The 1% Book Shelf — Human Magic by Johan Roos (Routledge, 2026).
Most technology shifts changed how we work. This one changes who decides. For decades the deal was simple: the tool helped, you chose. AI quietly rewrites that contract. Instead of using a system to inform a decision, we increasingly invite a system to make it for us.
Johan Roos sits with that shift in his 2026 book, Human Magic: Leading with Wisdom in an Age of Algorithms. It isn’t a manual about prompts. It’s an argument for staying awake. Roos frames the choice with two paths:
| The Erosion Curve | The Amplification Curve |
|---|---|
| Use algorithms without thinking | Challenge yourself on purpose |
| Core skills fade — curiosity, creativity, critical thinking, communication, teamwork | The same skills sharpen |
| You manage what the machine hands you | You use the machine to strengthen your own thinking |
Slide down the first and you stop producing ideas — you just supervise the ones the model gives you. Climb the second and the tool makes you better. Here are the four ideas worth keeping.
1. The AI concierge vs the professional citizen
Roos opens with a senior executive who realised she had stopped leading and started acting as an AI concierge — no longer making the big calls, just managing and forwarding what the algorithms produced. I see this constantly. We type a prompt, get a polished answer, and accept it on the spot. Roos calls this the shallow questioning loop: take the first answer that sounds good, and you let the machine end your curiosity. You lose the useful discomfort of not knowing — the discomfort that actually leads to insight.
His fix is to move from concierge to professional citizen. A concierge is efficient. A citizen takes the time to think.
| AI concierge | Professional citizen | |
|---|---|---|
| Mode | Manages and passes on the output | Thinks the problem through first |
| Default | Accepts the first plausible answer | Frames the problem before fetching |
| Curiosity | Outsourced | Protected — focused and broad |
| Trade-off | Fast, but owns nothing | Slower, but owns the reasoning |
The rule is deliberately blunt:
Frame first, then fetch. Before you ask the machine, write down how you see the problem and your first ideas.
That protects both kinds of curiosity: the focused kind that solves a specific problem, and the broad kind that explores new territory. AI is good at the focused task and poor at open-ended wandering. Let the algorithm define the question and you forfeit your creative freedom — and the unexpected insights that come with it.
2. The oracle effect and the death of “because”
We tend to believe things because they sound confident. Roos likens our trust in large language models to King Croesus consulting the Oracle at Delphi: the AI produces a well-written answer, and we accept it as true because it’s convincing. He calls this algorithmic deference.
That trust quietly erodes critical thinking, because it strips out the warrant — the logical “because” that connects evidence to a conclusion. I’ve sat in boardrooms where people treat metrics and dashboards as explanations, confusing the number with the reason. Roos reaches for Stephen Toulmin’s model of argument as the antidote: when AI hands you advice, make the reasoning visible. Lay out the claim, the evidence, the support, and the likely objections. Use the model to test your thinking — ask it to find three flaws in your plan — but don’t let it do the reasoning for you.
3. The preparation–performance boundary
We forget that leadership and communication are physical, present-tense acts. Roos points to the 2023 Vanderbilt University incident: after a tragic shooting, administrators used ChatGPT to write a condolence email. By machine standards the message was perfect. People found it offensive. It missed what the Romans called eloquentia corporis — the eloquence of the body.
His rule here is sequential amplification. Use AI in the background to prepare, research and organise your argument. But when it’s time to speak — a meeting, a negotiation, a crisis — put the device away. Real communication runs on space, voice, body language and presence. Reach for AI mid-conversation and you break the flow and the trust between people: your voice loses energy, you stop reading the room, and you start serving the script instead of the audience.
4. Bionic collaboration vs automated consensus
To show what real teamwork looks like, Roos goes back to the building of Canterbury Cathedral in 1178. When the master mason William of Sens was paralysed in a fall, the work carried on — the team had a collective mind. They worked closely, pooled different skills, and solved problems through creative friction.
Compare that with today’s “human-outside-the-loop” teams. As we hand agents the small jobs — scheduling, note-taking on project calls, tracking actions in Slack or Microsoft Teams — we get what Roos calls social thinning. On one project I worked on, automated tools handled every calendar invite and meeting record. Logistics got easier; the informal check-ins and honest disagreements got rarer. We lost the closeness and trust you need to take a risk or push back.
His alternative is bionic collaboration: pair human insight with machine power. Let the algorithms carry the routine, and keep people on the things that matter — uncertainty, ethics and meaning.
The 1% takeaway: practical wisdom
If you keep one idea from this book, keep practical wisdom — what Aristotle called phronesis. Intelligence works out what can be done; wisdom cares about what should be done.
Intelligence asks, “What can be done?” Wisdom asks, “What should be done?”
Roos illustrates the gap with the “Dashboard King.” Executives used an AI hiring dashboard to speed up recruiting and predict how long employees would stay. It looked impressive, the numbers improved, and the team felt in control — while the algorithm was quietly baking in age discrimination. They tuned the data. They never stopped to ask whether the system was fair. (The real-world echo is Mobley v. Workday, the 2025 collective action alleging Workday’s AI screening tools disproportionately rejected applicants over 40.)
Over the next decade your value will rest less on technical craft — coding, writing, analysis — and more on your ability to see what matters, balance competing values, hold your nerve under pressure, and act with courage.
Here I’d push back on Roos a little. He treats the ethical and emotional domain as uniquely human, and he may understate how far AI can shape — or partly imitate — aspects of judgement. As these systems sit deeper inside decision processes, they increasingly guide, or even decide, which values get highlighted and which actions get considered. Roos argues that machines can’t feel the moral weight of a choice or care about people, and I share that view — but it overlooks how algorithms can quietly embed particular norms and biases into automated systems. So caring stays a human responsibility, while the line between human moral agency and algorithmic influence keeps blurring. That’s exactly why how we design and use these systems matters.
Practically, teams can guard against hidden bias and ethical drift with concrete habits: audit AI outputs for fairness on a regular cadence, set up cross-disciplinary review panels, and build feedback loops where people can flag questionable decisions. Document how the algorithm reaches its recommendations and compare outcomes against your stated values, so problems surface early. Make those checks and conversations routine, and you give people the standing to question and improve the technology they depend on.
Application to work, leadership and change
Theory is cheap; the test is whether it survives contact with a delivery team. Roos’s five habits of professional citizenship map cleanly onto how we build and ship software. Here’s the working version, with a quick scenario for each.
| Practice | What it guards | Try this Monday |
|---|---|---|
| Deliberate Pause Protocol | Independent framing | Before any AI suggestion, ask what it optimises, what it sacrifices, who pays |
| Contrarian Requirement | Critical thinking | Write the scenario where the model is wrong; keep a double diary |
| Human-First Protocol | Collective intelligence | 3 minutes of silent framing, then a 5-minute dissent window |
| AI Ethics Hearings | Moral reasoning under pressure | Debate a real case with assigned sides |
| Weekly Capability Check | Self-awareness | 15 minutes every Friday: where did we lean on AI? |
| Quarterly Wisdom Review | Purpose over metrics | Ask whether the work served the mission or just the number |
1. The Deliberate Pause Protocol
Don’t open a meeting with “Here’s what the AI found.” That anchors everyone to the machine’s framing of the problem. Before accepting any AI suggestion, pause and ask three questions: What does this optimise? What does it sacrifice? Who pays the price? A minute of reflection turns passive acceptance into active thinking.
To wire it into a sprint or planning meeting:
- Set the expectation at the start of the sprint — before discussing any AI-generated input, the team pauses to reflect.
- When AI output is presented, the facilitator calls the pause: one minute for everyone to sit with the three questions individually.
- The team then shares what’s being optimised, what might be sacrificed, and who might be affected.
- Only after that does the group work through the AI’s suggestion in detail, adjusting or seeking alternatives.
Scenario. In sprint planning, the team receives an AI-generated story prioritisation. Instead of pulling it straight into the backlog, the product owner pauses the group and asks whether the top-ranked stories serve customer needs that don’t quantify easily — relationship-building, long-term maintainability.
2. The Contrarian Requirement
Before acting on an AI forecast, have someone write a realistic scenario in which the algorithm is completely wrong. Use a double diary: one column for what the AI says, one for what you think and why. It surfaces the weak signals the model missed.
Scenario. An AI estimate says a key system migration will have zero downtime. One engineer sketches a plausible case where an obscure dependency gets missed — and the team adds a manual cross-check before proceeding.
3. The Human-First Protocol
Open any big decision meeting with three minutes of quiet, so people can describe the problem in their own words before any AI output appears. Then run a five-minute dissent window in which the team has to put forward two genuine alternatives.
Scenario. In a feature design workshop, the facilitator hands out index cards and asks each person to write their own read of the user’s problem before the AI’s suggestions go up. Different perspectives surface, and the discussion gets richer.
4. AI Ethics Hearings
Rehearse ethical conflict with real material — for instance, the 2025 Deloitte Australia case, where an AI-assisted report shipped with fabricated references. Have engineers, product managers and compliance staff argue different sides. Assign a neutral facilitator — a team lead, or someone from HR or compliance — to guide the discussion, keep it on track and make sure everyone is heard. Set it up as a judgement-free space, with ground rules for respect and confidentiality, and step in if it gets tense. Practise the hard calls in safety so you’re ready when they count.
Scenario. A mock hearing: the AI proposes a feature that could lift engagement but also risks user privacy. Roles are assigned to investigate and debate the trade-offs.
5. The Weekly Capability Check
Spend fifteen minutes every Friday reviewing how much you relied on AI. Make it a shared ritual inside the retro, not a solo audit. Ask, together: Did we generate our own questions this week? Did we set aside time to create without AI? Were we present in the key meetings, or distracted by screens? Group reflection catches shared blind spots and builds mutual accountability.
Scenario. A retro closes with a quick roundtable on which decisions were AI-assisted and which came from team discussion — keeping visible where human insight still leads.
6. The Quarterly Wisdom Review
Check whether your team’s skill is serving a real purpose or just moving a number. Are you working on what matters, or only on what’s easy to measure?
Scenario. A quarterly offsite looks past velocity charts and asks whether the work supported the company’s bigger mission — surfacing value-driven improvements, not just throughput.
Change happens when these challenges become part of the daily work, not a poster on the wall. Peter Drucker drew the line cleanly:
Efficiency is doing things right; effectiveness is doing the right things.
Algorithms are very good at efficiency. Effectiveness is still your job.
Sources
- Johan Roos, Human Magic: Leading with Wisdom in an Age of Algorithms (Routledge, 2026).
- Plato, Theaetetus — on perplexity and the roots of wonder.
- Daniel Berlyne, Conflict, Arousal, and Curiosity (1960) — specific (focused) vs diversive (broad) curiosity.
- Stephen Toulmin, The Uses of Argument (1958) — claims, grounds and the “warrant”.
- Marshall McLuhan, Understanding Media (1964) — the medium as the message.
- Kathleen Eisenhardt and Donald Sull, “Strategy as Simple Rules”, Harvard Business Review (2001).
- Chris Argyris and Donald Schön on double-loop learning and the reflective practitioner (1974, 1983).
- Peter Drucker, The Effective Executive (1967) — efficiency vs effectiveness.
- Vanderbilt University’s Peabody College ChatGPT condolence-email incident (2023) — CNN.
- Mobley v. Workday, Inc. (2025) — algorithmic bias in hiring — Civil Rights Litigation Clearinghouse.
- Deloitte Australia AI report with fabricated references (2025) — The Register.
Leave a Reply