Organisational Architecture

How AI exposes the organisation you actually run, and what leadership must redesign

This is the working intellectual architecture beneath Challenger House’s practice.

Every function in your company is being redesigned for AI. Management is the exception. That matters because AI does not enter a clean organisation. It enters the one you already have: the formal processes, the undocumented workarounds, the trusted relationships, the unclear decisions and the knowledge people carry in their heads. AI makes this gap easier to see and harder to ignore. It can speed up a task in days. It cannot decide why the task exists, who should be answerable for it, what evidence deserves trust or which decisions must stay human. Those are questions of organisational architecture.

The landing page says why. What we do says what. This page is how we think, all of it, for anyone who wants to chew.

It looks different from the rest of the site on purpose: it is a reference, built to be read in parts, linked to by chapter and argued with.

It was first assembled in August 2026 and will change as client work produces better evidence. Some ideas are established parts of our practice. Others, especially in Part Seven, remain hypotheses to test. We build in the open because disagreement improves the architecture. If a chapter does not match what you see in your organisation, tell us which one and why.

Built to be read in parts, one slide at a time.

How to read this document

Press Next to choose a route.

Choose a route.

Start anywhere. Seven parts, thirty-six chapters, each with its own address. Three ways in, or simply press Next.

How to read this document

Press Next now for the argument in seven parts, or pick a route above.

The argument in seven parts

• The organisation you actually run. Why the formal organisation and the working organisation are never quite the same.
• Why AI exposes the gap. Why a manageable human workaround becomes an architectural problem when machines join the work.
• The five layers. What leadership needs to design, from purpose and decision rules to sensing and adaptation.
• How the organisation knows. How attention, information, interpretation and AI use shape decisions.
• How to see the real organisation. How the Org Mirror uses several lenses without pretending that one model explains everything.
• From theory to practice. What leaders must decide and which instruments make those decisions concrete.
• Where this may lead. A testable view of organisations that can sense and revise their own architecture.

This is not a finished doctrine. It is a working architecture built from organisational development, cybernetics, network theory, philosophy and twenty-five years inside large organisations. We expect client work to change it.

The organisation you actually run

Why the formal organisation and the working organisation are never quite the same.

• 1. You run two organisations
• 2. A system is what it does
• 3. The house
• 4. The Tower and the Square

You run two organisations

Every organisation has a formal version and a working version.

The formal version appears in the org chart, the process map, the governance model and the policy library. It tells people how work is meant to move.

The working version appears in everyday behaviour. It shows who people call when something is stuck, which meeting actually decides, whose objection can stop a project and which rule everyone quietly works around.

The difference is not evidence that people have failed to follow the design. No formal model can contain every exception, relationship or judgment call. Human beings have always filled the gap.

Imagine a procurement process with three written approval steps. In practice, an experienced project lead calls one trusted finance partner before submitting anything. That conversation prevents weeks of delay, but it does not appear in the process map. Remove the person or automate only the written process, and the system loses a piece of intelligence it never knew it had.

We call the space between the formal organisation and the working organisation the Missing Middle. It contains the decisions, translations and handovers that make the formal system work in practice.

A system is what it does

Stafford Beer gave systems thinkers a useful principle: judge a system by what it repeatedly produces, not by what it says it intends.

An organisation may say that it values challenge, for example. If people who challenge senior decisions stop getting invited into important rooms, the operating purpose of the system is conformity. The statement on the wall still matters, but behaviour tells us which rule is stronger.

This gives us two useful views:

• The stated self: strategy, values, policies, reporting lines and documented processes.
• The doing self: decisions, incentives, promotions, workarounds, informal power and the actual movement of work.

AI usually receives the stated self first because that is the material available to a system. It can read the policy and the procedure. It cannot automatically recover the ten years of experience that taught a manager when the procedure should bend.

Human thinking used to absorb this difference. People translated incomplete instructions, noticed exceptions and carried context from one conversation to another. The work was often invisible and rarely costed, but it kept the organisation viable.

AI reduces that buffer. It follows what has been made explicit, at speed and at scale. The more action we delegate, the more important it becomes to understand what the organisation actually does.

The house

The house is a simple way to see five parts of the organisation that leaders often discuss separately.

The attic: knowledge that was never written down

The attic contains tacit knowledge. This is the know-how people develop through experience: which warning sign matters, how to calm a difficult client, when a number looks technically correct but practically impossible.

Tacit knowledge is not merely undocumented information waiting to be uploaded. Some of it can be described. Some of it only appears while an expert is doing the work and explaining their judgment.

The stairwells: coordination nobody designed

The stairwells are the Missing Middle. They connect the formal floors of the organisation. This is where people decide who checks an output, what happens when the machine and the expert disagree, who carries the result into the next team and where the organisation stores what it learned.

When these links are weak, meetings multiply. Managers chase updates, translate between functions and repair handovers by hand.

The basement: the organisational shadow

The basement contains what the formal story leaves out: unclear ownership, competing incentives, informal power, protected territory and reasonable fears about status or redundancy.

We call this the organisational shadow. It is not a claim that people are obstructive. It recognises that ambiguity often serves a purpose. It can preserve relationships, protect options or keep a conflict below the surface. Making the work explicit therefore has political and emotional consequences, not just technical ones.

The load-bearing walls: what must remain human

Some decisions can move towards machines. Others should not. The load-bearing walls are the judgments and accountabilities an organisation chooses to keep human.

Examples may include dismissing an employee, accepting a safety risk or deciding which community bears the cost of a major investment. The exact list differs by organisation. What matters is that leaders choose it rather than inherit it from a vendor’s default settings.

The roof: new possibilities

Once the middle works, AI can do more than speed up existing tasks. Teams can explore more options, test assumptions earlier and create services the old workflow could not support.

The roof comes last. New value rests on the less glamorous work below: clear handovers, visible judgment, workable boundaries and a learning loop.

The Tower and the Square

Historian Niall Ferguson uses two shapes to describe how organisations coordinate.

• The TowerThe formal hierarchy. It includes reporting lines, budgets, committees, systems of record and the authority to allocate resources. It is visible to an auditor.
• The SquareThe network of relationships. It includes trust, informal influence, rapid conversations and the colleague who knows how to get something unstuck. It becomes visible when someone central leaves and the work suddenly slows down.

Neither shape is good or bad. Organisations need the Tower to hold accountability and the Square to move information through real relationships.

The Missing Middle is where the Tower depends on the Square without acknowledging the dependency. A manager may spend half a day translating between two departments because the formal process cannot carry the context. The work gets done, but the organisation does not learn how it got done.

This is why more automation can initially create more management work. The machine speeds up the formal process while people continue to repair the informal connections around it.

Next: why AI turns this gap from a normal feature of organisational life into a leadership problem that needs explicit design.

Why AI exposes the gap

Why a manageable human workaround becomes an architectural problem when machines join the work.

• 5. A print-age organisation meets network-age technology
• 6. Management has always written the constitution
• 7. Execution can move faster than accountability
• 8. You have to experience AI to lead with it
• 9. Productivity is not a sufficient reason
• 10. AI exposes weak foundations
• 11. The most senior accountable leader must join the work

A print-age organisation meets network-age technology

Alexander Bard describes four broad information eras: speech, writing, print and digital networks. His model is a philosophical lens rather than a settled history, but it makes one mismatch easy to see.

The modern corporation grew in the age of print. Information was expensive to copy, slow to move and difficult to coordinate across distance. Organisations responded with departments, reporting lines, standard forms, managers and layers of approval.

Much of management developed to route information. Managers collected updates, translated them into reports, passed decisions down and handled the exceptions that did not fit the form.

Digital networks changed the speed of communication. AI changes what can happen inside that communication. A system can now read, compare, draft, decide within limits and initiate further work. Information no longer has to pause at every human junction.

The old structure does not disappear overnight. Authority, incentives and legal accountability still live in it. This produces the central mismatch: network-age capability enters a print-age organisation.

Training helps people use the technology. It does not resolve who may delegate what, how evidence should be judged or how learning should move across the organisation. Those are architectural questions.

Management has always written the constitution

Here, constitution does not mean a national legal document. It means the rules that shape what the organisation is allowed to do and how it decides.

Every organisation has such rules, even when nobody has written them in one place. They cover purpose, participation, evidence, permissions, escalation, memory and accountability.

Indy Johar distinguishes between two kinds of action:

• Operational action happens inside an existing system. A team processes a claim, writes a proposal or schedules maintenance.
• Constitutive action shapes the system itself. Someone decides what counts as a valid claim, which evidence the proposal must include or when maintenance can be postponed.

Managers have always performed both kinds of work. They delivered results and also interpreted the rules around those results. Much of the second role remained tacit because a human manager could apply context in the moment.

AI forces more of the constitution into the open. An agent needs permissions, data, thresholds and instructions. Somebody decides what it may remember, when it must ask for help and who answers when it fails.

If the leadership team does not make these choices, they do not remain undecided. They move into software settings, procurement choices, IT controls and vendor assumptions.

Execution can move faster than accountability

An organisation can give an AI system a task in an afternoon. It can take months to agree who is answerable for the result.

Consider a customer-service agent that drafts and sends responses. The technical question is relatively simple: can the system access the case history and produce an answer? The organisational questions are harder:

• Which cases may it handle without review?
• What happens when the case includes vulnerable customers or legal risk?
• Who monitors patterns across thousands of replies?
• Who is accountable when a plausible answer causes harm?

The person operating the system, the person who designed the workflow and the executive who authorised it hold different responsibilities. Treating all three as “the human in the loop” hides the distinction.

The law reinforces the need for clarity. In Germany, section 43 of the GmbHG requires managing directors to exercise the care of a prudent businessperson and makes them liable to the company for breaches of duty. The EU AI Act now applies in stages, with certain transparency requirements in force from August 2026 and rules for high-risk systems following later. The exact duty depends on the system and context, but leadership cannot outsource its responsibility by buying a tool.

This is not an argument for keeping every decision human. It is an argument for moving execution only as fast as the organisation can state and support accountability.

You have to experience AI to lead with it

Reading about AI gives a leadership team information. It does not give them a reliable sense of what the technology can do, where it fails or how it changes their own thinking.

John Vervaeke distinguishes four kinds of knowing. The terms are academic, but the differences are practical:

• Knowing that: facts and propositions. You can explain what a language model is.
• Knowing how: skill built through practice. You can use one well on a real task.
• Knowing what matters: judgment. You can tell which output deserves attention and which warning sign matters.
• Knowing with others: understanding that forms through participation. A team discovers something while working together that none of its members could have specified in advance.

Most corporate AI programmes concentrate on the first kind. They distribute information, policies and basic training. Leadership requires all four.

For example, a team may understand that AI can compare contracts. Only practical use reveals which clauses demand expert judgment, how quickly the first draft anchors the discussion and where a confident summary hides uncertainty.

This is why Challenger House starts with real work. Experience expands the leadership team’s imagination and exposes the boundaries it needs to design.

Productivity is not a sufficient reason

Productivity is a sensible place to begin. It is not a distinctive destination.

If every competitor buys similar tools and asks for similar efficiency gains, the result is parity. Work gets cheaper or faster, but the organisation has not decided what it wants to become.

A useful reason for AI must connect to the organisation’s strategy. It might improve the reliability of a regulated service, shorten the time between a customer signal and a decision or help scarce experts serve more people without lowering the standard of judgment.

This choice belongs with the most senior accountable leaders because it changes priorities and trade-offs. A policy written two levels down can describe permitted use. It cannot decide which kind of value the organisation should pursue.

The practical question is simple: What can this organisation now do for customers, employees or society that it could not do before?

If the answer remains “the same work, more cheaply”, the organisation has an efficiency programme rather than an AI strategy.

AI exposes weak foundations

AI does not attack an organisation’s foundations. It depends on them.

A system needs usable knowledge, clear ownership, access rules, feedback and a definition of a good result. Where these foundations are weak, AI makes the weakness visible.

The most difficult issue is often tacit knowledge. People may be asked to document the judgment that makes them valuable so that a system can reproduce part of their work. It is reasonable for them to ask what happens next.

An organisation that treats this as resistance will drive the knowledge underground. A better response is to make the exchange explicit:

• What knowledge is the organisation asking people to contribute?
• How will contribution be recognised and rewarded?
• Which parts of the role will grow rather than disappear?
• What commitments will the organisation make about redeployment, development and attribution?

This is where the employment relationship enters the architecture. Knowledge capture is not a neutral technical exercise. It changes bargaining power, identity and trust.

The most senior accountable leader must join the work

Senior sponsorship is not the same as senior participation.

A sponsor can approve a budget, attend the launch and ask for progress reports. Participation means using the technology on real work, making visible choices and allowing the team to see where the leader’s own understanding changes.

This matters for three reasons.

1. Only senior leaders can resolve trade-offs across functions. IT may optimise security, a business unit may optimise speed and Legal may minimise exposure. Someone must decide for the whole organisation.
2. People judge priorities through behaviour. If the accountable leader delegates the learning, the organisation reads AI as another programme rather than a change in how leadership works.
3. The architecture reflects the judgment of its architects. Their fears, incentives and blind spots enter the rules whether they acknowledge them or not.

Challenger House therefore asks the most senior accountable leader to be in the room. This is a qualification for the work.

The five layers of organisational architecture

What leadership needs to design, from purpose and decision rules to sensing and adaptation.

• 12. The model at a glance
• 13. Source: where the initiative begins
• 14. Constitution: the rules that shape the organisation
• 15. Boundaries: decide what moves and what stays
• 16. Sensorium: how the organisation knows
• 17. Adaptation: how the organisation changes

The model at a glance

The model has five layers. They are not departments and they do not form a new reporting hierarchy. They are five sets of questions that every organisation must answer.

The unfamiliar word is sensorium. In a person, the sensorium is the whole system through which they perceive the world. In an organisation, it is the combination of data, conversations, observation and interpretation through which the organisation understands itself and its environment.

Read the model in two directions

The model works from the top down and from the bottom up.

• From the top down, commitments shape action. Source informs the Constitution. The Constitution sets the Boundaries. The Boundaries determine what the organisation may sense and how it may act.
• From the bottom up, reality tests commitments. Adaptation produces results. The Sensorium notices what happened. Leaders may then revise a boundary, a constitutional rule or, in rare cases, the initiative’s underlying purpose.

The model is a loop. A viable architecture holds its identity while allowing evidence to change how it operates.

Source: where the initiative begins

The word Source comes from Peter Koenig’s Source Principle, later explained by Tom Nixon in Work with Source.

The Source is the person who first steps into uncertainty and takes responsibility for bringing an initiative into being. This is not always the founder, the formal leader or the person with the largest budget. It is the person whose perception gives the initiative its original direction.

The distinction matters because organisations often replace a person with a category. They say “the business” owns a project or “leadership” provides direction. When the initiative reaches a difficult choice, nobody knows whose judgment can resolve it.

Large organisations contain many Sources. A company may have one Source, while a product, transformation or client engagement has another. The pattern is nested rather than centralised.

Naming Source does not give one person unlimited authority. It clarifies creative authorship so that authority, contribution and succession can be discussed openly. A healthy Source invites challenge and enables other Sources to lead sub-initiatives.

The practical questions

• Who first saw this need and accepted the risk of acting on it?
• Which decisions still depend on that person’s judgment?
• Where should that judgment be made explicit or handed over?
• What happens if the Source leaves?

Constitution: the rules that shape the organisation

The Constitution translates purpose into rules people and systems can follow.

It includes more than governance. It covers the outcomes the organisation values, the evidence it trusts, the people allowed to participate, the incentives attached to behaviour, the decisions that require escalation and the commitments that cannot be traded away.

Most organisations spread these choices across strategy decks, policies, employment contracts, software settings and unwritten managerial habits. AI reveals the contradictions.

For example, a company may say that customer trust matters more than speed. Its service agent may still be measured almost entirely on handling time. If an AI system optimises the measurable rule, it will make the contradiction more efficient.

A useful Constitution therefore names trade-offs. It might state that the organisation will accept a slower response when a vulnerable customer needs human judgment. It might prevent an agent from inferring employee performance from private messages, even if the inference is technically possible.

The Constitution should be stable enough to guide action and revisable enough to learn.

Part Six explains the AI Constitution →, our practical instrument for authoring this layer.

Boundaries: decide what moves and what stays

The question is not whether a task is human or automated. Most work will combine people and machines in different ways.

We use a Delegation Dial with five positions:

1. Human: a person performs and decides the work.
2. Augmented human: AI supports the person, who remains fully in control.
3. Human-led agents: a person sets the goal and supervises several machine actions.
4. Supervised autonomy: the system acts within limits while a person monitors exceptions.
5. Autonomous: the system acts without routine human review inside a tightly defined domain.

No position is inherently mature. A fully human decision can be appropriate for a high-consequence, irreversible judgment. Autonomous action can be appropriate for a low-consequence task that is easy to reverse and easy to monitor.

Set the dial using five factors: consequence, reversibility, confidence, regulation and trust earned through evidence.

The dial also needs four separate boundaries:

• Execution: who or what performs the work.
• Information: who or what may access the context required to act.
• Accountability: who must answer for the result.
• Identity: who belongs to the organisation and which commitments define “us”.

These boundaries do not have to move together. A machine may execute a task without receiving all available information. It may recommend a decision while accountability remains with a named human.

Sensorium: how the organisation knows

An organisation does not know something merely because the data exists.

It needs a way to notice signals, preserve context, compare interpretations and move important information to someone who can act. That whole arrangement is the Sensorium.

Dashboards form one part. So do customer conversations, frontline judgment, staff surveys, maintenance reports, external research and the colleague who says, “The numbers look fine, but something has changed.”

A strong Sensorium has four properties:

• It sees the work, not only the report. Leaders can reach direct evidence rather than relying on summaries alone.
• It preserves provenance. People can see where a claim came from and how it was interpreted.
• It allows disagreement. Several interpretations can remain visible until evidence resolves them.
• It protects people from surveillance. The organisation senses workflows and dependencies for a declared purpose rather than collecting behaviour because it can.

AI can make a Sensorium faster and broader. It can also centralise interpretation in a single model or leadership group. Part Four explains how to gain the first benefit without accepting the second risk.

Adaptation: how the organisation changes

Adaptation is where the architecture meets everyday work.

A team notices a problem, changes the workflow, tests the change and learns from the result. Agile methods, OKRs, transformation backlogs and improvement programmes all operate here. They are useful, but they sit downstream of the other layers.

If the Constitution rewards volume, an agile team will find faster ways to produce volume. If the Boundaries are unclear, a backlog will fill with local automation while accountability remains unresolved. If the Sensorium filters out frontline concerns, the organisation will adapt to an incomplete picture.

This is why method alone does not transform an organisation. The method acts inside an architecture that may reinforce the very behaviour leaders want to change.

Good adaptation creates evidence for the layers above it. A workflow test may show that a risk threshold is too cautious, that a model needs different context or that the original objective harms another part of the system. The architecture changes because the work teaches it something.

Next: the Sensorium in detail. What the organisation notices, how meaning gets made and how plausible machine output can distort both.

How the organisation knows

How attention, information, interpretation and AI use shape decisions.

• 18. The Attention Constitution
• 19. Viability: five things every system must do
• 20. Information latency
• 21. No one should control both the data and its meaning
• 22. AI sovereignty
• 23. The Plausibility Illusion

The Attention Constitution

Edgar Schein argued that leaders shape culture through what they repeatedly pay attention to, measure and control. AI makes this process executable.

A manager may ask about safety in a weekly meeting. An agent can scan for safety signals every hour, remember patterns and trigger action. That makes the original choice of attention far more consequential.

The Attention Constitution turns this into six explicit decisions:

1. What should the organisation notice? Define the signals that deserve attention.
2. What counts as evidence? Separate a useful signal from a claim, opinion or anomaly.
3. What may the system remember? Set limits on retention, sensitive context and institutional memory.
4. Who may challenge the first interpretation? Give affected people and alternative experts a route to contest it.
5. What threshold triggers action? State when a signal becomes an alert, escalation or decision.
6. What must never happen automatically? Name the judgments that require human responsibility.

Take employee absence as an example. An organisation might notice team-level patterns to plan capacity. That does not justify inferring an individual’s motivation from messages or turning a weak correlation into a performance judgment.

The Attention Constitution defines the purpose, evidence, boundaries and escalation before the system begins to watch.

Viability: five things every system must do

Stafford Beer’s Viable System Model describes five functions a system needs in order to remain viable. They are not five departments. People, teams or machines may perform each function in different combinations.

In plain English, the organisation must:

1. Do the work. Deliver the service, make the product or fulfil the mission.
2. Coordinate. Stop teams from colliding when they share customers, resources or constraints.
3. Hold the whole together. Balance local goals so that one function does not win at the organisation’s expense.
4. Watch what is changing. Sense the external environment and explore what the organisation may need to become.
5. Know who it is. Hold identity and policy, then balance today’s operation with tomorrow’s possibility.

The model gives us a practical diagnostic for AI work:

• Where do meetings compensate for missing coordination?
• Where do leaders rely on decks because they cannot see the operation?
• Where does the dashboard replace direct observation?
• Where does innovation discuss AI while operations continue unchanged?
• Where has nobody said what the organisation will never delegate?

Beer also described viable systems as nested. A team needs all five functions. So does a business unit and the wider organisation. This argues against building one central AI brain. Local autonomy works when each level has enough information and clear boundaries.

Information latency

Information latency is the time between something changing in the real world and that change reaching someone who can respond.

In many organisations, a customer signal passes through an employee, a manager, a meeting, a report, a presentation and an executive committee before it becomes a decision. Each step adds delay and removes context.

AI can shorten the path:

1. A signal appears.
2. The relevant context is attached.
3. A named person or authorised system receives it.
4. Action follows within clear limits.

Shorter is not always better. A weak signal may need time and comparison before it deserves action. The aim is a deliberate path that matches the speed and consequence of the issue.

Exceptional signals also need a route around routine reporting. A serious safety concern should behave like a fire alarm, not wait for the quarterly dashboard. Beer called this an algedonic channel: a direct path for urgent pain or opportunity.

Measuring information latency shows where the organisation pays a Talking Tax. People spend time chasing, translating and packaging information instead of acting on it.

No one should control both the data and its meaning

An organisation creates a dangerous concentration of power when one group controls the sensors, the data and the authorised interpretation.

In practice, this often means a central leadership team receives a simplified dashboard and treats its reading as the organisation’s reality. AI can intensify the problem by collecting more signals and presenting one polished explanation.

The alternative is not endless debate. It is a system in which important interpretations remain inspectable and contestable.

• Show where each inference came from. This is provenance: the visible origin of a claim and the steps used to reach it.
• Allow different models or analysts to disagree.
• Give the people affected by a reading a way to challenge it.
• Keep consequential decisions with named, accountable people.
• Separate the team that composes a view from the team that audits it when the stakes require independence.

These rules also protect the Org Mirror. It should sense the work before it senses the worker. It should map delays, dependencies and conflicting accounts for a declared purpose. It should never create a universal employee score or mine private conversation because the data happens to exist.

Plural interpretation is a form of fault tolerance. A leadership team with the same education, incentives, data and assumptions can be highly intelligent and wrong in the same direction.

AI sovereignty

We use AI sovereignty to mean the capacity to govern AI because you understand enough about how you use it, how far it reaches, what it costs and what it does to judgment. It has four dimensions.

What you use it for

Most organisations begin with production: drafting, summarising, analysing and searching. AI can also support reflection, challenge an assumption, simulate another perspective and help a team prepare for a difficult decision. Sovereign use means choosing the mode deliberately rather than accepting the most obvious feature of the tool.

How far it reaches

A personal assistant is the smallest possible use. AI may also change a workflow, coordinate several agents or enable a service the old organisation could not provide. Leaders need to understand the reach of each use case. A private drafting tool and an agent that acts across customer systems require different architecture.

What it is and what it costs

There is no single best model for every task. Systems differ in capability, speed, privacy, reliability and price. Sovereign organisations route work deliberately. They do not hold capable systems back from difficult work while spending heavily on simple tasks. Model choice becomes part of operational judgment rather than a procurement decision made once.

What it does to you

Repeated AI use changes attention and judgment. People can become anchored on the first answer, overvalue fluent language or stop doing the difficult work through which expertise develops.

Governance therefore needs to protect human capability as well as control machine behaviour.

The Plausibility Illusion

Language models are designed to produce plausible language. Plausibility is useful, but it can arrive before truth.

We call this the Plausibility Illusion: an output feels right because several familiar biases reinforce one another.

• Sycophancy: the model agrees with the user more than the evidence warrants.
• Anchoring: the first draft sets the frame for everything that follows.
• Framing: the wording of the question limits what the system considers.
• Fluency bias: polished language feels more reliable than hesitant language.
• Automation bias: people lower their scrutiny because a system produced the answer.

These effects can move from the individual to the organisation. If a whole leadership team starts from the same generated draft, the draft shapes the discussion before anyone has tested its assumptions.

Five practices help:

1. Ask for the strongest disagreement before improving the favoured answer.
2. Protect the person who says something feels wrong, then investigate the feeling rather than treating it as proof.
3. Keep human effort in the parts of work where judgment develops.
4. Match the model and level of review to the task.
5. Decide what stays human before convenience makes the choice.

The aim is not distrust of AI. It is disciplined use that keeps plausibility from becoming organisational truth by default.

Next: how the Org Mirror reads the working organisation without pretending to produce an objective final answer.

How to see the real organisation

How the Org Mirror uses several lenses without pretending that one model explains everything.

• 24. The Org Mirror
• 25. Morgan: eight metaphors for the same organisation
• 26. Oshry: Tops, Middles, Bottoms and Customers
• 27. Heifetz: technical and adaptive challenges

The Org Mirror

The Org Mirror is an MRI, not surgery. It helps a leadership team see patterns before deciding what to change.

It reads material that shows the organisation in motion: interviews, transcripts, workshop notes, strategy documents and accounts of real work. The org chart provides context, but it cannot be the main source because it describes only the formal organisation.

AI helps compare a large body of material, trace recurring themes and examine the same evidence through several theories of organisation. The result is a set of hypotheses with evidence, not a score or automated recommendation.

The distinction matters. A Mirror that claims to know the organisation better than the people inside it becomes another sovereign interpretation. A useful Mirror gives people something specific to agree with, challenge or correct.

We currently use three main bodies of thinking:

• Gareth MorganWhat kind of system are we looking at, and what does each metaphor reveal or hide?
• Barry OshryHow does the system feel from the position each person occupies?
• Ron HeifetzWhich problems have a known technical answer, and which require people to learn and change?

We are building this with pioneering clients, because we believe the real edge of AI adoption sits here.

Morgan: eight metaphors for the same organisation

Gareth Morgan’s Images of Organization offers eight metaphors. A metaphor is a way of seeing, not a claim that the organisation literally is one thing.

The Mirror uses evidence to show which views explain a pattern and where another view adds necessary context.

Suppose a transformation keeps missing milestones. The machine view may reveal a broken handover. The political view may show that two functions are protecting different incentives. The culture view may explain why nobody names the conflict in the meeting.

The combined reading is more useful than any single diagnosis.

Oshry: Tops, Middles, Bottoms and Customers

Barry Oshry’s Seeing Systems examines the experience created by a person’s position in a system.

These positions are not fixed job grades. A senior executive may be at the Top of the company and in the Middle between a board and a business unit. A frontline manager may be at the Top of a team and at the Bottom of a corporate change programme.

Oshry describes four common experiences:

• Tops feel burdened. They carry responsibility for the whole, receive competing demands and often respond by taking on more control.
• Middles feel torn. They translate between groups with different pressures and can lose connection with other Middles doing the same work.
• Bottoms feel unseen. They experience decisions made elsewhere and may push responsibility upward because they have little room to act.
• Customers feel neglected. Internal groups become absorbed in their relationships with one another while the customer waits outside the system.

The value of this lens is structural compassion. It helps people see that predictable behaviour may come from position rather than personality.

It is especially useful in the Missing Middle. Middles often carry the organisation’s translation work by hand. Mapping where they are torn, where information stops and which decisions keep moving upward reveals architecture that the org chart misses.

Heifetz: technical and adaptive challenges

Ron Heifetz distinguishes between technical problems and adaptive challenges.

• A technical problemhas a reasonably clear definition and a known form of expertise. A broken integration may need an engineer. A missing policy may need a legal decision.
• An adaptive challengerequires people to change priorities, habits, relationships or identity. Expertise can help, but nobody can supply the answer from outside because the people in the system must do the learning.

AI programmes often apply a technical solution to an adaptive challenge. The organisation buys a tool and runs training while leaving the underlying work, incentives and authority unchanged.

The Mirror therefore looks for:

• Misdiagnosis: where the proposed fix does not match the real problem.
• Resistance as information: what a refusal, delay or workaround may be protecting.
• Loss: what people reasonably fear they will give up if the change succeeds.
• Learning conditions: whether the organisation creates enough tension to move without overwhelming people.
• Perspective: whether leaders can step back from daily activity and see the pattern they are helping to produce.

AI makes this distinction more important because the technical answer can arrive quickly. A fast answer still fails when the organisation has not done the adaptive work around it.

From theory to practice

What leaders must decide and which instruments make those decisions concrete.

• 28. Three things change together
• 29. Five decisions for the leadership team
• 30. The AI Constitution
• 31. The instruments

Three things change together

AI changes the work, the team and the organisation at the same time. Treating any one of these in isolation moves the problem elsewhere.

The work changes

AI can rewire a workflow rather than simply accelerate a task. Research, drafting, checking and handover may move between people and systems in a new order. If leaders do not redesign the workflow, the same output arrives faster and somebody downstream inherits the checking and clean-up.

The team changes

AI use often begins privately. Some people become fast and confident, others hesitate, and the difference remains hidden until a handover fails. Teams need shared practices for showing assumptions, challenging output and asking for help. Psychological safety becomes part of the operating infrastructure because people must admit uncertainty about both their own work and the machine’s.

The organisation changes

When workflows and teams change, structures and incentives eventually follow. Roles shrink, grow or combine. Management layers may carry less information-routing work and more responsibility for judgment, conflict and learning.

The employment relationship also changes when the organisation asks people to encode valuable know-how. This is why AI cannot remain a technology programme.

Five decisions for the leadership team

The architecture becomes real through decisions. Five belong at the top table.

1. Fund the middle

Most organisations have budgets for technology and training. Few have a budget for redesigning how work, judgment and knowledge move. Create the line item. Without it, workflow redesign becomes unpaid extra work and the AI strategy remains a licence agreement.

2. Decide what machines may do

Set boundaries for action, information, memory and escalation. Name the decisions that stay human. In organisations with worker representation, bring the works council or equivalent body in early as a co-designer. A policy written without the people who live under it will often drive real use underground.

3. Decide how roles and teams should change

Do not let hiring freezes and isolated efficiency targets design the future organisation by accident. Use evidence from real workflows to decide which roles change, which capabilities grow and where coordination still needs human judgment.

4. Treat model choice as operational judgment

Match capability, cost, privacy and speed to the task. Do not cap powerful models reflexively after one expensive month, and do not use them for light work simply because they are available.

5. Connect every AI seat to real work

Access is not transformation. Give people AI capability alongside a real workflow they are expected and supported to improve.

The unit of progress is the amount of work that has become better, safer or newly possible.

The AI Constitution

The AI Constitution is the practical expression of the Constitution layer for AI-enabled work. It is not a longer acceptable-use policy. It connects strategic purpose, decision rights, boundaries, contribution and accountability.

Three groups author different parts:

The chief executive or most senior accountable leader

They write the manifesto. They state why the organisation is using AI beyond generic productivity. They connect the technology to strategy and name the outcomes that matter.

The leadership team

They author and own the constitution, turning that purpose into operating commitments: what the organisation will use AI for, what it will never delegate, which evidence and thresholds guide decisions, how people will challenge a machine-supported decision, who is accountable at each level, and how the Constitution will change when the organisation learns.

IT, Legal, Risk and worker representatives

They turn commitments into workable guardrails, controls and rights. Their involvement begins early because architecture designed without operational, legal or employee reality will not survive contact with the organisation.

They also open with an AI amnesty. Shadow AI is the elephant in the room: people from the top floor to the front line use unsanctioned AI daily, not out of recklessness, but to keep themselves and your organisation from falling further behind. Punish it and it goes deeper underground. Declare amnesty, learn from what surfaces, then offer guided freedom inside clear red lines.

The Constitution should also address contribution. If people help encode the tacit knowledge that makes AI useful, the organisation should state how that contribution is recorded, recognised and rewarded.

We are building this with pioneering clients, because we believe the real edge of AI adoption sits here.

Download the Constitution framework (PDF) →No email required. Proof before permission works both ways.

A constitution is authored, not downloaded. The framework is food for thought; the document your organisation will actually follow has to be authored, crafted and owned by your leadership team, and kept alive by them. We facilitate. You hold the pen.

The instruments

Challenger House is developing a set of instruments that sit at different layers of the architecture. An instrument gives a leadership team something specific to examine, discuss or change. It is not a maturity score.

• Org Mirror · SensoriumReads the working organisation through Morgan, Oshry and Heifetz. It shows recurring patterns and the evidence behind them so that leaders can challenge the reading.
• AI Constitution · ConstitutionRecords purpose, decision rights, non-delegables, guardrails and the commitments attached to knowledge contribution. It is authored by the organisation rather than downloaded as a generic template.
• Delegation Dial · BoundariesHelps a team decide how much of a specific activity should be human, augmented, agent-led, supervised or autonomous. It makes consequence and accountability part of the design.

• Talking Tax · Boundaries and SensoriumEstimates how much management time is spent chasing, translating and repackaging information instead of acting on it. It gives invisible coordination work an economic value.
• Attention Constitution · Constitution and SensoriumDefines what the organisation notices, remembers and treats as evidence. It also creates a route for alternative interpretations.
• Information latency · Sensorium and AdaptationMeasures how long reality takes to reach someone who can respond. It shows where delay, context loss and unnecessary reporting have entered the architecture.

Some of these instruments are live and others remain in development. Nothing becomes settled Challenger House doctrine until real client work has tested it.

Where this may lead

A testable view of organisations that can sense and revise their own architecture.

• 32. Self-revising, not AI-native
• 33. AI changes who holds power
• 34. When networks become cheaper than hierarchies
• 35. The Tower and the Square form a cross
• 36. What is your organisation optimising for?

Self-revising, not AI-native

“AI-native” usually describes an organisation that uses AI extensively. That tells us little about whether the organisation is well designed.

We prefer self-revising. A self-revising organisation can notice when its architecture no longer fits reality and change it before a crisis forces the issue. The loop is straightforward:

1. Sense: notice what is happening inside and outside the organisation.
2. Interpret: compare several explanations and preserve uncertainty where it matters.
3. Decide: make a choice with clear authority and accountability.
4. Act: change the work within defined boundaries.
5. Learn: compare the result with the intention and update the architecture.

Stafford Beer used the language of cybernetics, the study of how systems steer through feedback. The word can sound technical, but the idea is familiar. A thermostat senses a difference, acts and senses again. An organisation must do this while also dealing with purpose, politics, interpretation and human consequence.

A self-revising organisation does not constantly reorganise. It changes the right layer. A workflow problem may need a local adjustment. A repeated conflict may reveal a constitutional contradiction. The skill lies in knowing the difference.

AI changes who holds power

Every information system distributes power. It determines who can see, who can interpret and who can set direction.

Bard and Söderqvist describe three broad functions of power. Their labels are less important than the questions behind them:

• Who can gather and process information? In the industrial era, capital bought factories and reporting systems. Today, platforms and models can gather and connect information at enormous scale.
• Who can sense what people want or fear? Mass media once interpreted public attention periodically. Digital systems sense behaviour continuously.
• Who can set direction? Institutions and experts once held much of this authority. Networked groups and system designers can now shape the field in which decisions occur.

The industrial corporation often held all three functions inside one hierarchy. The same leadership structure owned the information, interpreted it and decided what should happen.

AI can concentrate these functions further. A single platform may store the data, generate the interpretation and recommend the action. It can also distribute them by giving more people access to evidence, modelling and coordinated action.

Neither outcome is automatic. Architecture decides which one the organisation gets.

When networks become cheaper than hierarchies

Ronald Coase asked why firms exist when markets can coordinate work. His answer was transaction costs.

Finding a supplier, negotiating a contract, checking quality and enforcing an agreement all take time and money. When coordinating through the market costs more than coordinating through a manager, the firm brings the work inside.

Walter Powell later argued that networks are a distinct form of coordination, not merely a midpoint between markets and hierarchies. Networks rely on trust, repeated relationships, shared knowledge and reciprocity.

AI lowers some of the costs that made hierarchy attractive. Systems can find capability, compare offers, coordinate tasks and monitor agreements more cheaply than before. The firm survives this. What changes is the boundary between what a company keeps inside and what a network can coordinate.

The likely result is not a world without managers. It is pressure on management work that mainly routes information, checks routine compliance or coordinates transactions a system can now handle.

Human management remains where the work involves identity, conflict, judgment, accountability, trust and decisions about the whole.

The Tower and the Square form a cross

The Tower and the Square are not separate worlds with an empty space between them. They cross inside every real role.

The vertical axis carries formal authority, accountability and resource allocation. The horizontal axis carries trust, relationships and real-time coordination.

Most leaders receive extensive training for the vertical axis. They learn budgets, reporting lines, governance and formal decision rights. They receive far less help with the horizontal axis, even though much of the organisation’s real coordination happens there.

The Missing Middle sits at the crossing point.

A middle manager, for example, may formally own a budget and report to a director. The same person may also connect engineers, customers and external partners who do not report to them. Their effectiveness depends on using both axes without confusing them.

New organisational architecture must make this crossing visible. It should say when formal authority decides, when a network must be convened and how learning from the network changes the formal system.

This synthesis is still young. It should be tested against real organisations before it becomes doctrine.

What is your organisation optimising for?

An objective function is the result a system is designed to maximise or minimise. The term comes from mathematics, but every organisation has an informal version.

It may optimise for quarterly margin, reliability, customer retention, growth, safety or the avoidance of visible failure. The stated objective and the operating objective are not always the same.

AI does not correct a badly chosen objective. It helps the organisation pursue that objective more consistently and at greater speed. This leaves leadership with two connected questions:

1. What is our organisation actually optimising for, and what happens when AI makes us much better at it?
2. Will our AI architecture concentrate power in a smaller group, or distribute useful agency to the people who remain answerable for the result?

Different organisations should answer differently. Variety creates resilience. If every company adopts the same models, measures and default settings, many organisations may become fragile in the same place at the same time.

Whoever decides what machines may notice, remember and act on is shaping the field in which everyone else works.

That is why organisational architecture belongs with leadership.

Sources and intellectual foundations

This architecture is interdisciplinary on purpose. The sources below inform it; none endorses Challenger House, and we do not agree with every part of their work.

Organisational development and leadership

• Gareth Morgan, Images of Organization. Eight metaphors that form the Org Mirror’s first lens.
• Barry Oshry, Seeing Systems. Tops, Middles, Bottoms and Customers, the Mirror’s second lens.
• Ronald Heifetz and colleagues, adaptive leadership. The distinction between technical problems and adaptive challenges.
• Edgar Schein, organisational culture and leadership. How leaders shape culture through what they notice, measure and reward.
• Peter Koenig, Source Principle. The person who carries the originating impulse of an initiative.
• Tom Nixon, Work with Source (2022). A structured account of Koenig’s research and its organisational implications.

Cybernetics and systems

• Stafford Beer, Viable System Model. Five functions a system needs to remain viable, recursion and direct escalation signals.
• W. Ross Ashby, requisite variety. A regulator needs enough variety to respond to the variety in its environment.

Networks, firms and power

• Niall Ferguson, The Square and the Tower. Formal hierarchy and living networks as two structures of power.
• Ronald Coase, “The Nature of the Firm”. Transaction costs and the boundary between market and hierarchy.
• Walter Powell, network forms of organisation. Networks as a distinct coordination form.
• Alexander Bard and Jan Söderqvist. Information paradigms, changing forms of power and the philosophical background to the print-age organisation.
• Indy Johar, “The Rise of Meta-Roles” (17 August 2026). Operational action, constitutive action and the power involved in configuring systems.

Knowledge and practice

• John Vervaeke, four kinds of knowing. Propositional, procedural, perspectival and participatory knowing.
• Charles Handy. The character and assumptions of organisational architects appear in the structures they create.

Legal references current at 19 August 2026

• German Limited Liability Companies Act, section 43: directors’ duty of care and liability
• European Commission: AI Act implementation timeline and risk categories
• EUR-Lex: Regulation (EU) 2024/1689, the Artificial Intelligence Act

The legal references provide context, not legal advice.

Practice

The architecture also draws on twenty-five years of work inside large organisations, including Microsoft, Telefónica, Sanofi and E.ON. These names indicate the environments in which the thinking developed. They do not imply endorsement of every claim on this page.

Challenger House · Building new organisational architectures for the AI transition. We spark courage to act now.