44% of UK workplaces use AI every day. That figure comes from Skills England research published on 10 June. Read the next line of it and the picture changes: most of those workplaces can't say what it has improved.

I think the explanation is simpler than the AI conversation makes it sound. A tool can only work with what it's been told. A model trained on the whole internet still knows nothing about your site. It doesn't know why your night shift runs the goods-in check in a different order to the day shift, or which order is correct. If nobody has written that down, the tool can't use it, and neither can the new starter you hired last month.

So the companies getting something out of AI this year have one thing in common. They already knew their own work well enough to point the tool at a real part of it. The rest bought the same software and got a faster way to write emails.

Three stories this fortnight show what knowing your own work actually looks like.

Lazer Logistics built an AI coaching tool, named it Uncle Phil after its COO Phil Newsome, and put it across about 750 sites. It reads telematics, in-cab video, inspection reports, labour plans and yard workflows, then coaches site managers in the moment, the way Newsome's 36 years would if he could be everywhere at once. It launched in April, so it's an example rather than this fortnight's news.

My take: The interesting decision is what they fed it. Newsome has spent 36 years learning when a yard is about to back up, which driver report to trust, when a customer complaint is going to escalate. To build the tool, somebody had to get that out of his head and write it down. Most companies have a Newsome. Very few could tell you, on paper, what he does that the others don't. That writing-down is the slow, unglamorous job, and it decides whether an AI project has anything real underneath it. The software is the cheap part.

Walmart now uses machine-vision forklifts to unload trailers in its distribution centres. The part I'd point at sits before the robot. The company builds each pallet using store-level data, so the goods a shop needs first come off the lorry first. Walmart calls the worker's job "conductor" now (Fortune, 12 June).

My take: They designed the packing job around the unloading job. The pallet gets built for the convenience of the person two steps down the chain. I've watched the opposite cost real hours for years. A load goes out correct on every system and miserable to receive, because the person who built it was measured on getting it out the door, and what happened at the other end was somebody else's morning. You don't need a robot to copy this. Ask one team what the team after them always has to redo. The answer is usually free, and usually already known by everyone except the people who could change it.

A manufacturing survey from Octave on 3 June found 47% of firms now use AI in quality work, up from 33% a year ago. The same survey found 85% of them said staff and skills shortages were dragging their quality down (2,263 managers across the US, UK and Germany; it's sponsored research, so read the framing accordingly). Adoption is going up. The experienced floor is getting thinner. Quality is slipping anyway.

My take: When people say "AI skills gap" they usually mean their staff can't drive the new tools. The Octave numbers point somewhere else. The skill that's leaving is the operational knowledge that used to pass from an experienced operator to a green one over 2 years of standing next to them. Nobody has 2 years now. AI gets sold as the cover for that loss. It can only pass on what's been made explicit, and almost none of this has ever been written by anyone. Before an "AI upskilling" line goes in the budget, I'd spend the cheaper money first and capture what your experienced people know while they're still on the payroll. Do it the other way round and you've paid to automate knowledge you never wrote down.

Audit the distance between your SOP and your best person.

Pick one task where getting it wrong is expensive. A release decision, a safety step, a handling procedure. Sit your most experienced operator down with the written version and ask them to read it out loud and stop wherever the document and the real job stop agreeing.

You'll get more than you expect. The step they do in a different order, for a reason. The check that isn't on the page but they'd never skip. The line they'll tell you to ignore because it went out of date 2 reorganisations ago. The thing they look at first that the SOP mentions last.

Everything they stop on is the gap between your written operation and your real one. That gap is the answer to this whole issue. It's why a new starter takes 2 years instead of 2 months. It's the knowledge a Walmart-style redesign needs before it can sequence anything. It's what an Uncle Phil tool runs on. You can't buy your way across it. You can map it, one task at a time, and it starts with a 30-minute conversation this week.

The Shift Report goes out by email too, if LinkedIn isn't where you do your proper reading, or you'd rather not depend on the algorithm: https://vinces-newsletter-5a6441.beehiiv.com/. Same fortnightly cadence, same ratio of signal to noise.

Reply if something landed, or if it didn't. I read everything.

Last thing. SOPwise, the thing I'm building, lives in that gap between the document and the person. Lazer spent a fortune mapping it for one COO across 750 sites. My bet is you shouldn't need a fortune, and that the knowledge separating your best operator from your newest one can be captured and made usable without a transformation budget or another job landing on someone's desk. If the read-out-loud exercise turns up a gap that's costing you, that's the conversation I want. Reply and tell me what it surfaced.

Issue #7 in two weeks.