Chapter 1 — What this book is

Two acts keep needing a human.

The first is framing. Deciding what the problem actually is. Deciding who it is a problem for. Deciding what would count as having solved it. A model given a badly framed problem produces a confident, well-written, wrong answer faster than any human could. Framing is not a bottleneck that better tooling will widen — it is the act that decides what the tooling is pointed at.

The second is deciding what a machine should and should not decide. Some decisions should not be automated at all, regardless of how capable the model gets. For everything else, the question is whether adequate controls — audit trails, human review, rollback triggers, sunset criteria — can be built to make the machine's decision trustworthy enough to delegate. Neither question answers itself, and both need a named human doing the answering.

So the real question isn't whether to use AI. It is is this an AI problem — and if so, where should a machine fit in the decision? Most AI books answer a different question, how do I build this, and presume the first one is already settled. This one doesn't. The reader shows up with a brief. The brief contains a name — a chatbot, an agent, a copilot, a retrieval system, a classifier — and an expectation that the name is the answer. The first move is to refuse that shortcut and ask what the problem actually is.

The book is a method for doing that. Five gates, two checks above them, a toolbox of twenty-six entries spread across them. The method produces one of six honest answers for any given problem. It does not pre-commit to "yes, use AI." It also does not pre-commit to "no." It tries to produce the right answer for the problem in front of it, and it shows its working.

1.2 What you get from reading it

A method for deciding whether — and how — to use AI on a specific problem.

The method is built for one kind of AI use — AI deployed inside an organisation, where the operator of the system and the person affected by its output are not the same. Chapter 2 names four other modes of AI use — generative work, conversation and reflection, building, augmentation — where this method is not needed. Everything from Chapter 3 onwards is aimed at the deployed case.

The method can produce six different answers for any engagement, each named and defensible, each reached at a specific point in the work. Two of the six are refusals produced above the method, at the Tier-0 checks. The other four are placements produced inside the method, at G3 Route, when a piece from the decomposition is matched to the substrate that should carry it.

• Off-limits to AI. Some problems arrive as decisions that should not be handed to AI at all, regardless of capability. Welfare-eligibility scoring, immigration-risk assessment, predictive policing, automated sentencing. For categories like these, the honest recommendation is to keep AI out, now and later — however tempting the capability curve gets. The method makes that recommendation available, not hidden.
• Not a tech problem. Some problems look technical but are not. Bad process, unclear ownership, a policy question dressed as a tooling question. The honest fix is policy, process, or people — not a tool. The method makes that redirection available as a named answer, not an embarrassment.
• Human-operated. Some problems pass both checks, but when the routing work is done, no automation earns its place against a person doing the work. The decision and the doing stay with a human. Plenty of engagements end here, and ending here is not a failure — it is a finding.
• Non-AI automation. Some problems are better carried by a rule, a script, or a constraint solver — a deterministic substrate that does not learn from data. The behaviour is repeatable, the logic is inspectable, the code is its own audit path. A large share of what arrives framed as an AI problem is actually one of these in disguise. (Classical machine learning — gradient-boosted trees, logistic regression, and so on — is AI, not Non-AI automation; it sits above the Non-AI / AI boundary and produces AI as assistant or Autonomous AI routes.)
• AI as assistant. Some problems pass both checks but, when routed, no AI level can be made responsible enough for the specific decision. AI stays in support — retrieval, summarisation, option generation — and the decision and its consequences belong to a person.
• Autonomous AI. The rest are routable to AI that makes the call under controls built for it. The method says what has to be true of the specific placement — level, controls design, rollout — for the machine's decision to be trustworthy enough to delegate. That is where the rest of the book earns its space.

Six answers, produced at the places in the method where they are honestly answerable. No pre-commitment to any one of them. Four of the six are non-AI outcomes — a reflection of what honest routing work actually finds.

See Illustration 1.1.

Illustration 1.1 — Six answers the method can produce for any given problem. The top row holds the two Tier-0 refusals, produced above the method from the brief alone. The bottom row holds the four G3 placements, produced inside the method once the problem has been observed and decomposed. Four of the six outcomes are non-AI; the method does not privilege any one of them over the others.

The second thing the book offers: a shape that ages in years, not months. The method sits above specific models and specific providers. When the leading names change — and they will, probably every eighteen months — the shape of the decision does not. The tool at a given place in the system may update; the decision about which place to use, for which piece, does not.

The third: two reading modes. First pass, cover to cover, in about half a day — enough to see the whole shape. Tenth pass, one stage at a time, ten minutes before a specific decision meeting — enough to refresh the one question that matters this afternoon. Both uses are intended. The book is structured to serve both without penalty.

1.3 What the book is built from

A curated collection of external sources, pulled into one method by the author.

Most of the method's pieces are not new. The refusal to treat a social problem as a technical one comes from Broussard [1]. The list of decision categories that should stay off-limits to AI comes from McQuillan [2]. The older claim that some decisions depend on a human bearing their outcome traces to Weizenbaum's 1976 argument [3], which the book keeps in part and modifies in part — explained where it matters, not summarised here. MECE, Five Whys, Ishikawa, Jobs-to-be-Done, the pyramid principle — each is a familiar piece of management consulting, credited to its actual origin where the popular attribution is wrong. The governance spine borrows from NIST and ISO. The interaction-design stack borrows from Amershi, PAIR, and Shneiderman. The book's contribution is not new pieces; it is the selection, the ordering, and the decisions about where each piece fits in a single, runnable method.

The author's own recommendations are distinguished from the sources'. Where this book says do X because of Y, the reader can follow Y back to the source that argues for it. Where the book says don't do X because it usually fails at step Z, the reader can check whether the claim is a widely held finding or the author's editorial judgement. Both are marked.

Every cited source carries a provenance tag in the reference list at the end of each chapter — verified, partial, secondary, vendor-origin, pre-canonical, single-author. The tag says how much weight the claim earns. Where an idea is commonly misattributed, the reference list says so. Small repairs, made visible.

A method that cannot be wrong about a specific engagement is not a method; it is a style. The one in this book can be wrong, at named gates, in named ways. The chapters that follow show where.

Next: the five modes of AI use, and which one this book is about.

Sources

[1] Broussard M. Artificial Unintelligence: How Computers Misunderstand the World. MIT Press; 2018. [verified]

[2] McQuillan D. Resisting AI: An Anti-fascist Approach to Artificial Intelligence. Bristol University Press; 2022. [verified]

[3] Weizenbaum J. Computer Power and Human Reason: From Judgment to Calculation. W. H. Freeman; 1976. [verified]