Kestrel Flight Academy is a fictional school, but its office could belong to any of a hundred real ones: a chief instructor, an office manager who is the actual backbone of the place, and a part-time IT contractor named Marco who is loved, mildly feared, and paid mostly in coffee. Between them they had quietly assembled nine small automated helpers — the booking export that ran each night, the weather-sheet updater, the roster email, the simulator-slot cleaner, and five more nobody could list from memory. It was, in other words, a textbook AutomationOps estate before anyone used the word. This is the story of the Tuesday that nearly went wrong, and of how the Atlas’s AutomationOps entry turned a lucky escape into a working system.
The Tuesday that nearly went missing
On a Monday night the slot cleaner ran, as it had for a year and a half. Its job was simple: find simulator slots marked unused for more than a week and release them. That week someone had renamed the shared calendar — “Kestrel Bookings” became “Kestrel Bookings (new)” — and the cleaner, reading a folder that no longer matched, had interpreted the confusion as a wide-open invitation. Priya, the office manager, arrived on Tuesday to a schedule that looked suspiciously light. Forty-one bookings, sitting in a stale copy on a backup drive, invisible to every instructor. She restored all forty-one by lunch, working through her own break, and nothing ever reached a student. But Marco, told the story over coffee, went pale in the specific way engineers do when they realise a guardrail saved them from explaining something at a board meeting. “The script did what I told it,” he said. “That’s the problem.”
What AutomationOps looks like for a five-person office
Marco spent an evening with the Atlas and came back with the least dramatic proposal in the world: treat the nine helpers as equipment. Each one gets a named owner, a written worst case, and a way to stop it that works in the dark. Dana, the chief instructor, liked the aviation logic instantly — “we don’t fly aircraft on vibes, why do we run bookings that way?” Nobody bought software. Nobody scheduled a meeting with an agenda item called “digital transformation”. They opened a spreadsheet, made nine rows, and started filling in the only three columns that mattered: who answers for it, what it could break, and what switches it off.
Naming the estate, then sorting it
The census took two evenings and produced the usual fossils: a report subscription from 2023, a sync that depended on Marco’s old laptop being awake, and two jobs doing the same thing twice. Then came the sorting. The weather-sheet updater could only display wrong numbers, so it sat at the bottom of the risk ladder. The roster email changed one thing reversibly — middle rung. The slot cleaner, now understood to be able to destroy a day’s schedule, moved to the top, where the Atlas’s rules say the treatment changes: a second person approves changes, deletions become recoverable, and rollouts happen in stages. Priya’s comment on that row — “so this one is a chainsaw” — became the office’s shorthand for consequence-based care.
Where the decision tree earned its keep
With risk sorted, Marco pulled out the Atlas’s decision questions and applied them to a backlog of “wouldn’t it be cool if” automation ideas, including Dana’s wish to automate the go/no-go weather call. The questions are blunt: does the task repeat, and is it actually understood? Can it be undone? Does it need judgement? Does it return real time, or is the prize consistency instead? The weather call failed at “needs judgement” — Dana stood by that, and by her claim that her instincts had saved two spring camps. The roster reminder passed everything and got built properly. The fuel-log reconciliation passed on consistency grounds, not time savings; it would never save hours, but it would catch a transposed digit the same way every time, which is worth more than speed in an audit.
AutomationOps guards, installed during one quiet week
The fixes themselves were small. Every change to a job now gets rehearsed first, against pretend targets, with a human reading the output — including a note that some rehearsal modes quietly skip command-level checks, so nobody trusts a green tick they haven’t inspected. The cleaner and the roster email learned the empty-list rule: an empty or wildly changed target list stops the run and asks. The booking export gained a batch cap, doing its work in slices so a bad run touches one slice instead of the whole archive. And every job received a failure budget of two, after which it halts itself rather than retrying its way through the night. Total build time: about a week of Marco’s evenings. Total cost: zero. Effect: the difference between a near-miss and a mystery.
The switch, and the drill that made it real
Each job got a documented pause control, plus one master halt for the whole collection — and the office agreed that whoever notices trouble pulls it first and explains after, with no 7 a.m. permission-seeking. Then came the part that surprised everyone: the drill. They picked a quiet Friday, pulled the master switch on purpose, and discovered two things. Stopping the helpers took four minutes, which felt fine. But the roster email also fed the whiteboard in the dispatch room, so “stopping” it needed a sticky note and a promise, or the instructors would have flown blind while believing they were informed. That is exactly the kind of fact a drill exists to surface, and it is now written on the job card.
The stall point, and how they dodged it
Marco described the trap they could have fallen into: teams that park themselves at “shared” — scripts in one place, plenty of goodwill, but no owners, no tiers, one login for everything. It feels comfortable, it can last for years, and it fails in an afternoon, usually at the worst possible moment in the booking season. He had a story for the office whiteboard, too: a trading firm that pushed unreviewed code into production in 2012 and watched roughly 460 million dollars evaporate in under an hour, not because people were careless but because nothing stood between a deployment and the market. Small offices are not exempt from big physics; they just have less cushion when physics wins. Kestrel’s answer was to keep the register alive — reviewed quarterly, owners updated when Priya went on leave, retired jobs actually deleted.
Why the story ends calmly
The honest moral is not “automation is dangerous”. The helpers still run, still save the office real hours every week, and nobody wants to go back to retyping rosters. The moral is that they stopped being lucky. A renamed calendar is exactly the kind of small, reasonable change that happens at every growing school, and the difference between Kestrel’s Tuesday and a genuinely bad week came down to a guard list anyone can copy:
- a named owner for every job, updated when people go on leave;
- a written worst case, so the chainsaws identify themselves;
- rehearsed changes, with a human who actually reads the output;
- empty-list refusals, so nothing interprets confusion as permission;
- batch caps, keeping one bad slice from becoming a bad archive;
- a master switch, drilled on purpose four times a year.
Flying schools grow on stories like this one, and on the hardware and recognition that mark real progress — the kind captured in announcements like a FRASCA 142 simulator joining a training programme for IFR work or a flight school earning a top award from AOPA. Behind every polished fleet update, including ones like a glass-cockpit twin-engine trainer arriving on the line, sits an office full of small machinery. The schools that scale happily are the ones that gave that machinery owners early — while the roster was still nine rows in a spreadsheet, and the coffee was still flowing.