On June 1, 2009, Air France 447 went into the Atlantic and 228 people died.

The proximate cause is well documented. Ice crystals blocked the pitot probes, airspeed indications went inconsistent, the autopilot dropped out, and the crew's response put the aircraft into a stall it never came out of.

Here is the part that should bother maintenance leaders. The probes on that airframe were Thales AA units. Airbus had identified 32 separate icing events involving those probes between November 2003 and the night of the crash. A service bulletin recommending the newer BA model had been out since September 2007, on an optional basis. Air France had already begun swapping them across its A330 and A340 fleet. The program simply had not reached the aircraft that flew as 447 yet. Within weeks of the accident the entire fleet was done, and EASA moved to make the change mandatory across every A330 and A340 in service.

The fix existed. The evidence existed. What decided the outcome was sequencing. Which tail number got the work first, and when.

That is a maintenance intelligence problem, and the industry still has it.

I run Kquika, where we build predictive maintenance systems for airlines. I came into this from genetics research, which turned out to matter more than I expected, because it meant I arrived without the assumptions that aviation insiders stop noticing they hold.

The paradox of precision

Commercial aviation is the safest transportation system ever built and one of the most wasteful. Those two facts share a cause. The safety record came from obsessive adherence to fixed maintenance intervals, and a fixed interval is, by construction, wrong for any individual component. It is either early or late. It is almost never right.

The numbers back this up. IATA put global airline MRO spend at $103.9 billion in 2024, around 11.5% of total operating cost, and that figure has climbed since. A substantial share of it goes to unscheduled work, which is exactly what the schedule is supposed to prevent. Meanwhile components come off aircraft with serviceable life still in them and go to scrap. Over-maintaining and under-maintaining, often on the same airframe in the same week.

The waste is not confined to the accounting. Delay costs run past $100 per minute once you count crew, fuel, compensation, and downstream rotations. AOG events cost far more than that. Emergency parts move by air freight because they have to, burning fuel to solve a problem that adequate forecasting would have made routine.

The environmental math is worse than the industry likes to acknowledge. Manufacturing a replacement part carries embedded carbon. Scrapping a component with life left in it throws that carbon away. Repositioning aircraft around hangar availability burns fuel for zero revenue. Aviation has committed to net zero by 2050 while running a maintenance model that quietly works against the target.

What genetics taught me about turbine blades

My route into this was not aviation.

During graduate work I built software called Sibs that read 25 to 30 years of familial data to predict genetic patterns and disease susceptibility. The interesting part was never the algorithm. It was the realization that degradation leaves a trail long before it produces a symptom, and that the trail is legible if you hold enough history and know the shape you are looking for.

Aircraft components behave the same way on a compressed timescale. A turbine blade's vibration signature shifts weeks before a crack becomes visible on inspection. Landing gear hydraulics show pressure irregularities well ahead of seal failure. Engine oil chemistry moves in a predictable direction as internal wear accelerates. Modern aircraft generate terabytes of sensor data per flight, so the trail is there.

Most of it gets logged and never read. That is the actual scandal.

None of this is speculative. Delta has reduced maintenance-driven cancellations using engine health monitoring at scale. Lufthansa Technik has been running machine learning against component data for years and folds predicted replacements into scheduled visits rather than absorbing the AOG hit. Real programs, published results. They are also isolated, which is the problem worth writing about.

The silo problem

Airlines treat operational data as proprietary. The reasoning is that maintenance insight confers competitive advantage, so it stays in house.

Coming from research, this struck me as absurd. Medical progress happened because researchers pooled patient data. If every hospital had kept its outcomes to itself we would still be guessing at oncology.

Aviation already knows this and applies it selectively. When an investigation finds metal fatigue in a wing spar, the finding goes to every operator of the type. When a study identifies a training deficiency, the industry adopts the correction. Safety knowledge has never respected commercial boundaries. Maintenance knowledge does, and nobody has offered me a defensible reason why.

There is a statistical argument here before there is a philosophical one. A mid-sized operator with 60 aircraft will never accumulate enough failure events on a given part number to model it well. The pattern lives at the level of thousands of airframes across dozens of operators. At the level of one fleet it is invisible, and no amount of internal analytics conjures it into view.

Fixing this means building systems where carriers contribute to and draw from a shared model without exposing what they consider commercially sensitive. That is the engineering problem, and it is solvable. It is what we have been building for years.

Why the timing is forced

Aircraft deliveries have run well behind plan for most of a decade, so airlines are holding older aircraft longer than they intended. IATA puts the average age of the global fleet at 14.8 years against a long-run average nearer 13.6, and slow fleet renewal keeps pushing it up. Maintenance cost does not scale linearly with age. It accelerates.

Environmental pressure is tightening at the same time. Aviation sits at roughly 2.5% of global CO2 and attracts scrutiny well out of proportion to that share. Intelligent maintenance is one of the few decarbonization levers that pays for itself immediately, because longer component life means less manufacturing and better scheduling means fuller utilization.

Margins leave no slack either. Carriers running low single-digit net margins cannot afford to scrap serviceable parts and eat unscheduled failures in the same fiscal year. Smaller operators feel it hardest, since they lack both the data volume and the engineering headcount to build proprietary prediction in house.

Collaborative models close that gap. A regional operator can reach prediction accuracy comparable to a major by drawing on patterns learned across a global dataset. Call it self-interest rather than charity. Network reliability is set by its weakest links, and in an interlined industry, everyone's schedule depends on somebody else's dispatch reliability.

The parts that are not technology

Software alone will not fix this.

Regulators need clear frameworks for AI-informed maintenance decisions. Today the certification pathway for a model-driven interval adjustment is vague enough that most operators will not attempt it, so proven technology sits idle while the rulebook catches up. What the industry needs is guidance on responsible implementation with the safety bar left exactly where it is.

Maintenance crews need training rather than displacement. A prediction is an input to human judgment. The engineer with 20 years on the floor is the one who decides whether the model is telling the truth about this airframe in this operating environment. Any vendor implying otherwise is selling a liability.

Leadership needs to drop the data hoarding, and I will put that in plain language because it deserves it. The belief that withholding failure data creates competitive advantage is wrong, and holding onto it costs the industry billions a year.

Where this lands

Airlines that pool intelligence will outperform the ones that do not. Environmental regulation will keep favoring operators who can demonstrate measured efficiency rather than stated intent. Passengers migrate toward reliability whether or not they can articulate why.

Come back to 447. The technology to prevent that specific failure existed and was already fitted to other airframes in the same fleet. What was missing was a way to decide which aircraft needed it first, driven by evidence instead of a rolling program. Seventeen years on, that decision is still made mostly by calendar.

We have already proven that sharing safety data makes flying safer. Point the same mechanism at maintenance and it makes flying smarter.