Controlled flight into terrain:
Why the dangerous AI isn't the one that breaks
1 The failure nobody puts in the budget
Picture the quietest way an operation loses money. Not a fire, not an outage. A single number in a planning system, an on-hand quantity, a lead time, a supplier's payment terms, that is simply wrong, and has been for months. In the old world, that number sat in a report nobody fully read and did a little damage slowly. In the new world, you hand it to an AI, and it makes ten thousand confident decisions on top of it before lunch.
That's the shift worth watching this year, and most of the industry is looking the wrong way. The headlines are about pilot purgatory, the AI project that demos beautifully in March and gets quietly defunded in September. Gartner expects more than 40% of agentic-AI projects to be cancelled by the end of 2027, citing runaway costs and value nobody can pin down. Read the post-mortems, though, and the model is almost never what failed. The data underneath it was.
Operators already feel this in their gut, and the surveys are finally catching up. In the 2026 State of Data Integrity and AI Readiness study Precisely ran with Drexel's LeBow College, 565 data professionals worldwide, only 12% said their data was actually good enough for AI to use well. Not 12% behind. Twelve percent ready. And here's the part that should stop you cold: 88% of leaders said they felt confident in that same data. The gap between "I'm confident" and "it's actually ready" is exactly where the expensive surprises live.
The bill is real, too. Gartner has long pegged the cost of poor data quality at an average of $12.9 million per company, per year. IBM's 2025 study found 43% of chief operating officers rank data quality as their single most important data priority, ahead of everything else, and more than a quarter of organizations put their annual losses to it north of $5 million.
So the failure nobody budgets for isn't the AI that breaks. It's the AI that works beautifully, fast, fluent, sure of itself, on a number that was wrong before it ever arrived.
2 Controlled flight into terrain
Aviation has a precise, sobering name for this. It's called controlled flight into terrain, CFIT. A perfectly good aircraft, no mechanical fault, an experienced crew, every gauge in the green, flown straight into a mountainside. Nothing malfunctioned. The airplane did exactly what it was told. The crew's picture of where they were just didn't match where they actually were, and a machine executing flawlessly on a wrong picture doesn't hesitate. It arrives at the wrong place right on schedule.
That is the shape of the risk when you point capable AI at operational data no one has calibrated.
The old spreadsheet-and-email way of running things had one accidental virtue: it was slow, and it passed through human hands. When a demand figure looked absurd, a planner squinted at it. When two systems disagreed about what a part weighed, somebody noticed the freight quote came back strange. That friction was expensive, and it was also a filter. Bad data ran a gauntlet of small human hesitations before it became a decision.
AI takes the hesitation out. That's the whole pitch, and it's a good one. But take the friction out of a pipe carrying dirty water and you don't get clean water faster, you get dirty water everywhere, faster. The field has even started renaming its oldest law to fit the moment: garbage in, garbage out has become garbage in, amplified garbage out. Bad data used to gather dust. Now it moves at machine speed, wearing the confidence of a system that reasoned its way there. Speed times confidence times a wrong input isn't intelligence. It's CFIT with a nicer dashboard.
3 Read first. Then act.
Here's the design decision underneath our whole product, and I'll say it the way I'd say it across a table. Every operation's data is dirty when we start. I've never once seen an exception, and I've stopped expecting one. So we built The Captain to do the thing a good operator does before trusting a number: read it, and notice when it doesn't hold up.
That's the job of our Data Quality Detection Layer. It watches for the six kinds of dirty data that quietly erode an operator's trust, stale records, missing identifiers, duplicate keys, two systems of record disagreeing about the same thing, schema drift, and reference-data gaps. Five of those, frankly, a decent single-system tool can catch. The sixth is the one that starts the rush orders and empties the shelves, your ERP says on-hand is 47, your WMS says 52, and you can only catch it from something standing above both systems, reading them together. The Captain already does exactly that, read-only, so she's watching for the contradiction the single-system tools are built to miss.
And this isn't a diagram waiting to be built. Today, on demand, The Captain runs these checks live against real SAP S/4HANA and Salesforce data. She reads the actual records, read-only, never a write back to your systems, and surfaces the ones that don't survive contact: a supplier invoice with no payment terms, an opportunity whose close date slid past weeks ago. She cites the exact records she read, tells you plainly that she sampled, say, 500 of them rather than pretending she read all million, and hands the decision to a named person. When she puts a dollar figure on it, it's labeled modeled, a projection from your live data carrying its own assumptions on its sleeve, never dressed up as money already banked.
Two connected systems disagree on on-hand quantity for 3 SKUs: your ERP reads 47 units, your WMS reads 52. A replenishment decision built on either number alone would be wrong. Reconcile before acting.
The operator decides:
Where we're headed next is the part I'm most excited about. Right now she does this when you ask. The roadmap is to have her doing it without being asked, a baseline sweep the moment a new system connects, a quiet check on every read, a nightly pass that comes back not with a wall of red but with a scoreboard: your team cleared 31 issues yesterday; your CRM is 94% clean and climbing. And we're widening the live coverage system by system, so the same live scan that reads SAP and Salesforce today reads the rest of your stack tomorrow. That's not someday-maybe language. It's the next few build cycles, and this newsletter will call each one as it lands.
The thing I want you to take from this, whether or not OpsATC.AI is ever on your list: an AI that acts before it reads is an autopilot with no ground-proximity warning. The one worth having reads first, tells you what it found, and lets you decide whether to pull up.
If you run operations at a distributor, contract manufacturer, storage OEM, hybrid, logistics carrier, or hub provider, here's a CFIT check you can run this quarter for free, no vendor required: take the one number your team most wants to hand to an AI, and trace it back through every system it touches. If two of them disagree, you just found the mountain before the AI did. Want to talk it through, design partner or not? Reach me at [email protected]. No sales motion attached.
Read your instruments. Trust the ones you've calibrated. Captain out.