and CAD had been on the senior engineer's desk for decades. The difference was that I started from the result we needed and worked backwards, and then gave the work to someone with no attachment to how it had always been done. She was willing to try it. He wasn't.
Why Your Experts Can't Find Your Use Cases
Here's the uncomfortable part. Most companies start their AI programs by asking their most senior, experienced people to find use cases. It seems sensible, because they know the work best. It's also why so many programs stall.
Expertise can lock in the familiar method. This isn't just my observation. In a well-known study, expert chess players who were shown a familiar but weaker solution kept looking at it, even while saying they were searching for something better. Their performance dropped to that of players far less skilled. Psychologists call this the Einstellung effect. Researcher Erik Dane calls the broader pattern cognitive entrenchment: the deeper your expertise, the more firmly settled your picture of how the work is done.
If someone isn't yet getting the most out of the last few decades of technology, they're unlikely to see what this one makes possible. Ask them where AI could help, and you'll get the eight-week answer: honest, competent, and focused on small improvements to today's process.
Incentives favour the status quo. Seniority, billable hours, headcount, professional identity and sign-off authority are all built around the current way of working. Asking the people who benefit most from that system to lead its replacement is a built-in conflict of interest. The data reflects this. In a 2026 WRITER survey of 2,400 executives and employees, 29% of employees admitted to sabotaging their company's AI strategy, and 79% of executives reported implementation struggles, including internal power struggles. Nobody needs to act in bad faith for an AI program to fail quietly. It's enough that nobody has a reason to push it forward.
Your Data Is Already Sitting There
A recent episode of the Everything AI podcast, Agentic AI x ERP: Layering Intelligence Beside Systems of Record, describes this well. Your ERP is an excellent system of record, but it was never designed to be a system of action. It can tell you an invoice exists. It can't decide whether to pay it. Everything in between, such as matching, checking and deciding on exceptions, is still done by people working in spreadsheets and email (The Agentics Co.).
My CAD files were a system of record too. Every instrument we needed to count was already in them. The eight-week estimate came from treating that data as something to read through by hand. The one-day answer came from my insisting that we treat it as something to query.
Mining operations are full of data like this: block models, LIMS, historians, maintenance systems, ERPs, CAD files. The facts are already there. What's missing is a way of working that puts them to use.
Don't Sideline Your Experts
None of this means senior people don't matter. My rough count got equipment ordered on time, but the detailed design still needed the senior engineer's judgement and his sign-off as the QP. The fast answer and the expert answer weren't in competition. We needed both, in the right order.
That's also how good AI governance works. The podcast's model places controls where actions happen, with permission checks on every action, full audit trails, human approval for high-risk decisions and a kill switch for every agent. McKinsey's mining team says the same: models recommend options, while specialists set the limits and stay in control.
Let new workflows do the first 80% fast. Let your QPs, senior geologists and engineers validate, challenge and sign off. Their expertise becomes the governance layer, not the bottleneck.
How to Get Your Day Instead of Eight Weeks
Start from the outcome, not the current task. Don't ask "Where can AI help with what we do?" Ask Bain's question instead: "If we were designing this process from scratch today, what would it look like?"
Lead the redesign from the top, and give it to people willing to run with it. Don't wait for your experts to find the new way of working. Someone with authority over the outcome has to set the new method and back it, as I had to on the pilot plant. Then put it in the hands of people who have no stake in keeping things as they are. Bain found that the most successful companies are 40% more likely to have cross-functional AI teams with clear ownership, combining subject matter experts with frontline users.
Map the data you already have. Before you buy anything, find out where the facts already live. Most mining operations have far more structured data than they use.
Go small, fast and measurable. Prove the effect on the P&L in a tightly scoped 4–8 week pilot before you scale. Maintenance is a strong starting point. At one open-pit mine, an AI troubleshooting workflow tripled the average time between failures. Procurement, grade reconciliation and ESG reporting are also good candidates.
Build governance in from day one. Agree on approval gates, audit trails and who owns the kill switch before the first agent goes live.
Change the incentives. Reward people for redesigning work, not just for doing it. If someone turns an eight-week task into a one-day task, that should help their career, not threaten it.
The Real Lesson
When PwC reports that most CEOs see no benefit from AI, that isn't really a verdict on AI. It's a verdict on how we're trying to use it.
My instrumentation engineer wasn't a bad engineer. But even with a faster method laid out in front of him, he couldn't see it and wouldn't do it. If you rely on the people who perfected today's process to find a better one, or even to accept one, you'll usually get an honest, well-reasoned eight-week answer.
AI asks us to rethink how we work from the ground up. The companies that see the returns will be the ones whose leaders are willing to set the new way of working themselves, hand it to people ready to run with it, and then let their experts sign off the result.
At LOMexcel, we help mining companies find, scope and deliver AI use cases that change how the work gets done. If your AI program is stuck at the eight-week answer, let's talk.