Full Transcript for The org chart in the drawer

cases where the rules ran out. AI doesn't remove uncertainty – it produces a new kind. Fluent answers that are sometimes confidently wrong. Someone has to decide whether to trust them. 16 00:03:31.920 --> 00:03:33.480 Yes. That's it exactly. 17 00:03:33.960 --> 00:03:40.120 So the middle doesn't disappear. Its job changes – from routing exceptions to verifying outputs. 18 00:03:40.640 --> 00:03:59.140 That's what happened with us. I didn't cut the team leaders. I changed their job. Before, they answered questions from the dispatchers. Now they check what the planning system proposes. Every morning they look at the twenty routes the system flagged as unusual and decide whether it's genius or nonsense. 19 00:03:59.660 --> 00:04:01.500 And how often is it nonsense? 20 00:04:03.380 --> 00:04:17.040 Less often than in the first month. But last winter it routed a truck over a mountain pass that had been closed for snow. The system didn't know. Ahmed, the team leader in Grenoble, did – he grew up there. 21 00:04:17.680 --> 00:04:34.280 That's the local knowledge that never enters the system. And it's why the essay asks a question I find very serious: who verifies the verifiers, when AI-checked work is itself checked with AI? 22 00:04:34.900 --> 00:04:35.860 Try an answer. 23 00:04:36.580 --> 00:05:00.220 My guess: at some point there has to be a person who is responsible and who checks a sample by hand. Not everything – a sample. Like an auditor. In Luhmann's terms, the organization needs a place where a decision is attributed to someone. A chain of machines checking machines has no such place. It's a chain without an end. 24 00:05:00.840 --> 00:05:13.400 In practice, I'd say: every verification step needs a name on it. Not a department – a name. If something goes wrong, someone has to be able to say "I looked at this, and here's why I let it through". 25 00:05:14.000 --> 00:05:19.620 A name on the verification. That's more practical than my answer and says the same. 26 00:05:24.160 --> 00:05:35.680 Second essay. Many companies have announced in the last two years that AI has made them flatter. Fewer managers, wider teams. Niels, you're sceptical. 27 00:05:37.060 --> 00:06:01.180 I'm always sceptical. It's my job. But here the data agree with me. The largest survey so far asked almost six thousand top executives in America, Britain, Germany and Australia. About seventy percent of their firms use AI. And more than ninety percent of the executives said it had no effect on employment over the past three years. 28 00:06:01.780 --> 00:06:05.300 So the flattening is real, but the AI isn't the cause. 29 00:06:05.876 --> 00:06:46.036 The thinning of management started before generative AI was widely used. It looks more like cost-cutting after interest rates and wages went up. And AI is the story you tell about it. A research firm has a name for this – "AI washing". In organization theory we've had a term for decades: management fashion. Eric Abrahamson described it in the nineties – collective beliefs that a certain technique is at the forefront of progress. And Meyer and Rowan, back in seventy-seven: formal structures often signal what the environment considers rational, not what the organization actually does. 30 00:06:46.656 --> 00:06:47.736 A rational myth. 31 00:06:48.456 --> 00:06:59.496 Exactly. "We removed layers to move faster in the age of AI" turns a cut into an investment. Bad news for employees becomes good news for investors. 32 00:07:00.556 --> 00:07:20.916 I have to be honest here. Two years ago our board wanted exactly that sentence in the annual report. "AI-driven efficiency." We had cut two regional offices for completely different reasons – fuel prices, two lost contracts. And the AI sentence was more attractive than the fuel prices. 33 00:07:21.496 --> 00:07:22.236 Did you write it? 34 00:07:22.836 --> 00:07:29.636 I wrote a softer version. "Supported by new planning technology." Which was true for about ten percent of it. 35 00:07:31.756 --> 00:07:34.136 Ten percent true is a very good rate for an annual report. 36 00:07:34.716 --> 00:07:35.856 Is it all empty, then? 37 00:07:36.736 --> 00:07:52.376 No – and here I have to push back on Niels. Even if the story is exaggerated, it changes things. Once you've told the board you're flatter because of AI, you have to actually make the AI work. The fashion forces the learning. 38 00:07:52.876 --> 00:08:11.996 That's – fair. Abrahamson himself said fashions aren't merely cosmetic. They can be a collective learning process. And expectations act before evidence – executives expect AI to cut jobs slightly in the next three years, and they decide on that expectation. 39 00:08:12.556 --> 00:08:21.996 The essay asks whether, like in the nineties, the flattening will be followed by a quiet rebuilding of coordinating roles under new names. 40 00:08:22.676 --> 00:08:37.596 My guess – yes, absolutely, and I can tell you the new names. "AI operations lead". "Workflow owner". "Automation coordinator". It's the team leader, back again, with a new title and a laptop. 41 00:08:38.156 --> 00:08:56.756 In the nineties it was "process owner" and "change agent". Two economists, Rajan and Wulf, found that large firms really did remove layers back then – and the CEO got more people reporting directly to him. Flatter on paper, more centralised in practice. 42 00:08:57.396 --> 00:09:04.716 So we'll find the same middle managers in five years, under different names, and Niels will put the new chart in his drawer. 43 00:09:05.316 --> 00:09:08.136 Next to the old one. They'll look very similar. 44 00:09:13.096 --> 00:09:15.456 Third essay. Niels, Mintzberg. 45 00:09:16.096 --> 00:09:49.516 Henry Mintzberg, the Canadian management thinker. He said there are basically five ways to coordinate work. People adjust to each other directly – that's mutual adjustment. A boss tells them – direct supervision. Or you standardise: the processes, the outputs, or the

skills people bring with them. Every organization is a mix, and the mix defines what kind of organization it is. 46 00:09:50.136 --> 00:09:56.636 Logistics is pure standardised processes. Mintzberg would call us a machine bureaucracy. 47 00:09:57.896 --> 00:10:00.656 He would, and he wouldn't mean it as an insult. 48 00:10:01.236 --> 00:10:02.096 And AI? 49 00:10:02.796 --> 00:10:34.376 The essay suggests AI is a strange new standardisation of skills. There's the customer-support study – five thousand agents. With an AI assistant, the novices became about a third more productive. The model had learned what the best agents do – the phrasing that calms an angry customer – and delivered it at the moment of work. Skills, standardised. Not in people's heads, but in the tool. 50 00:10:34.976 --> 00:10:45.816 And with AI, customers asked less often to speak to a manager. I saw that in our customer service too. Fewer escalations. Less direct supervision. 51 00:10:46.356 --> 00:11:06.556 But here's the darker side. In the Copilot study, developers with AI didn't only coordinate less with their managers. They collaborated less with each other. Mutual adjustment – the informal talking that holds an innovative team together – goes down when everyone talks to a machine instead. 52 00:11:07.136 --> 00:11:20.816 I've noticed that in the office. People used to shout across the room: "Does anyone know how to handle the Swiss customs form?" Now they ask the assistant. It's faster. And the room got quieter. 53 00:11:21.396 --> 00:11:28.856 And the person who knew the answer no longer gets asked. Which means nobody knows who knows what anymore. 54 00:11:29.456 --> 00:11:30.416 Is that a problem? 55 00:11:31.136 --> 00:11:49.156 It became one. Last spring, the assistant gave a wrong answer on a new customs rule. Before, someone would have shouted, and Marie from the Basel office would have said "no, that changed in March". This time nobody shouted. We lost two days of shipments. 56 00:11:49.836 --> 00:11:58.816 That's the point of the essay. When best practice lives in a model, someone must curate it. Who corrects the model when Marie knows better? 57 00:11:59.376 --> 00:12:11.356 Now Marie does. She has a formal role – she reviews the assistant's customs answers every week. It's the first time in twenty years anyone wrote her expertise into a job description. 58 00:12:11.776 --> 00:12:17.436 That's quite beautiful, actually. The machine made her knowledge visible by getting it wrong. 59 00:12:21.992 --> 00:12:25.832 Fourth essay. Niels, a very old question. 60 00:12:28.592 --> 00:12:57.352 Nineteen thirty-seven, Ronald Coase, a young British economist. He asked: if markets are so efficient, why do firms exist at all? Why not hire everyone for every task on the market? His answer: because using the market has costs. Finding a supplier, negotiating, writing contracts, checking that they're kept. When those costs are higher than organizing it internally, you build a firm. 61 00:12:57.972 --> 00:13:03.652 And if AI agents can find suppliers, negotiate and monitor contracts cheaply – 62 00:13:04.332 --> 00:13:15.112 Then, says the thesis, the firm shrinks. Some people already talk about the "firm of one" – a single person with an army of AI agents. 63 00:13:16.552 --> 00:13:19.772 I'd like to see a firm of one run two hundred trucks. 64 00:13:20.372 --> 00:13:56.652 And there's the antithesis, with the same economics. Two researchers, Hadfield and Koh, point out that the classic limits on firm size – communication, the difficulty of passing on tacit knowledge, monitoring people – are human limits. AI agents might coordinate more efficiently inside a firm, too. So firms could also grow. And the evidence so far leans that way: firms that invested more in AI grew faster, mainly through new products, and the growth was concentrated among the large firms. 65 00:13:57.292 --> 00:14:18.772 That matches what I see. We tried outsourcing some planning to a freight platform with AI matching. It was cheap – and it didn't know our customers. It didn't know that a certain chocolate factory needs deliveries before six because their loading dock floods at seven when it rains. That's in Ahmed's head and in our data. Not on the platform. 66 00:14:19.372 --> 00:14:41.432 That's the synthesis. What stays inside is what is specific – data, relationships, local knowledge that's worth more inside your company than outside. And accountability. If an AI agent on a platform makes a mistake, who's responsible? There's no registration, no liability infrastructure for agents yet. 67 00:14:42.032 --> 00:14:46.012 So we brought the planning back in. More expensive. Better. 68 00:14:46.692 --> 00:14:52.112 That's the first time I've heard a COO say "more expensive, better" without apologising. 69 00:14:54.772 --> 00:14:55.872 Don't tell my CFO. 70 00:15:00.452 --> 00:15:04.612 Last essay of the episode, and the title says it all. Niels? 71 00:15:05.312 --> 00:15:34.672 In nineteen ninety, the economic historian Paul David asked why computers weren't showing up in productivity statistics. And he compared them with electricity. Electric motors reached factories in the eighteen-eighties. But for decades, productivity barely moved. Because factories simply replaced the big steam engine with a big electric motor and kept everything else – the same layout, the same shafts and belts running through the building. 72 00:15:35.152 --> 00:15:35.652 And then? 73 00:15:36.212 --> 00:15:55.032 Then, a generation later, they understood that with electricity you could put a small motor at every machine. You could arrange the factory around the flow of work instead of around the power shaft. One-storey buildings, new layouts, new jobs. And only then did productivity jump. 74 00:15:55.632 --> 00:16:09.692 That's – painfully familiar. We did exactly that. We bought an AI planning tool and put it on top of our old process. Same handovers, same meetings, same

Excel sheets in between. And the gains were tiny. 75 00:16:10.332 --> 00:16:11.072 What changed? 76 00:16:11.592 --> 00:16:36.532 We rebuilt one process end to end – night dispatch, where I started. We asked: if the system can see all depots at once, why do we still have dispatchers sitting by region? So we reorganised by problem type instead – breakdowns, delays, customer changes. It took eight months. Two people left because they hated it. And then the night shift became the calmest part of the company. 77 00:16:37.052 --> 00:16:50.332 That's the J-curve, as economists call it. First you invest – in redesign, training, new processes – and productivity even dips. Then, if you've rebuilt the factory, it rises. 78 00:16:50.912 --> 00:16:54.992 The dip was real. My board asked me every month why we were slower. 79 00:16:55.532 --> 00:17:11.012 But I have to be honest – the J-curve is also a very convenient story. If returns are low, we're in the dip. If they rise, the curve is turning. A theory that explains every outcome is hard to test. For AI, it's still a forecast. 80 00:17:12.772 --> 00:17:15.772 That's very Danish of you. Ruin my success story. 81 00:17:16.532 --> 00:17:26.912 I'm protecting it. If you only believe in the curve, you'll wait forever. You had evidence – a calmer night shift. That's better than a curve. 82 00:17:27.312 --> 00:17:33.352 The essay asks what distinguishes the minority of organizations that actually cross the gap. 83 00:17:33.892 --> 00:17:52.992 My guess, from the inside: they start with one ugly, specific process and rebuild it completely – not a tool for everyone. And someone senior is willing to be slower for six months. Most boards don't survive the dip. They switch to the next tool before the first one paid off. 84 00:17:53.432 --> 00:18:09.592 I'd add one thing from the theory: they change the structure and the technology together. That was true for electricity, for computers, and I suspect it's true for AI. The firms that only buy licences stay on the steam engine. 85 00:18:09.832 --> 00:18:11.672 With an electric motor attached. 86 00:18:11.972 --> 00:18:14.272 A very expensive electric motor. 87 00:18:15.052 --> 00:18:59.312 Let me gather this. You agree that AI takes over part of what the hierarchy used to do – answering the routine exceptions – and that the middle doesn't vanish but changes: from routing questions to checking answers, with a name on every check. You agree that much of today's flattening is a story told about other cuts – but Claire thinks the story forces real learning, and Niels predicts the old roles will come back under new names. You agree that what stays inside a firm is what's specific to it. And that the gains only come when the factory is rebuilt, not when a new motor is bolted onto the old shaft. 88 00:18:59.972 --> 00:19:02.812 That's a lot of agreement for a first episode. 89 00:19:03.312 --> 00:19:07.212 We'll fix that next time. Decisions are my favourite fight. 90 00:19:08.872 --> 00:19:26.352 Then here's the question I'd like to leave with you, listening. Think of the last time your organization introduced a new tool. Did anything else change – the meetings, the roles, who decides – or did you bolt it onto the old chart? 91 00:19:27.332 --> 00:19:48.152 The essays behind this episode, with all the sources, are on ideation.zone. Next time we go to the heart of Niels' theory: decisions. Who decides when the machine proposes – and who's responsible when it goes wrong? Niels, Claire – thank you, and welcome. 92 00:19:48.772 --> 00:19:50.632 Thank you, Elisabeth. Tak. 93 00:19:51.152 --> 00:19:55.812 Merci – this was a pleasure. And Niels, keep that drawer closed.