Explain how to say no to an AI initiative without being labelled the person who blocks innovation.

Draymark Robotics builds the pathfinding and collision-avoidance model that drives a warehouse's fleet of small delivery robots between racks. Zubin Kalantari owns that model. Renate Odukoya runs field deployment and wants to switch Cadwell Grocery Distribution's whole fleet to fully autonomous "free-roam" routing in twelve days, the same week a rival announced they'd done it first. Zubin has to say no to Cadwell without sounding like the one person in the building who's afraid of shipping something new.

How to answer this, stage by stage

Nobody in the room is grading whether you sound confident saying no. They're grading whether your no is checkable, or whether it's just a feeling wearing a decision's clothes.

Let's learn

Five steps, and the whole argument sits inside the fourth one: the number the model reports, and whether that number is actually true where it's being reported.

What happens the first time a robot is sure about something it shouldn't be sure about?

Draymark's robots move totes through a warehouse. Right now most clients run them in "lane mode": geofenced paths, and a central system that sequences who goes first at any merge, especially the blind ones. It's safe and it's a little slow, because every robot waits its turn even when nobody's actually coming.

Three months ago, Draymark piloted "free-roam" at a produce site with wide, open aisles: no lanes, no central sequencer, just the model's own read on who has the right of way, decided robot to robot in real time. It worked. 40,000 merges, zero collisions, and robots got where they were going 18% faster because nobody was waiting on a turn that didn't need waiting for.

Same model, same confidence number, two completely different amounts of warning before it has to be right.

Renate wants that same free-roam mode live at Cadwell Grocery Distribution within twelve days, timed to a trade show demo, and to beat a rival who just announced their own "no geofencing" launch. Cadwell's aisles are narrow, 1.4 meters, with blind corners where two aisles meet. A robot rounding one of those corners has under a second between first seeing another robot and needing to know who goes first.

Here's the turn. The extra speed free-roam offers Cadwell isn't really what's on the table. What's actually on the table is a number nobody had checked in the one place it mattered. Draymark had quietly logged what the model would have decided at Cadwell's blind-corner merges for ten weeks, 3,200 of them, while the real marshal system still made the actual calls. Nobody had opened that log. When Zubin finally did, the model's "over 90% sure" calls were right only 79% of the time there, against 96% at the wide-open pilot site.

What I would leave alone: Cadwell's loading dock and staging yard are wide open, same shape as the pilot site, long sightlines, no blind corners. Free-roam there is already proven. Holding that up to wait on blind-corner data that has nothing to do with it would slow down the one part of the rollout that's already earned its yes.

The lesson: a model's confidence number is only as trustworthy as the ground it was measured on. You don't get to borrow a calibration number from one aisle shape and assume it holds in a narrower one, especially when the difference between the two is exactly the amount of time a robot has to react.

Now here is the same thing as a story

The short version above is what you'd actually say out loud in the room. Read this one for what it would have cost Draymark to find out the slow way.

The picking floor at Cadwell's regional distribution center gets loud around 4am, when the first trucks start loading and every aisle has two or three robots in it at once, gliding between racks of dairy crates and produce totes.

Zubin Kalantari has owned Draymark's pathfinding model for two years. He can read a merge log the way some people read a weather map, at a glance, for the shape of the thing rather than any one number.

The pilot went well enough that it stopped being a pilot in most people's heads. Free-roam ran for three months at a produce site with wide, easy aisles, and it just worked. Renate Odukoya, who runs field deployment, had been carrying that result into every client meeting since: zero collisions, eighteen percent faster, no lanes needed. She wasn't wrong to be proud of it.

Then a rival startup put out a demo video, robots weaving through a mock warehouse with no lane markings at all, and the caption said "fully autonomous, no geofencing, no waiting." Leadership wanted an answer fast. Renate picked Cadwell, the biggest account on the calendar, and set a twelve-day deadline tied to a trade show floor demo. Free-roam, fleet-wide, no lanes.

The rival's video and Renate's deadline landed the same week. What actually changed Zubin's mind arrived four days later, from a sister site nobody was watching that closely.

Four days into the twelve, word came in from Sedgewick Freight, a different Draymark client that had gone free-roam early under the same competitive pressure. A robot rounded a blind corner in a narrow aisle and clipped a picker's foot, not seriously, a bruise and a scare, but enough to trigger an incident review, a week with the whole fleet back on marshaled lanes, and a very uncomfortable call with Sedgewick's ops director.

Zubin didn't think Cadwell would follow the same script automatically. But he remembered something: months earlier, someone on his team had set up shadow logging at Cadwell, recording what the model would have decided at every blind-corner merge, while the real marshal system kept making the actual calls. It was meant to build a dataset for later. Nobody had ever gone back and read it.

He read it that afternoon. 3,200 blind-corner merges, ten weeks of them. He filtered for the calls where the model had said it was over 90% sure. At the pilot site, that confidence band was right 96% of the time, cleaner than the claim itself. At Cadwell, it was right 79% of the time. In the tightest corners, the ones with under a meter of visibility before the merge point, it dropped to 71%.

I want to say the problem was that free-roam didn't work. It worked fine, at the site it was tested at. That's not really the story. Nobody at Draymark had a rule that said "check the confidence number on the actual layout before turning it loose." They had a feeling that a strong pilot result travels, and it mostly does, right up until the geometry underneath it changes enough that the model's confidence stops meaning what everyone assumed it meant.

Neither of them was wrong about what they were holding. The deadline was real. So was the log. The question was which one got to decide first.

Zubin walked into Renate's office with the shadow log open on his laptop, not a warning, a number. He told her plainly: free-roam ships everywhere at Cadwell that looks like the pilot site, the loading dock, the staging yard, today, no delay. The dense aisles, the ones with blind corners, hold for eight weeks while the model gets recalibrated on real Cadwell data, with the old marshal system running underneath as backup the whole time. He named the number that would flip it: if the confidence gap at the tightest corners closes to under 5 points, he'd sign off on a phased pilot in one aisle before the eight weeks were even up.

Renate pushed back once, asked whether eight weeks would blow the trade show demo. Zubin said the open areas alone would still be a real demo, robots moving without lane markings through most of the floor, and that showing a rehearsed video of the blind corners would be a worse story than showing the real thing two months later. She agreed to split the rollout that same afternoon.

The recalibration work started that week: more blind-corner examples in the training data, a wider shadow-logging window, a bigger held-out sample to test against. Week zero, the gap at the tightest corners was 11 points. Week four, 8 points. Week eight, 4 points, under the line Zubin had set. Free-roam went live in Cadwell's dense aisles in week nine, marshaled fallback still running for the first two weeks as a safety net nobody ended up needing.

One version of this story ships free-roam everywhere on day twelve and spends the next month explaining an incident report to Cadwell's ops team. The other spends eight weeks closing an eleven-point gap and ships a version that's actually earned its confidence number. Same model, same deadline pressure, same trade show. The only thing that changed was whether the confidence number got checked against the ground it was actually going to run on.

What I'd tell myself, watching that shadow log sit unread for ten weeks: a number that's already sitting on your own server is not evidence until somebody actually reads it. It was there the whole time. It just hadn't been anybody's job yet.

PICK, for the day someone calls your no obstruction

Not a script for sounding confident. PICK only earns its keep if it makes you name, out loud, the one number that would actually change your mind, before anyone asks for it.

One path costs a scheduled eight weeks. The other already sent Draymark its real bill, from a different site, four days into the same deadline.

Four branches, one question: has the confidence number actually been checked on the ground it's about to run on, and does the gap clear the line.

The trade worth saying out loud: recalibrating on real blind-corner data before shipping cost eight real weeks nobody wanted to spend, against a twelve-day deadline that felt urgent for reasons that had nothing to do with Cadwell's actual aisles. That's slower, and it's worth paying, because the alternative, a robot that's sure of itself in exactly the spot it shouldn't be, only shows its true cost after someone's already standing too close to it.

And if you want to be sure it really works, try it somewhere else

Same four letters, a veterinary ER instead of a warehouse floor, and the missing check is a species, not a stretch of aisle.

Nyala Veterinary Emergency Network runs an AI triage model at intake: front-desk staff enter symptoms and vitals, and the model scores each animal's urgency to help route the sickest ones to a vet fastest. Mid flu season, with wait times climbing, someone proposes letting the model auto-discharge any animal it scores "low urgency" at 90% confidence or higher, no vet review, to free up staff for the animals that actually need one.

Same shape as Cadwell's aisles: a confidence number that was only ever proven on the majority case, about to be trusted on the minority one, unread.

Position: a no that names the exact subgroup the model was never checked on reads as judgment. A no about "AI in healthcare being risky" in general reads as obstruction, and it would be, since it doesn't point at anything specific. Impact: auto-discharge everything, and a genuinely sick cat gets sent home on a model that was mostly trained on dogs, since cats hide pain and distress far more than dogs do and made up only 18% of the training cases. Hold the whole initiative over general worry, and staff stay buried during flu season for no reason tied to actual risk. Cost asymmetry: pulling a held-out sample of cat cases and checking the model's calibration on them specifically costs a few days of a data scientist's time. A missed cat discharged as low urgency costs an animal's life and a lawsuit, not a symmetrical trade. Kill criteria: does the model's confidence match its real accuracy on cats specifically, within the same margin it already hits on dogs, about 95% agreement at the 90%-plus band. Today cats sit at 74%. Auto-discharge ships for dogs immediately, and holds for cats until that gap closes.

Swap the trigger and it still runs.
Speed: an interviewer caps you at ninety seconds. Skip straight to it: name the number that would flip your no, before anything else gets said.
Cost: no time to run the full recalibration before a real deadline. Fine, but test the highest-risk subgroup first, cats or blind corners, not a bigger pile of the same easy cases that already passed.
The model got better, for real: say a new version scores near-perfect on the overall average. Still don't skip the subgroup check. A clean average can hide one bad slice nobody separated out.

Where people run it wrong.
They say no with a feeling instead of a number, and then can't explain what would ever change their mind.
They say yes just to avoid the blocker label, and quietly hope the gap doesn't matter this once.
They cite general risk instead of the specific untested condition, which makes every future no sound the same, whether it's real or not.

How to use it live. If you're ever pushed to approve something on the spot, buy yourself a second by asking one plain question out loud: "has this been checked on the exact case we're about to ship it into, or just on the case that made the pilot look good." That question is the whole method, asked instead of stated.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0