The Capture Layer

At nine in the morning, a crew needed the answer to one question: had a piece of equipment been left in a safe state?

Nobody could confirm it. So they did the professional thing. They went back and verified it themselves — re-inspected, repeated every check, proved it from scratch. Hours of skilled work, by people who had other jobs waiting.

The answer surfaced that afternoon. It had been written down two days earlier, in a maintenance report, emailed to someone else's inbox. Every detail. Written before the question was ever asked.

The plant knew. The crew didn't.

I've watched some version of that morning happen in gold processing and in alumina refining, on more shifts than I can count. It is not a story about careless people — everyone in it did their job well. It is a story about a gap between what a plant knows and what the people on shift can actually reach. That gap is the most expensive thing in industrial operations that nobody has on a dashboard.

What capture actually is

Most plants are drowning in data and starving for information, and the reason is that three different things get called the same word.

Storage is where a number goes after a system records it. Plants are excellent at storage — the historian holds years of tags at sub-second resolution. Analytics is what you do to numbers you already have, and plants buy a lot of analytics.

Capture is the step before both: the moment a human observation becomes structured data. An operator notices a pump sounds wrong. A supervisor decides to keep feeding instead of holding. A technician finds a screen blocked ahead of a planned outage. Right there, in that second, the most valuable information in the plant exists — and in most plants it exists only inside a person.

That's the capture layer. It is the thinnest, least instrumented, most valuable layer in the whole operation.

We built plants that remember every number and forget every decision.

The historian can replay a trip to the millisecond. Nothing can replay the reasoning. Ask why the night shift held the unit instead of restarting it, six months later, and you will get a shrug and a guess — from a plant that can tell you that pump's discharge pressure at 3:47 a.m. to two decimal places.

Put a number on it. One control room, one twelve-hour shift: a supervisor makes — conservatively — thirty judgment calls before handover. Restart or swap to standby. Hold or keep feeding. Escalate now or watch it one more hour. That's two and a half an hour, which anyone who has actually run a shift will tell you is on the low side. Now multiply: around the clock, roughly 350 days a year, across the several areas a plant runs — digestion, precipitation, calcination, the powerhouse — each with its own senior decision-maker.

That is well over a hundred thousand expert judgment calls a year, in a single plant.

Now ask how many of them the plant could reconstruct a month later. Not the outcome — the reasoning. Almost none.

Three leaks

Frontline knowledge escapes through the same three holes in every plant I've seen.

The radio. The richest live stream in any operation is voice traffic — the state of the plant, transmitted continuously by the people standing in it. It has no storage. Every transmission evaporates the moment it ends. Nothing is written, nothing is searchable, and the only record is whatever the listener happened to retain.

The notebook. Walk any control room and you'll find shirt-pocket notebooks holding torque figures, valve positions, the trick that primes the standby pump — written at the moment of truth by the person who was there. No login, no lag, never crashes. The best capture device a plant has ever had, and readable by exactly one person.

Memory. The largest leak, and the slowest. A problem that beat the manuals for weeks gets solved in an afternoon because one operator remembered what the machine sounded like the last time it did this. That knowledge has one storage medium, and it retires.

In the United States alone, more than half the mining workforce — roughly 221,000 people — is expected to retire by 2029. The industry's own professional society names the consequence in its own words: a skill and knowledge gap. Over the same stretch, US mining and mineral engineering programmes fell from 25 to 15, and graduates dropped 39% since 2016.

The most valuable data in a plant walks out of the gate every time an experienced operator retires. We have known this was coming for twenty years, and our response has mostly been to hope the next person picks it up by standing nearby.

When the leak isn't just cost

Usually this gap is only expensive — duplicated work, slow decisions, a morning spent re-verifying something already written down. Sometimes it is not only expensive.

In September 2022, in the twelve hours before a fatal fire at the BP-Husky refinery in Oregon, Ohio, the plant raised 3,712 alarms. No crew can act on that many. Two workers, brothers, were killed. The warning was in the system. The system had no way to make it findable.

At Texas City in 2005, fifteen people died on a start-up that should never have happened. Everything needed to stop it was known one shift earlier — a tower filling far past a safe level, watched by a crew who then went home. The handover was thin, the logbook was vague, and a decision to hold the unit never crossed the gap between two crews.

And in 1977 a relief valve stuck open at the Davis-Besse nuclear plant and left its control room confused. Engineers inside the manufacturer wrote it up; by February 1978 they had drafted new operator guidance for exactly that failure. Thirteen months later the same failure hit Three Mile Island, and the operators misread it the same way. The guidance was still inside the company.

The answer to Three Mile Island existed before the accident. In writing. Addressed to the wrong place.

Capture worked in all three. Addressing failed.

Why "we'll do AI later" quietly fails

Here is the sequence I keep watching. A plant decides to get serious about prediction. Someone says, confidently, we have years of data. Then the project opens the data, and the data is tags — pressures, flows, temperatures. Every number the instruments produced, and almost nothing about what the humans saw, decided, or did.

You can build a model on that. You cannot build the model anyone actually wanted, because the interesting events — the near-miss, the workaround that became standard, the reason someone overrode the setpoint — were never captured in a form a machine can read.

Intelligence sits on top of a ladder, and the rungs are not optional:

• 1 · Structure the equipmentA register where every asset has one identity everyone uses. Without it there is nothing to attach history to.
• 2 · Structure the peopleWho was on shift, in which area, responsible for what. Attribution is what turns an event into evidence.
• 3 · Build the capture habitObservations logged at the moment they happen, by the person who saw it, in seconds — not reconstructed at the end of a twelve-hour shift.
• 4 · Accumulate real historyDepth on the same assets over time, including the reasoning, not just the readings.
• 5 · Then intelligenceOnly now does prediction have something worth learning from.

Almost every plant I've seen tries to buy rung five while standing on rung one. The vendor demo works because the demo data is clean; then it meets a real plant's history, quietly underperforms, and everyone concludes the technology isn't ready. The technology was ready. The data was never captured.

You cannot calibrate what you never measured. You cannot predict from history you never wrote down.

The constraint that expired

For most of my career there was a good reason the capture layer stayed broken: fixing it meant software, and software meant a vendor, a budget cycle, and a specification written by people who had never stood a shift.

Mineral processing was never a slow industry — we put expert systems on grinding circuits in the 1990s and cameras on flotation froth two decades ago. When intelligence arrives packaged from a vendor, we buy it. But the everyday layer — logs, tickets, handovers — could only be bought, never bent. So it froze.

That constraint has quietly expired. Software has been written with AI assistance for about five years now, and for the last three, the language you build in is plain English — the same words you'd use to brief a new technician. The people who know the plant best no longer need to hand their requirements to someone who doesn't.

I'm not claiming that makes everyone a software engineer. I'm saying the bottleneck moved. It used to be can we build it. Now it's do we know what to capture — and that question has always been answerable by the floor.

What good looks like on Monday

None of this requires a platform to start. It requires a few design principles, and any operation can adopt them:

• File information to the equipment, not to a personA report addressed to an inbox dies with that inbox's attention. The same report attached to the asset answers whoever asks, for as long as the asset runs.
• Make important things survive by defaultAn open item should stay open until someone closes it — not until someone forgets to re-write it. "He'll remember to mention it" is not a safety system.
• Let items show their ageIf a watch-item has crossed two handovers it isn't open anymore. It's stuck, and it needs replanning, not another hand-off.
• Answer every report visiblyOperators don't stop reporting because they stopped caring. They stop because the last three reports went nowhere. The response side of the loop is what keeps capture alive.
• Judge any tool by the 3 a.m. testIf it only works when the A-team is awake and unhurried, it doesn't work.

None of these are exotic. They are the disciplines a good supervisor already applies by force of personality — written down and made structural, so they survive that supervisor moving on.

The plants that get real value out of AI in the next decade won't be the ones buying the biggest models. They'll be the ones that started capturing properly three years earlier — because when the tools finally arrive, they'll be the only ones with anything worth pointing them at.

So the question I'd take onto shift tomorrow isn't what could AI do for this plant. It's simpler, and harder: when your crew has a question at nine in the morning, is the answer addressed to them — or to someone's inbox?

Sources — BP-Husky Toledo alarm count and fatalities: U.S. Chemical Safety Board final investigation report, June 2024. Texas City 2005: U.S. Chemical Safety Board report 2005-04-I-TX. Davis-Besse 1977 and Three Mile Island 1979: NRC and Kemeny Commission record. US mining workforce retirement, programme and graduate figures: Society for Mining, Metallurgy & Exploration, Workforce Trends in the U.S. Mining Industry.