The Most Important Technology on the B-17 Was a Piece of Paper
November 17, 2026 · Read Time: 8 mins
A 1935 bomber crash explains where AI will earn its keep in Army human performance.
On the morning of October 30, 1935, the most advanced bomber anyone had built took off from Wright Field in Dayton, Ohio. It got about 300 feet up before it stalled and fell back onto the field, where it burned.
The airplane was Boeing’s Model 299, a four-engine bomber the Army Air Corps badly wanted. Maj. Ployer Hill, who ran flight testing at Wright Field, was at the controls, and Boeing’s chief test pilot stood behind him to give advice. Both men died of their injuries. When investigators went through the wreckage they found nothing wrong with the engines or the structure. What they found, in the tail section that survived the fire, was a control lock that had never been released. It’s the device that holds the control surfaces still while an airplane is parked (link).
It was Hill’s first flight in the 299, though he was an experienced test pilot who had flown the early Martin B-10 bombers. The board stopped short of calling it pilot error. It concluded that on an airplane that size, almost any pilot taking off the same way would have discovered the locked controls too late.
The War Department’s General Staff was less forgiving. It argued the Boeing had crashed because it was too complex for anyone to fly safely, and in later retellings the bomber became “too much airplane for one man to fly.” The production order went to a smaller Douglas design. The Air Corps still wanted the Boeing and got 13 of them as test aircraft, which went to the 2nd Bombardment Group at Langley Field, Virginia (link).
So Air Corps crews wrote out what had to happen, in order. They ended up with four lists: takeoff, flight, before landing, after landing. That was the fix. The Air Corps had worked out that the weak point was a pilot’s memory, and a card could cover for it. Over the next few years, by the same account, those test bombers logged more than 9,200 flying hours without a major accident, and the Army went on to buy more than 12,000 B-17s.
I trained as a historian before I ended up in software, and this is one of my favorite stories from the period. The most sophisticated machine of its day was rescued by something that for all practical purposes cost nothing. I’ve been thinking about it lately because of how often I’m asked what AI is going to do for soldier readiness, and how rarely the true answer is exciting.
Then the lists took over
A good idea in a big institution tends to keep going long after it has done its job. The habit of writing things down so nothing gets missed left the cockpit a long time ago, and in the wider Army it grew into the tracker and the mandatory module. In 2002 an Army War College study added up everything a company commander was directed to train, mission-related or not, and got 297 days’ worth, to be fit into 256 available days. Ten years later the Army’s Inspector General inspected 16 locations and found that companies at none of them could complete all of their mandatory training and administrative tasks (link).
The Army hasn’t ignored this. In April 2025 it scrapped or made optional eight mandatory training programs and cut its training regulation from more than 250 pages to 132 (link). Ted Delicath, an Army Reserve captain, picked up the argument in Small Wars Journal this September. He calls what remains the checklist tax. His point is that a lot of it is work worth doing, like counseling a soldier or writing up an award, wrapped in drafting that doesn’t need a leader’s time. He suggests treating an AI model the way the Army has always treated a clerk: the clerk assembles the document, and the leader checks it and signs (link).
I like that framing, partly because it’s so unglamorous. About ninety years ago a card took a load off the pilot’s memory. Today the paperwork is the load, and the job hasn’t changed much. Get the routine out of the way of the person whose judgment you need.
What boring looks like on a team
I’m not a soldier and never have been. I work for a company that builds software for sports teams and military units, so most of what I know about this comes from training rooms and team operations offices.
Our CTO, Reed Shaffner, opened his session at our Human Performance Summit this year by asking for a show of hands. Most of the room used an AI assistant every day, and a good number had built an agent of their own. When he asked who had an AI application running in production, three hands stayed up. These were some of the most sophisticated performance staffs in the world (link).
The things that do reach production tend to be small. During March Madness, NCAA championship staff were fielding policy questions on site, the kind that used to mean scrolling through a 200-page manual on a phone. They asked us to build them a way they could ask the manual directly and get an answer with the passage cited, so they could check it. Our write-up of that project has a line I’ve stolen more than once: “Ctrl+F is not a workflow” (link).
An athletic trainer finishes an evaluation and talks through her note while the form fills in the required fields and waits for her review. A lineman takes a picture of his plate so he doesn’t have to type out his dinner (link). Nobody is going to make a movie about any of this. Reed’s summary is the best one I’ve heard: “It can save hours. It can’t make the call.”
He’s just as blunt about the glamorous end of the market. A model that predicts every athlete will be healthy every day is right more than 95 percent of the time and helps nobody, because on any given day injuries are rare. A seven-season UEFA study of top European soccer clubs counted about two per player per season (link).
Reed also passes along a number from Dr. Robin Thorpe, who told the same summit about a survey of 21 elite performance and medical directors. They estimated that between 75 and 90 percent of the data they collect never influences a decision. It’s a small survey and I haven’t seen it published, so take it as a practitioner’s read. It’s still the reason I think the boring work comes first. Before anyone needs a prediction, they need what they already collect to reach the person who can act on it.
What it could look like in a brigade
The Army describes its H2F Management System as a way to give units the kind of data-driven insight usually found in special operations and professional athletics (link). The people closest to it are already asking unglamorous questions, which I find encouraging. Command Sgt. Maj. Alex Kupratty of Army Transformation and Training Command told an audience in September that he wears a Whoop and an Oura Ring and worries about data fatigue. “I probably look at half of the data that it provides me daily,” he said, and he called that estimate generous. Jonathan Flinn, the H2F program director for the 199th Infantry Brigade, put it more plainly. Earlier attempts failed, he said, because soldiers didn’t want to log in (link).
Those are the right worries. An H2F team is small next to the brigade it supports, and every hour a physical therapist spends retyping a note is an hour she isn’t with a soldier.
So picture the dull version. A squad leader asks what the regulation says about a soldier on profile and gets the paragraph, with the citation, about as fast as he can ask. A physical therapist dictates between appointments and reviews the record afterward. A recovery session gets counted because it was booked in the system the staff already works in, which is the case I made in October for readiness data nobody should have to collect (link). The commander’s weekly summary arrives as a first draft built from records that already exist, and a person corrects it before it goes up the chain.
Soldiers aren’t waiting on anyone’s product roadmap, ours included. The Department’s GenAI.mil platform had about 1.1 million users within two months of its launch last December (link). When DefenseScoop asked troops early on what they were using it for, the answers were things like award packages and performance reviews (link). They went for the paperwork first.
Most of what I described in the last section is running for sports teams today. The Army versions have to earn their way through security accreditation before they reach a soldier, and that’s how it should work when the data belongs to soldiers.
Boring still has to be right
A clerk who is fast and wrong is worse than no clerk. Delicath says as much: feed the model a bad personnel record and it will hand back a tidy document that’s wrong. Staff officers from a 101st Airborne brigade wrote this spring about taking AI tools to the Joint Readiness Training Center. Their plans officer talked through a scheme of maneuver for about 20 minutes, and half an hour later they had a first-draft brigade operations order in the proper format. Over the same rotation the tools also mislabeled units and invented control measures that weren’t in the order, and it was the section chiefs who caught the mistakes (link).
That’s why I care more about the plumbing than the model. Reed’s argument, and the one we’ve built our products around, is that everyone will end up with access to roughly the same models. The difference comes from what sits underneath: whether a soldier shows up as one person across every system that holds a piece of his record, and whether that data is clean before a model ever sees it.
The rules we hold ourselves to would fit on a card. An answer links back to its source so you can check it. When the software fills in a record from a dictated note, a person reviews and approves it before it’s submitted. We also publish the questions we think a buyer should put to any AI vendor, starting with where the data lives and who can see it, and we expect to be asked them ourselves.
Walter Boyne, the retired Air Force colonel and aviation historian whose account of the crash I drew on above, described the prevailing attitude in the Army’s air arm as “If it has a stick and a throttle, go fly it.” Crews used the card anyway, and the airplane flew. I’d hold AI in human performance to the same standard. If it gives an H2F team an hour a day back with soldiers, and leaves a record a commander can trust when signing it, then it’s done its job, however dull the demo looks.