Glossary

Terms from the course, the newsletter and this website · as of October 2026

These are the terms that keep coming up in my writing: in the course “The Silent Power”, in the newsletters, on this website and in my comments on LinkedIn. Course and newsletter are written in German. The terms are grouped into four areas. Each gets a short explanation and a piece of history, the AI terms also what's changing right now, the others how I use the term. The history is here on purpose: a lot of what looks new today has a long backstory, and that backstory often shows more clearly what's really changing.

AI: Artificial intelligence (AI) · Language model · Generative AI · Predictive AI · Embedded AI · AI assistant · AI agent · Hallucination · Model and application · Context · Training on company data · Open weights · Shadow AI · Shadow tool · Decision in all but name · Human in the loop · AI literacy · Configuration power · The four zones · Pace rule

Organizational design: Organizational design · Structure (Ordnung) · Operating system · Organizational constitution · Decision path, approval rules, decision rights · Review point · Handoff · Reversibility · Binding layer · Self-organization

Management: Management · Management system and “unmanaged” · Insight · Clarity · Impact · Control reflex · Trust architecture · Checking load and the cost triad · Leftover load · Metric · Bottleneck · Governance · Accountability · In charge and responsible

Leadership: Leadership · Leadership system · Mission · Judgment · Challenge · Breadth of view · Participation · Commitment · Strain

AI

What AI is, how it works and the routes by which it gets into companies.

Artificial intelligence (AI)

In short: An umbrella term for programs that do tasks which used to need human judgment: recognizing, assessing, wording, deciding. Today it mostly means programs that learned their rules from large amounts of data. In the past, a person wrote every rule by hand.

Where it comes from: The term comes from a 1955 proposal for a summer meeting at Dartmouth College the following year, submitted by John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon. From the 1970s, researchers built expert systems with hand-written if-then rules, and in the 1980s companies used them widely. Twice, high expectations were followed by years of cut funding that were later called “AI winters”. From 2012, machine learning with neural networks took over, after one such network (AlexNet) won a major image recognition competition by a wide margin.

What's new (as of October 2026): Since ChatGPT arrived at the end of 2022, “AI” in everyday use almost always means generative AI. In companies, though, many older forms are still at work, such as forecasts and image recognition. What's mainly new is the route: AI increasingly comes as part of existing software (embedded AI) and as an agent that acts on its own.

In the course: weeks 1 and 2.

Language model

In short: A language model builds text piece by piece from probabilities. It splits text into small units, usually parts of words, and calculates for each possible next unit how likely it is to follow the text so far. Then it picks one, adds it and calculates again until the answer is complete. It learned these probabilities from very large amounts of text. It works with what sounds likely; whether something is true doesn't come into the calculation at all.

Where it comes from: In 1913 the mathematician Andrey Markov counted how often vowels and consonants follow one another in the first 20,000 letters of Pushkin's verse novel “Eugene Onegin”. It was the first time he applied his probability chains to a real text. In 1948 Claude Shannon used such frequencies to generate word chains that already sounded almost like English. For decades this was good enough for spell checkers and speech recognition, and hardly for longer texts. The breakthrough came in 2017 with an architecture called the Transformer, described by researchers at Google; it keeps the whole preceding text in view during the calculation. GPT (from 2018), Gemini, Claude, Mistral and Llama build on it. ChatGPT made this available to everyone at the end of November 2022, after the model had also been tuned toward helpful answers using human ratings.

What's new (as of October 2026): Since fall 2024 there have been models that work through intermediate steps before they answer, often called “reasoning models”; the first widely available one was OpenAI o1, released as a preview in September 2024. They solve math and logic tasks much better, though they take longer and cost more. Language models also increasingly process images, audio and video and call tools themselves, such as a search or another program. That's the basis for agents. The basic principle changes little, though: the intermediate steps of a reasoning model also come from probabilities.

In the course: week 2.

Generative AI

In short: AI that creates new content: texts, images, code, summaries, voices, videos. It's usually based on language models or comparable models for images.

Where it comes from: As early as 1966, Joseph Weizenbaum at MIT presented ELIZA, a program that imitated a conversation partner using simple text patterns. Weizenbaum was alarmed at how seriously people took the program and became one of the early critics of AI. In 2014 Ian Goodfellow and colleagues presented a method in which one network generates images and a second tries to spot the fakes. OpenAI showed images from text descriptions in 2021 with DALL·E. Such programs became usable for everyone in 2022 with DALL·E 2, Midjourney and Stable Diffusion, and text for everyone at the end of 2022 with ChatGPT.

What's new (as of October 2026): Video and voice are reaching a quality that's hard to tell apart from real recordings. That also means more fakes, and the EU AI Act provides for labeling duties for certain AI-generated content; what applies when is covered in week 7. In companies, generative AI now mostly comes inside programs that were already there (embedded AI).

In the course: weeks 1, 2 and 3.

Predictive AI

In short: AI that estimates, classifies or recognizes: How much will we sell in March, is this payment suspicious, is this part faulty? It learns from past data, and its errors can be counted as a hit rate.

Where it comes from: Its roots lie in statistics, long before anyone talked about AI. In the US, the FICO score has rated creditworthiness since 1989, credit card companies have used learning networks against fraud since the early 1990s, and in 2002 the programmer Paul Graham made learning spam filters widely known with his essay “A Plan for Spam”. In manufacturing, cameras using such methods have been sorting out faulty parts for years.

What's new (as of October 2026): Predictive and generative AI are often combined: the system delivers the forecast together with a text that explains it. The old problem remains: when the world changes, the learned patterns no longer fit. Anyone who planned in spring 2020 with a sales forecast based on previous years has seen this.

In the course: weeks 2 and 3.

Embedded AI

In short: AI built into standard software: the office suite, the ERP, the CRM or the HR software. It comes with the license or the update, without a rollout project of its own, and so it's the only kind of AI whose introduction nobody in the company has to decide on.

Where it comes from: Everyone knows the first forms: autocorrect and spelling suggestions. In 2018 Gmail began suggesting whole sentence endings with “Smart Compose”. Generative AI moved into the programs everyone uses daily with Microsoft 365 Copilot (for businesses from November 2023), SAP Joule (announced in 2023) and Gemini in Google Workspace; in January 2025 Google included Gemini in all Business and Enterprise plans.

What's new (as of October 2026): On July 1, 2026, Microsoft raised list prices for Microsoft 365 business plans, citing AI among other things, including agents in Word, Excel and PowerPoint. This makes embedded AI the most common route by which AI enters companies.

In the course: weeks 1 and 2.

AI assistant

In short: A language model with an input box inside a program that summarizes, drafts or searches on request, such as Copilot at Microsoft, Gemini at Google or Joule at SAP. An assistant usually acts only when told to; the line between it and an agent is blurry.

Where it comes from: Siri (2011) and Alexa (2014) were voice assistants with fixed, pre-programmed skills: setting alarms, giving the weather, playing music. Since 2023 there have been assistants based on large language models that can phrase things freely.

What's new (as of October 2026): Assistants are getting access to inboxes, calendars and documents, so they see more and more context, and they're increasingly allowed to perform steps on their own.

In the course: week 1.

AI agent

In short: A system that takes steps itself: it books, orders, sends, changes records or calls other programs, often in a chain where a person only looks again at the very end. What an agent may do is set by its permissions.

Where it comes from: In AI research, anything that perceives its environment and acts in it has long been called an “agent”; in 1995 Stuart Russell and Peter Norvig built their entire standard textbook on this concept. Companies had predecessors: software robots that have imitated fixed click sequences since the 2000s. In 2023 developers began letting language models work through their own task lists in loops, for example with AutoGPT in March 2023.

What's new (as of October 2026): Since late 2024, models can operate screens and browsers (Anthropic in October 2024, OpenAI with “Operator” in January 2025), and an open standard, the Model Context Protocol (from Anthropic, November 2024), connects them to company software. The research firm Gartner, itself a seller of guidance in this market, expects task-specific agents in around 40 percent of enterprise applications by the end of 2026, and at the same time expects more than 40 percent of agent projects to be scrapped by the end of 2027. There's also a new risk: agents read outside texts along the way, and a doctored email can redirect them. In June 2025 such a flaw in Microsoft 365 Copilot became public (EchoLeak); Microsoft had already closed it by then.

In the course: weeks 2 and 9.

Hallucination

In short: A language model makes up details that sound plausible: sources, quotes, figures, court rulings. This follows directly from its basic principle, because it forms likely sentences and doesn't check for truth. Some experts prefer the term “confabulation”, since the model doesn't perceive anything it could see wrongly.

Where it comes from: Research on machine translation and image captioning was already using the term in the late 2010s. It became widely known in 2023, when two New York lawyers in the case Mata v. Avianca cited rulings that ChatGPT had invented; the federal court in New York fined them and their firm $5,000. In 2025 Deloitte Australia refunded part of its fee to a government department because a report contained invented references and an invented court quote.

What's new (as of October 2026): Web search and source citations make such errors rarer, though they don't go away. For some reasoning models, the providers' own tests even measured higher fabrication rates than for their predecessors.

In the course: weeks 2 and 3.

Model and application

In short: The model is the computing core, such as GPT, Gemini, Claude, Mistral or Llama. The application is the program you use it in. A provider can swap the model behind an application without the interface changing.

Where it comes from: For a long time, model and program belonged together. Since providers began renting out their models through interfaces, OpenAI from 2020 with GPT-3, many companies build their applications on other firms' models.

What's new (as of October 2026): Models are replaced every few months, and even an update can change their behavior. At the end of April 2025, OpenAI rolled back a GPT-4o update after a few days because the model was conspicuously telling users what they wanted to hear. Whether a provider may swap the model is therefore a question for the contract.

In the course: weeks 2 and 9.

Context

In short: Everything a system sees at the moment it answers: the question, attached documents, emails, calendars, earlier conversations. The more context a system gets, the more useful it becomes, and the more it knows about the company.

Where it comes from: In 2020, GPT-3 could take about 2,000 word pieces into account at once: a few pages.

What's new (as of October 2026): Since 2024, leading models process a million word pieces and more, the equivalent of several books at once (Google Gemini 1.5, February 2024). On top of that comes a “memory”: assistants remember details beyond individual conversations.

In the course: week 2.

Training on company data

In short: Whether a provider may use inputs and documents to improve its models is settled by the contract, and the technology has no say in it. For consumer accounts it's often switched on by default, for business accounts usually excluded. So for employees' private accounts, the consumer terms generally apply.

Where it comes from: A large share of the training data for large language models comes from the public internet. That's disputed; in December 2023 the New York Times sued OpenAI and Microsoft over the use of its articles.

What's new (as of October 2026): Providers keep changing their default settings. If you want to know what applies to your own accounts, you have to read the current terms; a one-off check is hardly enough.

In the course: weeks 1 and 2.

Open weights

In short: For some models, the learned numerical values, the “weights”, can be downloaded, so you can run the model on your own computers. That reduces dependence on the provider, and it requires operating skills of your own.

Where it comes from: Meta made its Llama model available to researchers in February 2023 and released Llama 2 for broader use in July 2023. The Paris company Mistral published its first open-weights model in September 2023.

What's new (as of October 2026): In January 2025 the Chinese model DeepSeek R1 showed that open models can come close to the best closed ones. In August 2025 OpenAI also released open-weights models. When it comes to how dependent a company becomes on a few providers, open models are one of the few alternatives.

In the course: weeks 2 and 9.

Shadow AI

In short: Employees use AI for their work without approval, usually through private accounts. There's often a good reason behind it: the official route is too slow, or there isn't one.

Where it comes from: The term is derived from “shadow IT”: the private devices, cloud storage and programs that employees have used around the IT department since the 2000s.

What's new (as of October 2026): In a global survey by the University of Melbourne with KPMG of around 48,000 people, more than half of the employees surveyed said in 2025 that they hide their own use of AI at work. According to a January 2024 decision of the Hamburg Labor Court, using private AI accounts doesn't trigger co-determination (Mitbestimmung), the works council's right to a say, because the employer doesn't control the tool at all. Honest numbers come only to those who promise beforehand that an honest answer will have no consequences.

In the course: week 1.

Shadow tool

In short: A tool that employees have built themselves alongside the official systems, such as a spreadsheet with its own formulas, a macro or, more recently, a workflow with AI. It often describes what the official system lacks more precisely than any requirements list.

Where it comes from: Such homemade tools are part of shadow IT. German banking supervision calls them “individual data processing” (individuelle Datenverarbeitung): applications that staff in a business unit develop or run themselves. The requirements for IT systems apply to them there too, graded by how important they are for operations (BaFin, MaRisk AT 7.2, Circular 05/2023).

What's new (as of November 2026): With AI, far more people build such tools today than before, often in an afternoon. The rules then live in instructions to a language model that often nobody except their author knows.

Decision in all but name

In short: A system formally only prepares a decision and still shapes it, because its selection determines what the decision-maker gets to see at all: which ten applications, which five emails, which three offers. I use the term in “Cheap to Make, Costly to Check”.

Where it comes from: Political science has known the idea since 1962. Back then, Peter Bachrach and Morton Baratz described “two faces of power”: alongside the visible power to make decisions stands the silent power that determines what comes up for decision at all. A year later they called what never even makes it onto the agenda “nondecisions”.

What's new (as of October 2026): In 2025 a research group at the University of California, San Diego found that AI summaries of reviews changed the framing of the source in about a quarter of cases; in a small experiment with seventy participants, 84 percent chose a product after the summary and 52 percent after the original reviews (Alessa et al., 2025). Legally, the preparation stays in the shadows for now: the EU AI Act exempts systems that only perform a “preparatory task” from classification as high-risk (Art. 6(3) AI Act).

In the course: week 1.

Human in the loop

In short: The rule that a person sees and approves a system's output before it takes effect. Many AI policies rely on it. Whether the person really checks depends on whether they have the time, the knowledge and the right to say no.

Where it comes from: The expression comes from control engineering. For decades the field has distinguished between a person who intervenes in a control loop and one who merely watches. In 1983 the psychologist Lisanne Bainbridge described in “Ironies of Automation” how well-automated plants in particular make people forget what they need to be able to do when something goes wrong.

What's new (as of October 2026): A 2024 MIT meta-analysis of 106 experiments found that humans and AI together perform worse on decision tasks, on average, than the better of the two alone (Vaccaro, Almaatouq and Malone). A proposal that's almost always accepted therefore works more like a signature than a check.

In the course: week 6.

AI literacy

In short: The skills, knowledge and understanding that let people use AI in an informed way and assess its opportunities, risks and possible harms. For a company it takes more than knowledge: whoever adopts an AI result also needs the time and the authority to check it.

Where it comes from: A widely cited working definition of AI literacy was presented by Duri Long and Brian Magerko of the Georgia Institute of Technology at the CHI conference in 2020: competencies that let people critically evaluate AI, work with it and use it as a tool. The EU AI Act defines AI literacy in Art. 3(56), and Art. 4 AI Act has placed an obligation on providers and deployers since February 2, 2025.

What's new (as of November 2026): Under the amending Regulation (EU) 2026/1744, Art. 4 AI Act requires measures that support the AI literacy of staff. It doesn't prescribe a specific level per person.

Configuration power

In short: The influence of those who set up, update and reweight a system: thresholds, templates, default settings, the weights in a ranking. It appears on no org chart and is rarely assigned explicitly.

Where it comes from: I use the term in “Cheap to Make, Costly to Check”. The idea is older: in 1999 the legal scholar Lawrence Lessig wrote “Code is law”. What software allows and what it prevents regulates behavior as effectively as a statute.

What's new (as of October 2026): Part of this power lies with the provider. OpenAI's Model Spec, for example, explicitly ranks the provider's rules above the instructions of the companies that deploy the model. The other part usually belongs to two or three colleagues in IT or a business unit who know the settings.

In the course: weeks 1 and 6.

The four zones

In short: How strictly a task involving AI is handled. In Zone A the AI may work alone. In Zone B it proposes and a person checks. In Zone C the person decides and the AI supplies groundwork. In Zone D human judgment is mandatory and is trained specifically. The zone is decided explicitly as soon as one of five triggers applies, for example when a decision would be hard to reverse.

Where it comes from: The classification comes from “Cheap to Make, Costly to Check”. A related idea is the scale of automation levels that Thomas Sheridan and William Verplank developed in 1978. In the revised version of 2000 it runs from “the computer offers no assistance” to the computer deciding everything and ignoring the human.

What's new (as of October 2026): With agents, the question moves to the front. In Zone B someone checks the finished result; with an agent, what it may do at all has to be settled beforehand.

In the course: weeks 5 and 6.

Pace rule

In short: The commitment on what pace a type of decision may run at. Not making it is making it anyway, because then the software's pace applies. “Cheap to Make, Costly to Check” distinguishes four paces: machine pace with no people in between, review pace with fixed stopping points, human pace, in which the speed follows whoever decides, and maturing pace for processes that need their time and break when rushed.

Where it comes from: In manufacturing, takt time is the time within which a work step has to be finished. After the Second World War, Toyota made takt time a core part of its production system. The pace rule applies the idea to decisions.

What's new (as of October 2026): AI speeds up producing, while checking keeps its pace. Where nobody sets the pace, the difference ends up with the people at the review points.

In the course: weeks 4 and 6.

Organizational design

How an organization is built: its reporting lines, processes and rules. I sum all of this up as its structure.

Organizational design

In short: The deliberate design of how an organization works: who is responsible for what, how decisions are made, which rules apply and how the parts fit together. The org chart is only the most visible part of it.

Where it comes from: For a long time people looked for the one right structure. In 1967 Paul Lawrence and Jay Lorsch showed that the right structure depends on an organization's environment. In 1977 Jay Galbraith named the levers by which an organization can be designed and later developed them into his “Star Model”; in 1979 Henry Mintzberg sorted out the basic forms in which organizations are built.

How I use it: For me, organizational design means finding the spot where things are stuck and rebuilding there. That rarely takes a new org chart. With AI, it also includes the question of where results are checked and who answers for them.

Structure (Ordnung)

In short: My umbrella term for everything in an organization that governs how work gets done: reporting lines, processes and rules. It includes the rules that were formally decided, the ones built into software and the lived ones that are written down nowhere and apply anyway.

Where it comes from: In “Economy and Society”, published after his death in 1921/22, Max Weber described bureaucracy as authority based on rules that apply equally to everyone, and considered it the most efficient form of administration. Since then, organizational research has shown how far the decided and the lived structure can drift apart. In 1964 Niklas Luhmann spoke of the “useful illegality” without which no organization works.

How I use it: Many problems that get described as leadership or culture problems lie in the structure. That's why I first ask what the rules force people to do, and only then what the people do.

Operating system

In short: A picture for an organization's structure: what runs in the background and determines what the programs on top of it can do at all. People's behavior follows this operating system. An appeal to them therefore changes little.

Where it comes from: The image comes from computer science and moved into management literature. In 2007 Gary Hamel argued that most companies work with management practices from the early 20th century. In our book UNMANAGED, Lukas Michel, Herb Nold and I describe how many organizations run on an operating system built for a stable environment.

How I use it: From appeal to operating system. If you want to change behavior, first change what the structure rewards and what it enforces.

Organizational constitution

In short: The few commitments an organization goes by when in doubt, and the procedure by which they may be changed. It ranks above individual rules and approvals. At its core are the limits that apply even when it gets expensive; the Bonn Charter (in German) calls this part the binding core, “Cheap to Make, Costly to Check” calls it the core commitments.

Where it comes from: The idea comes from constitutional law. A constitution governs how rules are made and protects a core from the majority of the day; Germany's Basic Law (Grundgesetz) even puts it beyond any constitutional amendment in Article 79(3). In companies there are forerunners in articles of association, mission statements and the family constitutions of many family businesses.

How I use it: An organization can't make its core unchangeable; it can only make changing it harder. That's exactly where the value lies: everyone knows which hurdle stands in front of which change.

Decision path, approval rules, decision rights

In short: Three things that often get mixed up. The decision path is the actual process, visible in minutes and lead times. The approval rules are the document that sets who may decide up to what amount without asking, called a schedule of authority or signature policy, depending on the company. The decision right says who has the final say on a matter.

Where it comes from: Externally, the German Commercial Code (HGB), in force since 1900, governs who may bind a company through the Prokura, a statutory form of commercial power of attorney. The Prokura itself already existed in the General German Commercial Code of 1861. For the internal side, grids such as RACI or the RAPID model from the consultancy Bain (2006) came later; for every decision they record who recommends, who agrees and who decides.

How I use it: An unclear decision path is often answered with new approval rules, and afterwards nothing changes. So I look first at the path decisions actually take.

Review point

In short: A step in the process where a person checks or signs off a result before it counts. Blocks that a system sets on its own are built-in rules and don't count as review points. Review points can be counted and removed.

Where it comes from: The four-eyes principle has long been standard in accounting. In 1924 Walter Shewhart developed the control chart for manufacturing. It monitors a running process at fixed points before finished parts have to be sorted out, and with it he founded statistical quality control.

How I use it: Every rule is a promise to check, because someone has to check whether it's followed. With AI, more arrives at the review points while the time there stays the same. That's why I count the existing review points before I set up new ones. A reviewer who never rejects anything isn't reviewing.

Handoff

In short: The transition between a fast and a slow part of an organization, for example between a sales team that writes offers with AI and a legal department that reads every contract. This is where it's decided how much arrives and what knowledge comes along.

Where it comes from: In 1976 the organizational researcher Karl Weick described schools as “loosely coupled systems” whose parts respond to each other without tying each other down. In 1984 Charles Perrow showed, using industrial accidents, that systems that are both tightly coupled and complex pass errors on unchecked.

How I use it: Where AI makes one part faster, the load moves to the next handoff. If you don't adjust there, you'll experience the gain as a traffic jam.

Reversibility

In short: The question of whether a decision can be taken back later. It can be taken back if that's possible with reasonable effort. It's hard to take back if it ties up money, time or dependencies for the long term. In between lies what can only be partly taken back.

Where it comes from: In his shareholder letter for 2015, Jeff Bezos compared decisions to doors: you can walk back through some, others close behind you. Decision theory speaks of the option value you give up as soon as you commit.

How I use it: What can be taken back may be decided quickly. What's hard to take back needs time and a higher hurdle. Many organizations treat both with the same slowness.

Binding layer

In short: The level above small, self-reliant units that keeps them together, with a few clear rules, responsibilities of its own and funding that fits its tasks. Without it, independent teams fall apart into islands. If it gets too powerful, the teams lose their independence.

Where it comes from: In 1931 Catholic social teaching formulated the principle of subsidiarity: what the smaller unit can do, the larger one shouldn't take away from it. In 1990 Elinor Ostrom showed, using pastures, forests and irrigation systems, that communities can maintain common goods over centuries, and she named design principles for this; one of them is several levels that interlock. In 2009 she received the Nobel Memorial Prize in Economic Sciences for this work.

How I use it: Small units that can work as communities, with a binding layer above them that has its own rules: this design principle applies in a company just as in federalism or in dual vocational training (duale Ausbildung), where companies, chambers and vocational schools shape the system together.

Self-organization

In short: Teams decide on their own work without every decision going through a manager. That works where the rules for it are clear. Without such rules, an invisible hierarchy forms.

Where it comes from: In 1951 researchers at London's Tavistock Institute showed in British coal mines how a new mining method broke up established, self-regulating work groups and cost health and performance as a result. Follow-up studies up to 1963 showed that semi-autonomous groups work better. In 1972 the activist Jo Freeman described in “The Tyranny of Structurelessness” how groups without rules form informal centers of power. Later came Holacracy (2007) and Frederic Laloux's book “Reinventing Organizations” (2014).

How I use it: Self-organization needs more structure, if anything. Above all it needs accountability among peers. In most organizations, accountability only runs upward.

Management

For me, management is the task that concerns the structure: designing, maintaining and changing reporting lines, processes and rules.

Management

In short: The task of designing, maintaining and changing an organization's structure, with its goals, reporting lines, processes, rules and resources. Leadership is aimed at people, management at the structure they work in. Often the same people do both.

Where it comes from: In 1916 Henri Fayol described the tasks of management as planning, organizing, commanding, coordinating and controlling. In 1954 Peter Drucker made management a discipline of its own with “The Practice of Management”. In 1990 John Kotter drew a line between management as coping with complexity and leadership as coping with change.

How I use it: If you say “leadership” where management is meant, you look for the solution in individual people's behavior, even though the structure forces that behavior.

Management system and “unmanaged”

In short: The management system covers the rules, routines and instruments an organization is steered by: goals, budgets, reports, approvals, meetings. “Unmanaged” means this system was never deliberately designed. It grew over time and often dates from a period with a more stable environment.

Where it comes from: The term management system spread with the quality standards that began with ISO 9001 in 1987. Since the 2000 edition, the standard has spoken explicitly of a quality management system. In UNMANAGED (LID Publishing), Lukas Michel, Herb Nold and I describe how to recognize a management system that nobody designed and how to rebuild it.

How I use it: The managers in an undesigned system are rarely the problem. They work with instruments that were built for different conditions.

Insight · Clarity · Impact

In short: The three steps I work by. Insight means seeing how the organization really works and where it's stuck. Clarity means deciding what should change and who answers for it. Impact means putting the change into practice so that it lasts in daily work, and checking that it does.

Where it comes from: The three steps follow the old pattern of diagnosis, decision and implementation that applies in medicine as much as in quality management. In the 1950s William Edwards Deming popularized the Plan-Do-Check-Act cycle. It goes back to earlier work by Walter Shewhart.

How I use it: The sequence makes the difference. If you start with impact, you implement measures before it's clear what they're for.

Control reflex

In short: The habit of responding to uncertainty with more control: extra approvals, reports, attendance rules, review loops. Each single measure makes sense. Together they cost time and trust, and in the end often the very safety they were supposed to provide.

Where it comes from: In 1911 Frederick Taylor built scientific management on separating thinking from doing, and control was part of it from the start. In 1960 Douglas McGregor described “Theory X”, the assumption that people won't work without supervision, and showed how it confirms itself. In 1997 Michael Power called the growing culture of audits the “audit society”.

How I use it: The control reflex is a practice, and it says little about the character of individual people. You recognize it when a mistake is followed by a new review step and no old one goes away. Its counterpart is trust architecture.

Trust architecture

In short: A structure that makes trust possible without blindly assuming it, with clear responsibilities, decision rights that match accountability, and checks where they actually catch something.

Where it comes from: In 1968 Niklas Luhmann described trust as a mechanism that reduces complexity: if you trust, you don't have to check everything yourself. In 1974 the economist Kenneth Arrow called trust an important lubricant of a social system.

How I use it: Trust can be built. If you set who answers for what and where checks happen, you get by with less control everywhere else.

Checking load and the cost triad

In short: The checking load is the work that arises when someone has to check a result that others or a machine produced. The cost triad asks about three positions: who saves, who checks, who pays? AI has made producing cheap. The time saved appears in the metrics, the checking time further down the chain usually doesn't.

Where it comes from: Both terms come from “Cheap to Make, Costly to Check”. The observation behind them is older. In 1987 the economist Robert Solow wrote that you can see the computer age everywhere except in the productivity statistics.

How I use it: The person who produces saves time that the person who checks then pays for. Because checking costs appear in no metric, a system often looks better than it is.

Leftover load

In short: What's left when producing gets cheap and checking and answering for it don't. The leftover load doesn't disappear. It lands somewhere, and usually nobody decided where.

Where it comes from: The leftover load is one of the three core ideas of “Cheap to Make, Costly to Check”, where it goes by the name checking load. Economics has described the pattern since Arthur Pigou (1920); today it's called external costs: costs that don't fall on whoever causes them.

How I use it: For every AI project I ask who ends up with the leftover load, and whether that person can bear it. Usually it falls to the person whose signature is due.

Metric

In short: A number by which an organization wants to tell whether something is working. What gets measured gets attention, and what stays unmeasured disappears from the meetings.

Where it comes from: In 1975 the economist Charles Goodhart observed that a number loses its meaning as soon as you steer by it. Shortly afterwards, the social psychologist Donald Campbell described the same for social policy. Today the idea is known as Goodhart's law.

How I use it: With AI, checking time is missing from almost every set of metrics. The savings are in the report; the work they cost elsewhere is written down nowhere.

Bottleneck

In short: The point that sets the pace of the whole. An improvement anywhere else speeds up nothing as long as the bottleneck remains.

Where it comes from: Eliyahu Goldratt popularized the idea in 1984 with his novel “The Goal” and built the Theory of Constraints from it.

How I use it: The gain from AI gets stuck at the organization's bottleneck: at decisions that go round in circles, at knowledge locked in one head, at reviewers without time. If you only improve the tool, you speed up steps that weren't slowing anything down.

Governance

In short: The rules for how an organization is steered and overseen: who decides, who monitors, who reports to whom. AI governance is the part of it that sets who decides on the use of AI and who answers for its results.

Where it comes from: In the UK, the Cadbury Report set out principles of good corporate governance in 1992. In Germany, the German Corporate Governance Code has existed since 2002.

How I use it: An AI policy is only where governance starts. Whether it works depends on whether someone has the time and the right to enforce the rule, and whether the advisory board finds out if the checks are effective.

Accountability

In short: The duty to explain to others what you did and why, and to answer for it. In most organizations it runs only upward, to the next manager.

Where it comes from: The German word, Rechenschaft, comes from arithmetic. Giving Rechenschaft originally meant presenting a statement of accounts, and the English “accountability” also goes back to reckoning, via the Old French “aconter” and the Latin “computare”. In public administration it became a core concept of democratic control.

How I use it: If you want to work in a self-organized way, you have to set up accountability sideways. Many teams simply lack practice at it, because for years there was one person who did it for everyone.

In charge and responsible

In short: You're in charge when a task has been assigned to you. You're responsible when you bear the consequences if it goes wrong. The two often drift apart, especially where responsibility is assigned without time, knowledge or decision rights coming with it.

Where it comes from: German organizational theory has long demanded that task, authority and responsibility match, and calls this the congruence principle; back in 1916 Fayol already required authority and responsibility to correspond. German company law (the GmbH Act and the Stock Corporation Act) keeps management responsible for due care even when it delegates tasks.

How I use it: Many responsibilities in companies are assigned and never equipped. Then there's a name on the org chart while the load ends up somewhere else.

Leadership

For me, leadership is the task that concerns people: making sure they can judge, object and take responsibility within the structure.

Leadership

In short: The task that concerns people: giving direction, enabling judgment, settling conflicts, supporting development. Leadership works within a structure, and even good leadership can hardly make up for a poor structure in the long run.

Where it comes from: In 1939 Kurt Lewin and colleagues distinguished authoritarian, democratic and laissez-faire leadership in experiments with youth groups. In 1978 James MacGregor Burns contrasted transactional leadership, based on exchange, with transforming leadership; Bernard Bass coined the now common term “transformational” in 1985. Since the 1960s, situational approaches, for example by Fred Fiedler, have asked under which conditions which kind of leadership works.

How I use it: When things aren't working in a team, it can be a problem of poor leadership. Far more often it's down to the structure the team works in.

Leadership system

In short: The interplay of people, roles, bodies and rules through which an organization is led. It reveals itself less in normal operations than in an emergency, when decisions have to be made quickly and the usual routes are too slow.

Where it comes from: Since the 19th century, armies have thought of leadership as a system of mission, situation and reporting lines. Crisis teams in public administration and companies have adopted this way of thinking.

How I use it: Whether a leadership system stands up shows in the first twenty minutes of a real crisis. That's when you see whether it works or only exists on the org chart.

Mission

In short: What someone is meant to achieve, for what purpose and within what limits, without every step being prescribed. A clear mission sets the purpose and leaves room for the how.

Where it comes from: In the 19th century the Prussian army developed mission-type tactics (Auftragstaktik). Junior commanders were given the goal and the intent and decided on the route themselves, because orders arrived too late in battle; Helmuth von Moltke the Elder is its best-known proponent. Today the Bundeswehr calls the principle “leading by mission” (Führen mit Auftrag).

How I use it: A tool makes things faster, and the mission stays as unclear as before. A process without a clear mission gets faster with AI and sounds more convincing along the way.

Judgment

In short: The ability to reach a well-founded decision in a situation for which there's no rule. It develops through practice on real cases.

Where it comes from: Aristotle called this practical wisdom “phronesis” and in the Nicomachean Ethics set it apart from knowledge of general laws. In 1790 Immanuel Kant described judgment as the faculty of thinking the particular as contained under the universal.

How I use it: Judgment fades when a system permanently takes over the groundwork. In four Polish endoscopy centers, doctors who had worked with AI for three months found fewer precancerous lesions without it (Budzyń et al., Lancet Gastroenterology & Hepatology, 2025). So every use of AI comes with the question of what is regularly done without the system.

Challenge

In short: The right and the duty to raise objections to a decision, a proposal or a machine-produced result, together with a path along which the objection actually arrives. “Cheap to Make, Costly to Check” calls this the duty to challenge.

Where it comes from: In 1970 Albert O. Hirschman described two responses to decline: exit and voice. According to Hirschman, when quality falls it's often the most attentive members who leave first if their voice achieves nothing. In 1999 Amy Edmondson showed that teams learn more when raising objections has no downside. Taking up an older term from Edgar Schein and Warren Bennis, she called this state psychological safety in the team.

How I use it: Whether people report uncomfortable findings depends on what happened to the last uncomfortable findings. An objection that never arrives anywhere won't even be raised next time.

Breadth of view

In short: The range of what actually goes into a decision. Who's at the table answers only half of that. What matters is whether more than one way of knowing a thing gets through to the decision.

Where it comes from: In 1972 the psychologist Irving Janis described groupthink, using the Bay of Pigs invasion among other examples: committees full of smart people decide badly when agreement matters more to them than examining the situation.

How I use it: AI summaries and pre-selections can quietly narrow the breadth of view. They weight things, and the decision-maker doesn't see what falls away.

Participation

In short: How far people take part in decisions that affect them, from being informed to deciding together. Taking people along is something other than deciding with them.

Where it comes from: In 1969 the American policy adviser Sherry Arnstein arranged citizen participation on a ladder with eight rungs: nonparticipation at the bottom, tokenism with information and consultation in the middle, citizen control at the top. In German companies, elected works councils (Betriebsräte) have existed since the Works Councils Act of 1920. Today the Works Constitution Act (Betriebsverfassungsgesetz) governs co-determination at company level (Mitbestimmung), first in 1952 and in its current basic form since 1972.

How I use it: Many participation formats promise joint decisions and deliver consultation. That costs more trust than saying openly which rung you're on.

Commitment

In short: A promise someone makes to others and agrees to be measured against, whether to customers, employees or the public. Whether a commitment is real shows when it costs something.

Where it comes from: In lectures at Harvard in 1955 (published as a book in 1962), the philosopher of language John L. Austin showed that some sentences are themselves an action: whoever promises binds themselves. In 1986 Terry Winograd and Fernando Flores described organizations as a network of requests and commitments.

How I use it: With published goals, I ask whether the commitment ever appeared in a decision paper and changed the outcome there. For many commitments that never happened.

Strain

In short: The state in which a work system runs continuously faster than the people in it can come to rest. It becomes visible in interruptions and in how hard it is to switch off, long before results suffer.

Where it comes from: In 1979 the work scientist Robert Karasek showed that high demands are psychologically straining above all when people have little say over their work. Later studies found the same link for heart disease.

How I use it: AI raises the pace where things are produced; the strain arises where they are checked. Leadership here means setting the pace yourself before the software does.

To the course: The Silent Power · The Course (in German) · To the newsletter: The Silent Power (in German)