A Parts Desk Story: Learning to Run AI-Assisted Operations

A Parts Desk Story: Learning to Run AI-Assisted Operations

The story below is a composite, drawn from how small aviation suppliers typically stumble into modernisation, but the shape of it will feel familiar to anyone who has worked a parts counter. It is the story of a five-person operations team at a regional aircraft parts distributor — call them Meridian Aero Supply — and the eight months it took them to become comfortable with AI-assisted operations. Not the software part. The human part.

The Tuesday that started it

It began, as these things do, with an AOG call. A charter operator’s Caravan was grounded for a starter-generator, Meridian’s system showed two in stock, and the shelf showed none. Somewhere across four months, two inventory records had drifted from reality, and nobody had caught it because nobody’s job was to catch it. Priya, who ran the warehouse, spent that Tuesday calling competitors to source the unit at cost, just to keep the customer. Her boss, Tomas, spent it staring at a spreadsheet, wondering how many other records were quietly wrong.

Every parts business carries this risk. Inventory drifts, certs get misfiled, a supplier’s lead time stretches from ten days to thirty without anyone updating the quote templates. The failures are rarely dramatic; they are slow leaks that surface at the worst moment — usually attached to a grounded aircraft. The buying-side version of these traps is something we have catalogued before in our rundown of avionics purchasing errors that cost real money, but Meridian’s problem was the mirror image: the seller’s own operations going quietly out of tune.

The purchase that almost repeated the pattern

Tomas did what most owners do — he went looking for software. He found plenty: platforms promising intelligent inventory monitoring, demand forecasting, automated reorder triggers, exception alerts. In a demo, one vendor showed him a system learning the normal rhythm of stock movements and raising a flag when a record behaved strangely, the way those two starter-generators had. Another showed automated responses: low stock on a fast-moving consumable triggering a purchase order draft without a human touching it.

He nearly signed that week. What stopped him was a question from Dev, the youngest person on the team, who managed the webstore listings. “Who’s going to tune it when it’s wrong?” Dev had watched a friend’s e-commerce company install a similar system, drown in false alerts for a month, and shut it off. The software had been fine. The team had never learned to work with it — to adjust its sensitivity, to teach it about seasonality, to decide which of its suggested fixes could run unattended. The tool was a instrument nobody could read.

Learning before buying

So Meridian did something unusual for a company its size: it delayed the purchase and spent the budget’s first slice on people. Dev enrolled in a structured program — AIOps School’s certification track for AI-driven operations — starting at its Foundation level, a thirty-day course he fit around work at ten-odd hours a week. Priya audited the same material more casually, focusing on the units about how detection models decide what counts as abnormal. The program’s ladder continued upward through an Engineer stage and beyond toward Professional and Architect levels, but Tomas was clear-eyed: a five-person team needed one competent practitioner and one informed skeptic, not a department of architects.

Dev later moved into the second rung, a 45-day stretch heavier on lab work, where he built monitoring pipelines and wrote auto-remediation rules against simulated environments. He described it to the team as “learning to fly the autopilot” — understanding not just what buttons exist but what the machine is assuming when it acts. The hands-on labs mattered more than the lectures, he said; breaking a fake system teaches you what alerts from a real one mean.

The quiet go-live

When Meridian finally deployed its platform, the rollout was almost anticlimactic — which was the point. For the first two months, the system only watched. It learned the pulse of the business: how gasket stock moves in annual-inspection season, what a normal day of webstore orders looks like, how long each supplier really takes against their promised lead times. Dev reviewed its flags every morning over coffee and graded them in a shared doc: real, noise, or interesting.

The grading sheet became the team’s trust ledger. By week six, the “real” column included catches no human had made: a slow mismatch between received-goods logs and shelf counts in one bin range — the starter-generator failure mode, caught this time at a drift of two units instead of a crisis. A fastener SKU whose sales had tripled after a type club’s maintenance bulletin, headed for a stockout nobody saw coming. Fasteners, as it happens, are where small errors hide best; our deep reference on aircraft hardware and fastener selection exists precisely because that corner of the catalogue is dense enough to defeat memory.

Only after the trust ledger showed a month of clean grades did the team allow the system to act on its own — and only in the safest lanes: drafting reorders for human sign-off, re-syncing webstore stock counts, holding orders that tripped a certificate mismatch until someone reviewed the paperwork. Priya insisted every automated action write a log entry in plain English. “If I can’t read what it did,” she said, “it didn’t happen.”

Eight months on

The numbers were respectable: inventory accuracy up, two stockouts averted in a quarter, the AOG response feeling calm instead of frantic. But the change Tomas talks about is temperamental. The team stopped being surprised. Mondays lost their archaeology — no more reconstructing what went wrong over the weekend, because the weekend’s anomalies arrived pre-flagged with context. Priya, the most skeptical at the start, became the system’s best trainer, feeding it the warehouse intuition she had carried in her head for a decade. Control-systems orders, the trickiest category in the building — the one whose pitfalls fill our guide to avoiding control-system component purchasing traps — got their own tightened alert rules, written by Dev, reviewed by her.

And Dev got a credential that outlives this job, which Tomas regards as the deal’s hidden clause: training people makes them more employable elsewhere, and the only answer is being a place worth staying.

What Meridian would tell another parts house

Three things. First, your operations already produce the data; the gap is continuous attention, and machines are better at vigilance than people. Second, buy the training before the tooling — one month of structured fundamentals for a skeptic and a practitioner beats any feature list. Third, let the system earn autonomy one reversible action at a time. The technology is the easy part of AI-assisted operations. Readiness is the project.