Why your customers now get told when AI is involved, and your shift does not

Since 2 August, a chatbot has to tell a customer it is not a person. AI-generated content has to carry a mark. The Commission’s AI Office and the national regulators started enforcing that day.

Nobody has to tell your night shift anything.

The duty points outwards. It protects the person at the other end of a conversation with a business. It stops at the gate.

Inside the gate, a supervisor works from a ranked exception list and a new starter reads a briefing note that somebody produced last Thursday. Neither of them is entitled to know which parts of those were written, ordered or rewritten by a model, and in most operations neither could find out if they asked.

People check differently depending on what they think produced the thing in front of them.

An experienced operator reading a colleague’s handover note reads it for the things that colleague usually gets wrong. He has worked next to the man for 6 years and he applies that without thinking. Hand him the same note drafted by a model and none of what he knows about his colleague helps, because the errors sit somewhere else: the confident sentence about a piece of equipment that was replaced in 2023, the step that is correct at the other site, the warning that was never in the original because nobody wrote it down.

He would look for those, if he knew. Nothing on the page tells him to.

You already label everything else. A stock adjustment carries a user ID. A work order has a status history. A quality record shows who released it and when. We built all of that because knowing who did something changes how the next person treats it. The ranked exception report that arrives every morning with no author is the odd one out, and it is the one increasingly shaping what the shift does first.

The second half of this surprised me.

The same 2 August was the date the AI Act’s high-risk rules were due, the ones covering systems that allocate tasks, monitor performance and inform decisions about people’s jobs. Those rules include telling workers when they are subject to such a system. A regulation that came into force on 27 July moved them to 2 December 2027.

So the disclosure aimed at your customers arrived, and the disclosure aimed at your people slipped by 16 months.

You do not have to wait for it. Telling a supervisor which parts of his morning report a machine produced costs nothing, needs no software, and is the difference between a person reviewing a document and a person receiving one.

Three things worth knowing

1. The transparency rules are live, and your customers will start asking.

The European Commission confirmed on 31 July that enforcement began on 2 August, covering disclosure for interactive AI systems, machine-readable marking for generated or altered content and labelling for deepfakes. Cooley’s alert of 3 August adds the detail worth having: the guidelines were adopted on 20 July, fines run to €15m or 3% of worldwide turnover, the duties apply to systems already on the market, and providers of existing generative systems have until 2 December 2026 for the marking obligation. It applies to anyone placing AI on the EU market or whose AI outputs are used in the EU, which includes plenty of UK suppliers.

2. Half of logistics leadership teams say they are not ready to use AI in decisions.

The Adecco Group’s Where Talent and Technology Collide, reported by MHL News on 6 August, found 51% saying their leadership team is not well-prepared to use AI tools in decision-making, 42% of organisations delivering no formal training to reskill or upskill people in AI, and 23% with no AI policy at all. 2% came out as “future-ready”. It is a staffing company’s report on its own market, so read the framing accordingly.

3. A chemical logistics firm was fined £425,000 because a written instruction never reached the ramp.

Peter Hutchinson, 60, fell 1.5 metres from a mobile loading ramp onto a concrete floor and died. HSE found Bertschi UK Ltd had failed to plan the loading activity properly and had not fitted a handrail, despite the manufacturer’s instructions clearly requiring one. Full release.

One thing to try this fortnight

Take the last 5 things your team acted on that somebody else produced.

The morning exception list. A revised work instruction. A translated safety notice. A customer complaint summary. A shift handover note.

For each one, answer 2 questions yourself, from the record alone, without asking the author:

1. Which parts were written by a person, and which were produced, ranked, translated or rewritten by a tool?
2. If it turned out to be wrong next Tuesday, what in the record would tell you where the error entered?

You will get one of 3 results. You can answer both, in which case you are ahead of almost everyone. You can guess but not evidence it. Or you cannot tell at all, which is the common outcome and the useful one.

Then do the small version of the fix. On the next 5 outputs, add one line at the top: what produced this, from which systems, at what time, and who read it before it was sent.

Ask the people receiving them whether the line changed what they checked. That answer is the whole point of the exercise, and you will have it inside a fortnight.

Final thought

A rule arrived on 2 August that makes software introduce itself to the public. The rule that would have made it introduce itself to the people running your site is now 16 months away. In between sits an operation where the instructions arrive every day with no author on them.

That’s close to the problem I’m building with SOPwise: keeping approved operational knowledge, its source and its review state visible at the point where somebody needs an answer or a training update.