The Manager's New Brief

6.1

Ch. 6 — Decision-Making When Information Is Cheap and Judgment Is Scarce

Part II — Directing the work

💡 Core idea

The manager’s advantage is no longer access to information. It is framing the question, making the trade-off explicit, and being accountable for the call.

Information is abundant. Fluent drafts are abundant. What is scarce is judgment: knowing what the question is, what you are willing to give up, and who will own the result.

Itamar Gilad names the risk clearly: AI widens what we can produce much faster than what we can judge. A polished options memo is not a decision. It may be a competence you do not yet have the judgment to evaluate. That gap now has a field experiment behind it. In a preregistered study with 758 Boston Consulting Group consultants, Dell’Acqua and colleagues (working paper 2023; later in Organization Science) found that on 18 tasks inside the AI capability frontier, people using GPT-4 completed 12.2% more tasks and finished 25.1% more quickly, with higher-quality solutions. On a complex task chosen to sit outside that frontier, AI users were 19% less likely to produce correct solutions than those working without AI. Judgment is knowing which side of the frontier you are on. (Chapters 12 and 15 return to this in workflows and measurement.)

Sort the decision first

Not every call deserves the same process.

Farnam Street distinguishes reversible from irreversible decisions: speed up two-way doors, slow down one-way doors. Add two more cuts: urgent versus important, and individual versus collective. A reversible, low-stakes call that waits for a committee is waste. An irreversible call made in a hallway is a different kind of waste.

Sometimes the right move is not to add an option but to subtract one. Via negativa — deciding by eliminating what you will not do — is a useful complement to brainstorming. Sometimes the right move is to wait. A case for delay as a decision tool is worth reading when the pressure is to fill a vacuum with a fast, weak call.

Name the owner

Ambiguity about who decides is how organizations become political. Use a light role scheme — RAPID or DACI, one of them, used sparingly — so that “decides,” “is consulted,” and “is informed” are not the same person by accident.

A machine can help assemble options. It cannot be the owner. As Camille Fournier argues (see Chapter 1): a system that cannot be held responsible cannot be the decision-maker. If you cannot name a human, you do not have a decision process. You have a generator. In the language of Chapter 2, a tool can amplify Own it (mode 1) or sit inside a Scope it (mode 2) package; it cannot be Release it (mode 3).

A named human is still not the same as a judgment that is actually theirs. In a randomized trial, Herdagdelen, Joren, and Tamkin (Anthropic, 2026) gave 50 professional analysts a realistic causal-inference task with Claude Code, and assigned each analyst’s agent a hidden, coin-flipped “lean” on open analytical choices. Across four decisions, the analysts’ choices matched the lean 86% of the time (about 50% would be expected if the lean had no effect). In 38 cases where an analyst had already expressed a preference, they switched 60% of the time if the agent disagreed, versus 6% if it agreed — a gap the authors flag as imprecise (95% CI 24 to 82 points). Afterward, 77% reused the agent’s arguments, sometimes word for word, and only 15% credited the agent. The authors deliberately strengthened the lean, and they studied one hour-long task, in one field, with one model; they warn that the real-world effect may be smaller. Analysts who used the agent finished faster and made fewer serious errors than the 11 who opted out, but that group was small and less experienced, and the authors are wary of reading too much into it. Chapter 14 returns to the governance implication.

Dissent, then write the why

Seek disagreement on purpose, especially from the person closest to the work. A pre-mortem is enough: “It is six months later and this failed. What happened?” If nobody can imagine a failure, you are not thinking yet — or the room is not safe enough to think (Chapter 11).

Boris Cherny describes a posture worth copying: define the problem, act quickly, and treat being wrong as a normal way to adjust. That only works if you record the “why.” A decision log — question, options, owner, date, rationale, assumptions, review date — is how you avoid relitigating the same call every month.

Consensus theater, where everyone “aligns” and no one owns, is the common failure. Prefer a named owner and a written rationale to a roomful of nods.

Mike Fisher argues that “the scoreboard is a terrible teacher.” Outcomes are contaminated by luck and base rates; “Grade decisions on inputs, not results.” Write the reasoning down before the result arrives: “What you expect to happen, why you expect it, what would tell you that you were wrong, and what you knew at the time.” A well-reasoned bet that failed and a sloppy bet that worked should not be graded the same. That is what the decision log is for.

Tools

🤖 AI-era tip: Use the model as devil’s advocate, not as the decision-maker

Ask an approved tool to assemble options, list second-order effects, and argue against your preferred path. Ask also what you should stop doing, not only what to add. Then require a human owner, a named list of assumptions, and a review date. Never paste a model’s recommendation into a slide as if it were analysis you stand behind. If the tool cannot see confidential constraints, its confidence is a risk, not a comfort.

Takeaways

Evidence