On 22 May, Harvard Business Review published findings from a seven-week Accenture study: 85 frontline workers, video diaries, Australia, UK, US. More than three-quarters were dissatisfied with how AI had been introduced to them. The manufacturers pulling ahead did three things: involved workers in mapping how roles would change, trained in the flow of real work, measured humans and machines performing together.

All of it requires asking the people doing the job before the decision is made.

So here's the question to carry into every demo, every pilot review, every "we're rolling out X next quarter" this year. Are we building this with them, or for them? If you can't point to one frontline person who shaped the decision, you already have your answer — and your most likely failure mode. 

On 29 May, the Guardian reported on a TUC-backed report from the IPPR calling for workers to have more influence over how AI gets introduced at work. The survey behind it found 20% of workers say AI has made their working life better, 21% say it's made it worse, and 4% believe they've already lost a job to it. The report's useful distinction: the same technology can improve a job, degrade it, or remove it — and which one you get depends on how it's brought in and who had a voice in the design.

My take: The same AI tool can support a supervisor or quietly turn into another layer of monitoring. The same checklist can sharpen adherence or collapse into box-ticking. Which one you get depends almost entirely on whether anyone asked the people doing the job before it landed. A think tank report won't give your team a say — asking them will. The ones who do that first will be ahead of whatever guidance eventually arrives.

On 19 May, Food Safety Magazine published findings from Registrar Corp's 2026 Global Food Safety Training Survey - 1,226 professionals across more than 3,000 facilities. Facilities with above-average training programmes were found to be twelve times more likely to maintain strong, consistent adherence to their protocols, and five times more likely to prevent incidents before they happen. The survey also flagged low uptake of learning management systems and AI, which leaves most operations with little visibility into whether their training is actually working. (It's sponsored research, so read the framing with that in mind — but the link between training quality and floor behaviour is one most of us already know in our bones.)

My take: Twelve times more likely to hold the line on adherence. That's a control metric, not a wellbeing one — the difference between an audit you walk into calmly and one you dread. Weak training rarely announces itself. It shows up as small variation: a missed check, a copied shortcut, a sign-off that says trained followed by a blank look when you ask what the step means. The knowledge to fix it usually already exists in the business. The hard part is converting it into something people can use without creating another admin job for someone who's already flat out.

If you want to see the principle in working clothes, Supply Chain Dive reported on 22 May that the US grocer Albertsons has put an AI tool into its distribution centres to grade the quality of strawberries and grapes. A warehouse worker photographs the punnets on a tablet, and the tool scores them against the company's own quality standards — deep red strawberries high, discoloured ones lower. Note how the chief supply chain officer described it: they built it "to support our team of talented quality inspectors," and early results show it's "incredibly helpful in increasing the consistency of quality rating."

My take: This is the kind of AI example I pay attention to, and not because of the technology. There's a real product, an existing quality standard, and a person who already has to make that judgement on shift. The tool was built to support that person and tighten the consistency of a decision that already gets made — not to replace the inspector or invent a new process. That's "with them" in practice. It also quietly makes the case for why your SOPs matter: a tool like this only works because there's a clear standard for it to grade against. A vague standard is just noise. A clear one, linked to the checks and the people, is something an AI can actually work with.

This is the thread running through everything above, and it's worth saying plainly. The HBR article found trust rose when workers helped map their own roles — in other words, when the real process was made explicit instead of left in people's heads. AI needs that context to be useful. If the official system says one thing and the actual work happens somewhere else — a spreadsheet, a side message, the supervisor everyone asks — the tool only sees half the job.

My take: This is the boring bit, and the bit that pays. Clear procedure, clear decision, clear evidence, clear owner. None of that has ever gone out of fashion, and AI raises the stakes on it rather than removing it. A confident, wrong answer from a tool fed on messy inputs is more dangerous than no answer at all. Before you ask what AI can do for a process, it's worth knowing how that process actually runs.

A small act of knowledge capture. No software, no budget.

Find one experienced operator or supervisor — the person everyone quietly depends on — and ask them this:

Write down exactly what they say. That answer is judgement that currently exists in one head and walks out of the building at the end of every shift.

Then turn it into three things:

That's the whole exercise. Do it once and you've captured something useful. Do it ten times across your most-depended-on people and you've started to build a real picture of how the work actually gets done — which, not by coincidence, is exactly the picture any AI tool would need before it could help. Diagnosis first. The order matters.

The Shift Report is also going out by email. If LinkedIn isn't where you do your proper reading — or you'd rather not depend on the algorithm — you can subscribe here: https://vinces-newsletter-5a6441.beehiiv.com/. Same cadence. Same ratio of signal to noise. Just in your inbox.

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

Last thing: I'm building SOPwise — it turns the SOP library into something that actually trains the people using it. It's a with them tool by design: it starts from the way your people already work and the knowledge already in the business, not the way a system wishes they did. The Albertsons example above only works because there's a clear standard for the AI to grade against — and that's the same bet here. If the "how do you know something's wrong before it shows up on anyone's report?" answer is the kind of thing that exists in heads on your site and nowhere else, that's the exact gap I'm working on. Reply and tell me about it.

Issue #6 in two weeks.