From the information-processing theory of the firm to the axes of organizational design — and their extension into AI-augmented and AI-native work.
Four parts · use the sidebar to navigate
Organization theory is not a pile of unrelated models. It is one long argument that slowly converged on a single idea: coordinating knowledge work is fundamentally a problem of processing information under uncertainty. Here is the line of thought.
Four moves turned a theory of individual decision-making into a theory of organizational design. Read left to right.
Simon: people have limited attention and compute, so they satisfice rather than maximize.
March & Simon; Cyert & March: slack, sequential attention, loosely coupled sub-units.
Galbraith: reduce the need for information, or raise the capacity to process it.
The canon extends into ties, tacit knowledge, modularity and post-bureaucratic forms.
Fourteen works that define the field. Click any point on the timeline to read why it matters.
Each node is a foundational contribution. Together they trace the shift from "the firm as a production function" to "the firm as a knowledge-creating, coordinating system." Click a node to explore.
The synthesis groups the literature into twelve clusters. Expand each to see the core idea and its key thinkers.
At the whole-organization scale, if structure is a set of choices, what are the choices? A common starting point uses three axes — decision rights, communication channels, and knowledge. The literature says that is valid but incomplete. Below: a four-spine core, three orthogonal modifiers, and the contingencies that decide which setting is right.
Move the four core sliders — or load a known archetype — and watch the profile update. These four "spines" are the cleanest, most independent design levers.
Adjust the sliders to characterize a design.
The literature throws up a dozen candidate "dimensions," but they aren't all independent. Consolidating for statistical distinctness leaves eight genuine design dimensions — plus two contingency inputs the design must fit rather than choose.
| Design dimension | Poles | Type |
|---|
Spine independent core axis Modifier orthogonal add-on Composite merges several correlated dimensions
These describe the work, not the design. You don't set them — you fit the eight dimensions to them.
| Input | Range | Role |
|---|
Because the tradeoffs invert with the work, there is no universally best setting. Pick the nature of the task and see how the design should shift (Galbraith / Daft & Lengel contingency). Uncertainty is missing data; equivocality is disagreement about what the data means — they demand different remedies.
The clearest evidence that these are choices with consequences, not abstractions, comes from Hansen, Nohria & Tierney's study of consulting firms (HBR 1999).
Invested heavily in document repositories and reuse: solutions codified once, sold many times, staffed with leveraged junior teams. Won on scale and price for repeatable problems.
The strategy only pays where volume and repetition are high — the contingency table above, in action.
Invested in networks of people, not documents: knowledge moves person-to-person, staffing is expertise-matched, fees carry a premium. Won on bespoke judgment for one-of-a-kind problems.
Hansen's rule: pick a dominant strategy (~80/20 mix). Firms that straddled the middle underperformed on both.
The macro axes describe the whole organization. But most real work happens in pairs — two peers, a manager and a report, and now a human and an AI. The dyad is its own design unit, and it is where AI actually lands day to day.
They share a spine but differ in their primary risks and mechanisms. Pick a type to see its profile.
Synthesis across theories yields a compact set of properties — each with a validated measure. (A fuller ~13-construct set is in the dyad synthesis.)
Design moves that repeatedly show up across the literature. Expand each.
Shares the human dynamics (grounding, trust, role clarity) but adds its own apparatus: joint cognitive systems, levels of automation, and the calibration of reliance.
The twin failure modes are over-trust (using AI when it's wrong) and under-trust (ignoring it when it's right). The goal is appropriate reliance — deployed systems sit well below the ideal (RAIR ≈ .54, RSR ≈ .64).
Human correctly updates to good AI advice — raises RAIR.
Human follows incorrect AI advice.
Human ignores correct AI advice.
Human overrides wrong AI advice — raises RSR.
A staged playbook. Click a stage.
From diagnosis to redesign-or-exit. Click a stage above to see what it involves.
The information-processing view (Part 1) assumed human decision-makers with fixed cognitive limits. AI changes the cost, speed, and division of the organization's core loop — how it senses, makes sense, decides, and acts — and how it adapts that loop over time. The design question becomes: who runs each part of the loop, how fast does it turn, and what bounds its autonomy?
The core cycle every organization runs (Boyd's OODA), oriented by intent, drawing on memory, wrapped by governance, and adapted by a meta-loop. Click any stage or element.
Four operating stages turn continuously — oriented by intent (the reference signal), fed by memory (state), adapted by the meta-loop, and bounded by governance. The human/AI division of labor differs at every stage; click to explore.
The human ↔ AI division of labor is not one setting — it differs stage by stage. Assign each stage and see whether it fits, and the reliance risk.
How much of a stage AI can run depends on the work — the operational form of the Part 2 uncertainty/equivocality contingency. Slide it.
Part 3 gave the dyad validated instruments and thresholds. The loop deserves the same discipline: if the loop is the model, these are the gauges. None of these is a validated scale yet — they are operational proxies, DORA-style, and should be treated as a starting instrument panel, not a standard.
| Gauge | What to count | What it tells you |
|---|---|---|
| Signal-to-action latency | Elapsed time from a material signal (market, regulator, incident, customer) to a changed action in the field | How fast the operating loop actually turns — the headline cadence number |
| Decision latency & queue depth | Age and count of decisions waiting on a named human or committee | Where decision rights are mis-assigned; which gates are the bottleneck |
| Experiment throughput | Process changes proposed / trialed / institutionalized per quarter (mutate → select rate) | Whether the meta-loop turns at all — and whether Select keeps up with Mutate |
| Eval coverage | % of AI-run workflow steps with automated evaluation and monitored quality | The feasible ceiling on autonomy — you cannot open the loop past what you can see |
| Reliance calibration (RAIR / RSR) | Sampled accept/override decisions on AI output, scored against ground truth | Over-trust vs. under-trust per loop stage — the human side of the ceiling |
| Time-to-institutionalize | Time from a validated improvement to it becoming the default workflow | Governance drag: distinguishes deliberate locks from mere organizational friction |
Reading the panel: latency gauges price the operating loop, throughput and institutionalization gauge the meta-loop, and coverage and calibration bound how far governance can safely open either. A locked process with fast institutionalization is a choice; a locked process with slow everything is decay.
Why the loop is changing at all: AI lowers the cost of running each stage and of coordinating across people, agents, and firms (Coase, Galbraith). The master variable is whether you use it to automate or augment — Brynjolfsson's "Turing Trap." Toggle it.
When execution gets cheap, the bottleneck — and the value — migrates. Where it goes:
A second-order loop watches the operating loop and changes it — observe, evaluate, mutate, select. Single-loop learning fixes actions within goals; double-loop revises the goals and assumptions themselves (Argyris). AI compresses this from months to minutes. Click a function.
Unlike biological evolution — random mutation across generations — this runs within the organization's lifetime, with directed mutation and conscious selection. Click a node.
Self-modification without governance is pathological — in biological terms, cancer. Every process sits on a spectrum from locked to open, deciding how much the loop may change without human authorization. Classify each; see if it fits.
Mapping to machinery you already own: in regulated firms, Locked / Experimental / Open is not a new control framework — it maps onto model-risk tiers (SR 11-7), the three-lines-of-defense structure, and EU AI Act risk categories. "Locked" processes are those your second line already attests; "Open" is what internal audit has classified as immaterial. Extend the existing machinery; do not duplicate it.
Each stresses a different element of the control system. All four are from public reporting; none involves exotic technology — which is the point. The failure mode is organizational, not algorithmic.
A manual deploy left obsolete order-routing code live on one server; the automated loop executed it at full speed. ≈ $440M lost in 45 minutes; the firm did not survive independently.
Pre-LLM, and precisely why it matters: any loop that acts autonomously needs change-control and a kill-switch before it needs intelligence. "What must not change" is question four for a reason.
The airline's support chatbot invented a bereavement-fare policy; a tribunal held the airline liable for its agent's answer, rejecting the argument that the chatbot was "a separate legal entity."
The clean legal statement of this module's rule: accountability stays human (and corporate) even where AI acts. Question two — who owns the outcome — is not optional.
A model-driven buying loop kept acting on forecasts as the housing market shifted; evaluation lagged action. Wind-down with > $500M in write-downs and a 25% workforce reduction.
The loop sensed and acted fast but the Evaluate–Select functions could not keep pace. Speed of learning means the whole loop — accelerating half of it is worse than accelerating none.
Announced its assistant did the work of ~700 agents; a year later leadership acknowledged quality had suffered and began re-hiring humans for higher-touch service.
Not a failure of automation but of the matching rule: codified tickets fit "AI executes"; ambiguous, trust-sensitive ones needed the co-pilot configuration. The correction — re-tuning the mix — is the meta-loop working.
The macro spines (Part 2) still apply — but AI shifts their poles, costs, and contingencies. Toggle the view.
Part 2 established three orthogonal modifiers. Boundary permeability got its renegotiation above (Coase). The other two — incentives and psychological safety — are hit hardest by AI and discussed least. Same toggle applies.
Nonaka's SECI spiral assumed converting tacit to explicit knowledge was slow, human, and costly. AI shifts several of these conversions — with real risks attached.
Leadership shifts from designing fixed processes to governing a dynamic system — setting the reference (intent and identity), tuning who runs each stage of the loop, and bounding the meta-loop's autonomy. The thesis is not maximal adaptability (that dissolves into chaos) but governed adaptability: continuous adaptation within constraints that preserve coherence, accountability, and identity — homeostasis, not rigidity.
Open questions the 1970s categories can't answer on their own:
The four pushbacks every executive raises — plus the one this framework raises against itself. Expand each.
A framework you cannot apply is a lecture. The module ends where your decision starts: Part 5 embeds the diagnostic instrument that operationalizes everything above.
Twenty-two questions about one unit's work, constraints, current operating baseline, and intent — never its structural preferences — returning target, feasible-now, the three best-fitting AI-organization configurations, rank, calibrated band, distance gap, tensions, exclusions, influences, and roadmap. Run it per unit and read the results as a portfolio.
Continue to Part 5 — run the diagnostic →Also available standalone ↗ for sharing with a respondent who shouldn't need the course first.
Everything so far becomes actionable here. The instrument below asks twenty-two questions about one unit's work, constraints, current operating baseline, and intent — never about its structural preferences — and returns target, feasible-now, the three best-fitting AI-organization configurations, rank, calibrated band, distance gap, tensions, exclusions, influences, and roadmap.
Answer for a single business unit or workflow family, not the whole firm. Fit varies by task — that is Part 2's entire argument. A bank's regulatory-reporting unit and its research desk should get different answers.
The respondent should see the work directly, not only its reporting. Better still, have two roles answer independently — where they disagree, you have just measured equivocality (Part 2) rather than opinion.
Re-run quarterly or after any meta-loop change; a configuration is a setting, not an identity. And log every "none of these fit" — clusters of misfits are how the preset library earns its next archetype.
The configurator opens in its own clinical styling — deliberately: it is an instrument, not a lecture. It scrolls independently below. Open it standalone ↗ to share with a respondent.
Engine v0.2 — treat rank, calibrated band, raw-distance gap, tensions, exclusions, and roadmap as the payload. The instrument's generated design spec and known limits live in ai-org-configurator.md alongside it.