Chapter 3 — The method at a glance¶
The entire method fits on one page. This chapter is that page, expanded.
What follows is the whole shape of the book, walked at a slow pace. Two gates sit above the method. Five gates sit inside it. A toolbox spreads across the gates. Three indices and ten overlays stand to the side. Every later chapter does one of these things in more detail; nothing is added that is not shown here first.
3.1 The method on one page¶
See Illustration 3.1.
Illustration 3.1 — The method at a glance. Two Tier-0 checks sit above five Tier-1 gates. The toolkit carries a tool palette for each gate; the AI capability levels are picked only at G3. Loopbacks from G3 back to G2 are cheap and expected.
The top band holds the two Tier 0 gates. The row below it is Tier 1 — the five triage gates, left to right, in the order they run. Below the row is the toolkit strip, which carries a tool palette for each gate — diagnostic frames for G1, decomposition frames for G2, the AI capability levels for G3, sequencing lenses for G4, ownership discipline for G5. To each side are quiet rails pointing to the three access indices and the ten overlays. Those are for later. The method itself runs without them.
The illustration shows the whole shape of the method; nothing else is being revealed on the page.
3.2 Tier 0 — the two checks¶
Two checks stand above the method. They are not refinements of it. They decide whether the method runs at all.
The first is the Technochauvinism check [1]: is this a technical problem at all, or a social problem renamed? Both answers are valid; the check forces the question to be asked before a tool is reached for. Failing it produces Not a tech problem — policy, process, or people, not tools of any kind.
The second is the Abolition check, drawn from McQuillan's Resisting AI [3]. The one-sentence test: is this decision in a category where no plausible set of controls could make AI delegation acceptable — regardless of future model capability? Predictive policing, automated welfare sanctioning, immigration risk scoring, automated sentencing. For categories like these, the honest answer is that better governance is not the fix. The test is categorical, not case-by-case, and answerable from the brief alone. Failing it produces Off-limits to AI.
A note on what the Abolition check is not. The obvious broader test — should any judgment-heavy decision be handed to a machine? — is Weizenbaum's 1976 line [2], and it is tempting to place at Tier 0 alongside the two checks. This book does not. Weizenbaum mixed two claims: that machines could not compute judgment-heavy decisions, and that they should not, because accountability depends on a human bearing the outcome. The capability half has not held up; a modern model can produce sentencing-style reasoning, therapy-style dialogue, custody-style weighing credibly. The normative half has held up. But it is too wide to sit above the method as a pre-check: most problems need decomposition and routing before anyone can honestly answer whether the controls available — audit trails, human review, rollback triggers, sunset criteria — are adequate for the specific decision in front of us. That engineering question belongs inside the method, at G3 Route, and is threaded through Sequence and Commit as a cross-cutting controls overlay (see Chapter 3.5). Tier 0 keeps only the categorical refusal. The split — narrow categorical refusal above, controls viability inside — is this book's modification of Weizenbaum, not a restatement of his line. He would likely have kept both halves above the method; this book does not, because without the split the pre-check refuses problems that careful engineering genuinely does make routable.
The effect on the method is simple. Everything else waits for the two Tier-0 checks, which can produce two outcomes: Off-limits to AI (the Abolition check fails) or Not a tech problem (the Technochauvinism check fails). If both checks pass, Tier 1 opens. Four further outcomes are available inside the method, all produced at G3 Route when each piece from the decomposition is matched to the substrate that should carry it: Human-operated (no automation earns its place; the work stays with a person), Non-AI automation (a rule, a script, or a constraint solver is the right substrate — deterministic behaviour, inspectable logic, the code as audit path; classical ML sits above the Non-AI / AI boundary and routes through the next two outcomes), AI as assistant (an AI level fits but no level clears the controls bar for machine-held responsibility — a human holds the decision and AI supports with retrieval, summarisation, or option generation), and Autonomous AI (an AI level clears the controls bar — the machine makes the call under the controls designed for it). Six outcomes in all, produced at three places — the two Tier-0 checks, and G3. Four of the six are non-AI outcomes; that distribution is not a quirk of the taxonomy, it is a reflection of what honest routing work actually finds.
Chapter 4 carries the full treatment — the four "dressed-as-tech" patterns at the Technochauvinism check, the four axes at the Abolition check, the tools each uses to produce its one-paragraph determination.
A word on G1 Observe. Illustration 1.1 in Chapter 1 shows the six answers; Illustration 3.1 shows the five gates. G1 folds what older method traditions split into framing (stating the problem in the world, not the brief's version of it) and diagnosis (finding the system the symptom lives in) — because both produce one artefact, the situation statement, through one iterative pass. Chapter 5 treats the craft of framing and diagnosis together.
3.3 Tier 1 — the five triage gates¶
Inside the method, five gates run in order. Each gate has a name, one question, and one artefact it produces. Nothing more.
G1 Observe. What is actually happening? Most briefs arrive pre-diagnosed by the person writing the brief. G1 undoes that. It goes to the work — the operator's desk, the queue, the actual customer call — and writes down what is happening in the operator's or the customer's words, not the sponsor's. The output is a short paragraph: what is happening, to whom, since when, with what second-order effects, what changed recently. This paragraph is the situation statement. It is the only artefact G1 produces. Hammer's warning about automating the cow-path [4] applies here: you cannot route a problem you have not seen.
G2 Decompose. Independent pieces, or one tangled thing? G2 takes the situation statement and breaks it into pieces that can be routed separately. The backbone is Minto's pyramid [5] and its MECE discipline — mutually exclusive, collectively exhaustive. Specialist frames sit under the backbone: Ishikawa [6] for multi-cause brainstorms; 5 Whys, attributed via Ohno [7] to Sakichi Toyoda, for simple chains; Ulwick's operational Jobs-to-be-Done [8] when the shape of a solution is unknown; fault trees when failure is safety-critical. The output is an issue tree: the problem as a set of independent sub-problems, each small enough to route.
G3 Route. Where does each piece belong, and can we build the controls for that routing? The verb is chosen on purpose. Every gate is named after an action, and G3's action is routing: each piece arrives from G2 with no default destination, and something must be sent somewhere. The alternative — placement — is passive, and the whole point of G3 is that no piece rests in a natural spot; the destination is chosen. Chapter 7 opens with a fuller note on the word. This is the gate most engagements fail at. Given the pieces from G2, each piece is routed to the right computational substrate — a rule, a statistical model, classical ML, an LLM feature, retrieval-augmented generation [9], a single agent, a tool-using agent, a multi-agent system — or handed back to a human. The default is not agentic. Each sub-problem needs a positive case to be routed above rules or statistics, and a plausible controls design for the route — audit trail, human review point, rollback trigger — before the route is accepted. Four outcomes are possible at G3, one per piece. If no automation earns its place against a person doing the work, the piece produces the Human-operated outcome. If a rule, a script, or a deterministic solver fits, the piece produces the Non-AI automation outcome — repeatable behaviour, inspectable logic, the code itself as audit path. Classical machine learning, though simpler than an LLM, sits above the Non-AI / AI boundary because it learns from data; a classical-ML route is an AI route and routes through the next two outcomes. If an AI level fits but no level clears the controls bar for machine-held responsibility, the piece produces the AI as assistant outcome — a human holds the decision; AI supports with retrieval, summarisation, or option generation. If an AI level does clear the controls bar, the piece produces the Autonomous AI outcome — the machine makes the call under the controls designed for it. The output is a routing map: each piece with its destination named, its controls design sketched, and its route justified. The AI capability levels are picked at this gate and only this gate — earlier gates use their own analytical tools (observation discipline, decomposition frames); G3 is where the computational level is chosen.
A note on what these outcomes do not say. G3 names the runtime substrate for each piece — what makes the decision when the system is live. The outcome does not dictate how the substrate was built. A piece routed to Non-AI automation may have its rules drafted with an LLM copilot; a piece routed to Human-operated may have its training material, its onboarding aids, or the operator's lookup tool generated or supported by AI. That is Chapter 2's Building mode, not this method's subject. The method governs deployed use; build-time AI use is legitimate, common, and out of scope. A Non-AI automation outcome does not mean AI is banned from the project — only that the runtime decision does not ride on a model.
G4 Sequence. In what order, given dependencies? G4 draws the dependency graph across the routed pieces. Three lenses read the graph: what depends on what; what is reversible and what is not; what the blast radius is if a piece fails. A piece that is irreversible and high-blast-radius runs last, behind a shadow-mode pilot and a staged rollout. A piece that is reversible and contained can run first, even if it is the larger piece. The output is a dependency order — a numbered sequence with reversibility and blast-radius notes attached.
G5 Commit. Who carries the decision? A commitment without a named owner is not a commitment. G5 names one person — not a committee, not a role, not a rotating seat — who owns the outcome, controls the resources to change course, and signs the rollback triggers. The output is a one-page document: owner, triggers for rollback, review cadence, sunset criteria. If the gate cannot produce that page with a real name on it, the engagement does not cross into implementation.
The five gates run in order, but they are not one-way. The arc on Illustration 3.1 from G3 back to G2 is doing real work. Routing often discovers that the decomposition was wrong — a piece that looked independent turns out to share a data pipeline with another, or a piece that looked routable to rules turns out to hide a judgment. When this happens, the loopback is cheap. Pretending it did not happen is expensive.
3.4 The AI capability levels enter at G3¶
Every gate has tools, and every gate's tools live in the toolkit. G1 draws on observation discipline, the Critical Decision Method, ShadowBox, and the stakeholder map. G2 draws on Minto's pyramid, MECE, Ishikawa, 5 Whys, Jobs-to-be-Done, and fault trees. G4 draws on dependency graphs, reversibility classes, and blast-radius estimation. G5 draws on named-owner discipline and rollback-trigger design. These are analytical tools — craft techniques for producing each gate's artefact — and they run where their gate runs.
The narrower claim, and the one that matters: the AI capability levels — rule, classical ML, LLM feature, retrieval-augmented generation, single agent, tool-using agent, multi-agent — enter at G3, and only at G3. No piece is routed to an AI level before G3. This is deliberate, and it is worth one paragraph of defence.
Routing errors swamp level-choice errors. A correctly chosen LLM, pointed at the wrong sub-problem, produces a well-evaluated answer to a question nobody asked. A correctly framed sub-problem, pointed at an imperfect level, produces a close-to-right answer that a human can repair. The first is expensive to detect and more expensive to fix. The second is visible and cheap. The method holds the AI capability levels back until G3 because the cost of a wrong level is bounded by the cost of replacing it; the cost of a wrong route is bounded by the size of the engagement.
The five most common routing errors sit together because they are variants of the same mistake: choosing an AI level before the routing is clear. Promoting a rule-based piece to an LLM because LLMs are newer. Demoting a judgment piece to a classifier because the classifier performs well on the training distribution. Reaching for an agent when a single function call would do. Routing to classical ML without a label budget. Routing to retrieval-augmented generation when the knowledge base is not retrievable-quality yet. The toolkit entries for each of these levels say this in the when not to use section. Chapter 7 treats them one by one.
The toolkit itself is not a glossary. Every entry — analytical tool or AI level — follows the same eight-section template: purpose, anatomy, example, pitfalls, when not to use, provenance, related tools, verification tag. Twenty-six entries recur across the method, distributed across the five gates. They are the toolbox — no more, no fewer.
3.5 Overlays and indices¶
Two quiet bands sit to the side of the main method. Most engagements do not need them on first pass. Large or regulated engagements cannot function without them.
On one side are the three access indices — the subject of Chapter 10. They help decide where a problem enters the method. The first index is task codifiability: a spectrum from tasks with explicit rules or labelable outcomes (payroll) to tasks that rely on judgment (settling a contested divorce). The second is weight class: featherweight tools (a 5 Whys takes minutes) to industrial programmes (ISO 42001 [10] certification takes months). The third is starting points — a catalogue of eight common entries into an engagement, from a new build to an incident review to a compliance mandate. A problem triaged through all three indices arrives at G1 with its rough shape already named.
On the other side sit the ten cross-cutting overlays — the subject of Chapter 11. Each overlay is a discipline that most teams re-invent badly and that runs across the whole method rather than living at one gate. The ten are: a data readiness gate; evals-as-code; level-indexed total cost of ownership (the published agent-multiplier figure is roughly ten to twenty times a single LLM call); a three-stage rollout pattern (shadow, canary, progressive); an adaptation decision tree (prompt, then retrieval, then fine-tune, then agent, in that order, with stop rules); a privacy control ladder; the NIST [11] and ISO governance spine; a retirement protocol; the interaction-design stack (HAX [12], PAIR, Shneiderman's two-dimensional human-centred AI frame [13]); and data contracts. Each overlay has a home chapter or toolkit entry.
Several of these overlays — evals-as-code, the three-stage rollout, the governance spine, the retirement protocol, the privacy ladder — together form what this book calls the controls discipline: the engineering answer to the question can adequate safeguards be built to make AI delegation acceptable for this specific problem? That question is first sketched at G3 Route (the controls design that justifies the level choice), committed to at G4 Sequence (the rollout gates, rollback triggers, and staging that let the controls actually bite), and closed at G5 Commit (the named owner, review cadence, and sunset criteria that keep the controls alive). The controls discipline is the book's modification of Weizenbaum's normative line: where he placed a single categorical refusal at the gate, the book places a categorical refusal and a case-by-case engineering discipline that runs across three gates of the method. The second is an extension of his work, not his original claim.
A closing note on scope. The indices and overlays are the reason this is a 220-page book and not a 40-page pamphlet. They hold the method up under realistic conditions — regulated industries, messy data, governance regimes, long time horizons. They are for later use. The method runs without them.
3.6 How the rest of the book is shaped¶
Part 2 teaches the method in the order it runs. The two Tier-0 checks come first; then the five Tier-1 gates — G1 Observe (framing and diagnosis folded), G2 Decompose, G3 Route, then G4 Sequence with G5 Commit. Part 2 is the book's spine.
Part 3 adds the views that sit across the method: the three access indices, the ten overlays, the five governance failure modes, and a chapter on retirement — what to stop doing. Part 3 is where the method meets realistic conditions.
Part 4 is the toolkit. Twenty-six tool entries, each in the same template. Read once in order if you have never met the tools before. Thereafter, look up entries on demand.
Front matter and back matter are thin. A preface, a reader's guide, a glossary, and a compiled sources list. They earn their space by being short.
Chapter 4 opens Part 2 with the two Tier-0 checks — Technochauvinism and Abolition — and the tools each uses to produce its one-paragraph determination. Chapter 5 picks up G1 Observe; in practice, the framing conversation and the Tier-0 checks usually happen side by side, but the book treats them in separate chapters to keep the work of each visible.
Sources¶
[1] Broussard M. Artificial Unintelligence: How Computers Misunderstand the World. MIT Press; 2018. [verified]
[2] Weizenbaum J. Computer Power and Human Reason: From Judgment to Calculation. W. H. Freeman; 1976. [verified]
[3] McQuillan D. Resisting AI: An Anti-fascist Approach to Artificial Intelligence. Bristol University Press; 2022. [verified]
[4] Hammer M. Reengineering Work: Don't Automate, Obliterate. Harvard Business Review. 1990 Jul–Aug. [verified]
[5] Minto B. The Pyramid Principle: Logic in Writing, Thinking and Problem Solving. Pitman; 1987. [verified]
[6] Ishikawa K. Guide to Quality Control. Asian Productivity Organization; 1968 (Japanese) / 1976 (English). Diagram dates to 1943 at Kawasaki Steel. [verified]
[7] Ohno T. Toyota Production System: Beyond Large-Scale Production. Productivity Press; 1988. [secondary]
[8] Ulwick A. What Customers Want. McGraw-Hill; 2005. [verified]
[9] Lewis P, Perez E, Piktus A, et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In: NeurIPS; 2020. arXiv:2005.11401. [verified]
[10] ISO/IEC 42001:2023. Information technology — Artificial intelligence — Management system. International Organization for Standardization; 2023. [verified]
[11] NIST. AI Risk Management Framework 1.0 (NIST AI 100-1). National Institute of Standards and Technology; 2023. [verified]
[12] Amershi S, Weld D, Vorvoreanu M, et al. Guidelines for Human-AI Interaction. In: Proc. CHI 2019. DOI:10.1145/3290605.3300233. [verified]
[13] Shneiderman B. Human-Centered AI. Oxford University Press; 2022. [verified]