PDF Summary:The AI-Driven Leader, by Geoff Woods

Many leaders already use AI. They’re just using it for the wrong things—drafting emails, summarizing documents, and answering questions that don’t require much judgment. In The AI-Driven Leader, leadership strategist Geoff Woods argues that AI’s real potential is in strategic thinking: the decisions and analysis that determine whether your organization thrives or falls behind. He offers a practical framework for closing the gap between how most leaders use AI and how the best ones do.

In this guide, we’ll explore Woods’s method for getting the most out of AI as a leadership tool. We’ll also examine what AI is really doing when it appears to reason, we’ll consider how human cognitive bias and AI design can sometimes amplify each other, and we’ll bring in research that strengthens and complicates Woods’s argument.

(continued)...

Seeing the Future Isn’t the Hard Part

The Great Mental Models Volume 1 frames long-term strategy as second-order thinking—the practice of tracing both where a decision leads and where those effects lead in turn. Research suggests that the obstacle to this in organizations is mostly structural: Corporate incentives make protecting short-term results the more defensible choice, as the consequences of disappointing earnings expectations are typically sharper and faster than the opportunity cost of forgone long-term value.

There’s also a second complication on the AI side. Frontier LLMs show some forecasting advantages over humans in certain structured tasks, but that advantage rests on pattern recognition across large bodies of data. Decisions that have no close precedent in the model’s history are the cases where long-horizon thinking matters most—and the ones where AI’s scenario-tracing is least reliable.

The solution, like the problem, may be institutional: Organizations must create space and incentives for people to act on what long-horizon analysis reveals, not just understand it. That distinction matters for Woods’s claim that AI can help resolve the tension between short-term pressure and long-term strategy. If the obstacle were mostly analytical—if leaders simply couldn’t see the downstream consequences of their choices—better tools could plausibly close the gap.

Where Human Judgment Is Irreplaceable

Woods is careful to distinguish between AI as a powerful analytical tool and AI as an authoritative source of truth. He characterizes AI as a prediction machine, not an oracle: It generates responses by identifying patterns in the data it was trained on and predicting what should come next. That makes it remarkably capable at synthesizing information, generating options, and structuring thinking, but it also means that AI can be confidently wrong.

When AI fills gaps in its knowledge with plausible-sounding but inaccurate information—a phenomenon called hallucination—it does so without any indication of uncertainty. Leaders must always use their own judgment to validate, contextualize, and apply AI output. This is why the Thought Leader role isn’t optional. AI expands the range of material your judgment can work with. It can’t supply the judgment itself.

What the Prediction Machine Can’t Predict

Other experts agree with Woods that AI can be understood as a prediction machine. What an LLM is actually doing is generating text that fits: constructing the response that sounds most like a knowledgeable answer, based on which words are statistically most likely to come next, without any reference to whether those words are true. That isn’t the same as retrieving a fact or checking a claim, and nothing in the system tells the model that it’s wrong. Linguist Emily Bender and colleagues coined the term “stochastic parrot” for this, “stochastic” meaning “determined by probability.” AI output looks equally confident whether the model is on solid ground or reaching past what it reliably knows—it won’t flag its own uncertainty.

This is why hallucinations look indistinguishable from correct answers: A model that invents a “fact” out of thin air and a model that accurately recalls one are, at a mechanical level, doing the same thing. The only difference is whether the topic had enough coverage in the AI’s training data for the model to land on a real answer. Hallucinations, therefore, don’t occur randomly, but cluster where the model’s statistical confidence in the correct answer is weakest: in poorly documented domains of knowledge, where the LLM’s pull toward writing fluent text can overpower a fragile factual signal. The upshot is that you should apply more scrutiny to factual claims in niche areas.

Use AI as Your Strategic Thought Partner

Understanding AI’s strategic potential and knowing how to unlock it are two different things. Woods argues that the gap between them is almost always a communication problem. This section examines the four-part framework he recommends for structuring your prompts, how that framework works in practice, and the three roles you can ask AI to play to help you in your most important leadership work.

What AI Needs from You

Woods contends that the quality of an AI tool’s output is determined almost entirely by the quality of your input. Most people approach AI like a search engine—typing a question and waiting for an answer—and get results that feel generic or shallow. He argues that this isn’t a limitation of the technology, but it illustrates a mismatch between what an AI model is designed to do and how most people ask it to do things. An AI doesn’t perform well when it lacks context, has no sense of what role it should be playing, or is asked to jump straight to a conclusion without enough information to reason from. To address this, Woods introduces a four-part framework he calls CRIT: Context, Role, Interview, and Task.

Context is where you set the scene. Rather than posing a bare question, you give the AI a detailed picture of your situation—the background, the constraints, what you’ve already tried, and what’s at stake in the decision you need to make. The more specific the context you provide the model, the more relevant its response will be.

Why Context Quality Beats Context Volume

Context quality determines output quality because, as stated earlier, an LLM calculates the most probable continuation of whatever text you’ve given it. When you set the context for the model, you establish a starting point for a long chain of probability calculations. When your prompt is a bare question, as Woods warns against, the range of text that could plausibly follow it is enormous. When, instead, your prompt specifies who you are and what you’ve already tried, the range of plausible continuations collapses toward outputs that fit your situation more closely.

However, it’s not accurate to say that more context is always better: Studies have documented sharp drops in performance and accuracy, called “context rot,” when irrelevant information is added to prompts. The model treats the totality of what you’ve provided as potentially meaningful input and has no built-in way to set aside what doesn’t apply. As the volume of the context you provide grows, relevant details have to compete harder for the model’s limited attention, and the signal gets diluted by everything surrounding it.The useful question to ask about any piece of context you’re considering adding is: Does this change what the model is likely to produce? If not, you’re adding noise rather than signal.

Good context is specific: It pulls the model’s response away from the broad, generically plausible middle toward your specific situation, your constraints, and your definition of what a useful answer would accomplish. Quality context also tells the model what not to produce. Telling the AI what you’ve already tried, what won’t work, what you’re not asking for is just as important as being clear about what you’re trying to achieve. Researchers who study prompting find that context and constraints work together. Prompts that do only one of these jobs—or neither—account for most of the frustrating, generic output that makes people feel like they’re talking past the tool rather than with it.

Role is where you assign the AI a relevant expertise. Because AI models are trained on vast volumes of text across many fields, you can direct one to respond as a market analyst, a financial strategist, an executive coach, or any other expert whose perspective would be valuable as you consider the decision at hand. This shapes not just the vocabulary of the response but the angle from which it approaches your problem.

(Shortform note: When you assign a role, you’re changing where the model starts its reasoning: which questions it leads with, whose concerns it treats as central, and how it carves the problem up. These patterns are learned from the enormous volume of professional writing in the AI’s training data, and for common, well-represented roles, specifying a starting point is useful. The problem is that orientation and expertise aren’t the same thing: Assigning a role in a prompt produces no improvement in factual accuracy compared to prompts with no role. This gap matters most in niche or technical domains, where the nuances that separate real expertise from a well-oriented answer are harder to learn from surface patterns.)

Interview is the element Woods considers most important—and the most underused. Rather than asking the AI to jump straight to drawing conclusions, he recommends that you ask it to interview you first, posing one question at a time until it’s gathered enough information to be useful. This step slows the process down in exactly the right way: The questions the AI asks often highlight assumptions you’ve made or gaps in your information that you hadn’t noticed. Your answers give the AI the material it needs to reason well, rather than respond generically.

What Answering Does for the Interviewee

Woods’s interview step solves two problems at once. The first is that the AI model lacks context about your problem, and questioning you helps it gather enough information to be useful. But second, when a model asks you a focused question about your task, the person who gains the most from your answer is often you. That’s because the cognitive work of constructing an explanation actually improves your understanding. This is why The Pragmatic Programmer (1999) recommends a practice called “rubber duck debugging,” which involves explaining a problem to an object that won’t answer—a rubber duck, a coffee mug, an empty chair—and typically discovering a new insight that helps you solve the problem in the process.

This effect has been tested in AI research. One study compared what the researchers called “Thinking Assistants”—LLMs that ask reflection questions before offering advice—against AIs that only answer, ask, or do neither. The ask-first-then-advise design significantly outperformed the others at helping users reach better decisions. The one-at-a-time constraint also turns out to matter: Training a model to use sequential iterative questioning—where each answer reshapes what the next question addresses—helps the model produce more useful responses. That said, you have to ask the AI model to interview you because it hasn’t been trained to do that automatically.

Task is where you specify the output you want: You might ask the AI to produce a strategic analysis, a list of alternatives, a risk assessment, or the draft of an important piece of communication. Woods notes that being precise in how you describe the output you want keeps the AI focused on what you actually need.

(Shortform note: There are two reasons to ask for an output format that preserves your role as evaluator. First, the format you request determines whether you expect to judge the output: Ask for a recommendation, and your brain is ready to accept whatever the AI delivers; ask for a list of options, and you know you need to choose among them. Second, format instructions behave more like suggestions than commands. Ask for a bulleted list and you’ll usually get one, but ask for a list with five items, each no longer than a sentence, and the AI will miss at least one of those requirements. The more specific the instructions, the more things can go wrong. For both reasons, it makes sense to treat the output as a working draft, not a finished product.)

How CRIT Works in Action

Woods’s CRIT framework for putting AI to work on strategy becomes clearest in action. Let’s consider an example: An orchard owner has five newly available acres and has to decide which apple varieties to plant—a decision that will shape her revenue for the next decade. She currently grows primarily Honeycrisp apples and sells mostly to regional grocery chains, with a smaller share of direct-to-consumer sales at a farmers’ market. So, using CRIT, she gives the AI context about her operation and customers, assigns it the role of an agricultural market analyst with expertise in specialty produce, asks it to interview her, and sets the task: Develop a planting strategy that accounts for her existing contracts while positioning her for growth.

When the interview begins, the AI model’s first question might be something like, “What proportion of your revenue comes from grocery contracts versus direct sales?” “Roughly 80-20,” she answers. Second question: “What do your farmers’ market customers ask for that you can’t currently provide?” She names heirloom varieties and cider apples—things they can’t find at the supermarket. Third question: “How do the margins compare between the two channels?” That question surprises the orchard owner. When she works through the numbers, her direct sales generate roughly three times the margin per pound that her grocery contracts do.

This means that the business that represents 80% of her revenue is, pound for pound, less valuable than the smaller, more direct relationship she has with customers who want something she doesn’t yet grow. When she sat down to talk through her strategy with the AI tool, she started with a question about which varieties to plant. But because she gave the model context about her business, the interview reframed that initial ask into a question about whether the five acres she’s about to plant might give her an opportunity to rethink her customer mix entirely.

AI Can Find What’s Hidden in Plain Sight—Not What’s Missing Entirely

What the Interview does well is find connections that are already implicit in the context you’ve provided: The AI asked the orchard owner about margin comparison because she mentioned two revenue channels. But providing rich, specific context shifts an AI’s advice by less than you’d think: Models draw mostly on patterns prominent in their training data rather than the particulars of the situation at hand. The result is that in situations like the orchard’s, they can easily produce what experts call “trendslop”: advice that reads as tailored but is mostly determined by the statistical weight of their training data. The margin question felt like a discovery because it reflects a common framework, applied to a case where it happened to fit.

There’s also a structural limit that no prompting skill can overcome. Machine learning systems are designed to handle quantifiable uncertainty—the kind where you know what the possible outcomes are and can assign probabilities to them. What they can’t handle are situations where you don’t yet know what the possibilities even are, let alone how to weigh them. The practical upshot is that the interview can only make explicit what’s already implicit—and that’s useful, as long as you don’t mistake it for something more.

Three Jobs Worth Giving to AI

The orchard example illustrates how AI performs in what Woods calls the Interviewer role: drawing out your thinking through focused questions, rather than generating answers from scratch. He considers this the most transformative of AI’s roles precisely because it shifts the dynamic from information retrieval to collaborative reasoning.

He identifies two other roles that you should ask your AI tools to play. As a Communicator, AI takes complex or loosely organized thinking and shapes it into clear messaging tailored to the audience you’ve specified—useful when you need to translate a nuanced strategic position into something a board, a team, or a customer can act on.

As a Challenger, AI stress-tests your thinking by generating counterarguments, identifying weaknesses in your plan, and simulating how a skeptical stakeholder might respond—a close approximation to a trusted advisor with no stake in telling you what you want to hear.

Woods emphasizes that getting AI to perform effectively in any of these three roles depends on the same foundation: giving the tool a well-structured prompt that gives it enough context to reason from, a role that focuses its expertise, and a task specific enough to produce something useful. He explains that CRIT isn’t just a technique for one kind of conversation: It’s the underlying architecture for all of them.

Are You Giving AI a Well-Posed Problem?

Woods presents CRIT as a communication technique—a better way to talk to AI—but it’s more precisely a task-design discipline. Computer scientists distinguish between tasks that are well-posed for a given tool and tasks that aren’t. For LLMs, that distinction matters more than almost anything else about how you phrase a question, since (as stated earlier), they perform most reliably when the task is clearly bounded, the context is rich, and the kind of output being requested has strong precedents in their training data.

Each element of CRIT does something specific to satisfy these conditions, and the three roles in Woods’s framework have correspondingly different profiles against this standard. The Communicator role, which asks the model to translate complex thinking into messaging for a defined audience, is naturally well-posed: Language transformation for a specified audience is exactly where LLMs are most competent. The Interviewer role is well-posed for a different reason: Rather than requiring the user to front-load all the context before knowing what’s relevant, it lets the model use its own pattern-matching to identify what information is missing through questions it’s well-placed to generate.

The Challenger role requires the most care. A model trained on human text is, by definition, drawing on a record of what’s already been said and argued—which makes it better equipped to recombine existing positions than to depart from them. A Challenger prompt that specifies which assumptions to interrogate and how to do so is asking for something the model can deliver, but a vague prompt will just produce the standard objections typically raised against what you’re proposing.

Build an AI-Driven Organization

Getting results from AI yourself and getting results across a whole organization are two different challenges. Woods argues that the second requires rethinking not just which tools your team uses, but how you structure roles, lead changes, and keep making progress. This section examines each in turn.

Redesign the Work Before You Introduce the Tool

Woods’s starting point for building an AI-driven organization is to find clarity about what each human role is actually for. He argues that most job descriptions function as comprehensive lists of responsibilities, when the more useful question is: Which 20% of a given role drives 80% of its outcomes? He recommends that leaders explicitly identify this critical 20% and make it the organizing principle for how work gets assigned and evaluated. The remaining 80%—the lower-leverage tasks that fill most people’s days—can then become the target for simplification and, eventually, automation. The sequence matters: Clarity about what each role is for should precede any decisions about what to streamline.

(Shortform note: The 80/20 audit Woods recommends is a question about roles: What does this job need to produce? In Multipliers, Liz Wiseman argues for asking another question alongside it: What is the specific person in each role uniquely good at? Wiseman finds that the leaders who get the most from their teams aren’t the ones who manage job descriptions carefully, but those who search for what she calls each person’s “native genius”—a skill they perform so naturally that they don’t recognize it as a skill at all. Applying both Woods’s and Wiseman’s criteria together sharpens the question of what to simplify: Target those tasks that are neither among the role’s critical outputs nor connected to your worker’s native genius.)

To simplify that 80%, Woods adapts a five-step process improvement sequence he associates with Elon Musk’s engineering work at Tesla and SpaceX. The first step is to question whether each task or process needs to exist at all—a harder question than it sounds, since many organizations continue doing things simply because they always have. The second is to delete as much of the remaining process as possible. The third is to simplify what survives. The fourth is to accelerate it. Only then—fifth—should you consider automating. Woods emphasizes that automating before completing the earlier steps just produces bad results more efficiently.

Where the Five Steps Came From

The sequence Woods describes didn’t originate with Musk, but is a version of the Lean manufacturing philosophy, originally developed by Toyota, that informs process improvement thinking in many industries. The Toyota Production System rests on the idea that waste has to be identified and eliminated before you can optimize anything, and automation belongs at the end of the process. What distinguishes Lean manufacturing from Musk’s method isn’t just the sequence, but the care with which each step is applied—and the assumption that the people closest to the work are partners in getting it right. The two approaches are worth comparing.

The first step, to question every requirement, is an invitation to challenge bureaucratic dead weight. But Musk’s version of this step instructs leaders never to accept a legal or safety requirement, treating regulatory standards as presumptively dumb starting points rather than as floors beneath which deletion is off limits. Tesla’s Autopilot system, developed under that philosophy, prompted a federal investigation and the largest recall in Tesla history. Aggressive deletion, step two, calls for removing what has no purpose, expecting to restore some of it. But at Twitter, Musk cut 80% of the workforce and dismantled safety infrastructure before determining whether any of it was essential or irreplaceable.

Steps three and four—simplification and acceleration—are where Musk’s approach and the traditional Lean manufacturing philosophy most closely converge: SpaceX’s engineering record suggests Musk applies these with genuine discipline. Step five, finally, is where Musk himself supplies the most useful cautionary note: In building the Tesla Model 3, he acknowledged violating his own sequence by automating before deleting and simplifying, which resulted in what he called “production hell.” That admission contains the sequence’s central argument: that automating a flawed process only makes it fail faster.

Once lower-leverage work has been cleared away, the challenge is ensuring that employees use the freed time for strategic work rather than simply refilling it with more tasks. Woods argues that this requires setting the expectation that employees contribute their own analysis and propose solutions rather than simply identifying problems. He identifies three practices that encourage this shift. First, respond to questions with questions—push employees to reason through problems themselves before you offer your answer. Second, when you do offer guidance, explain your reasoning so that employees develop judgment rather than just compliance. Third, hold people to well-defined standards rather than loose expectations.

Ancient Practice, Modern Evidence

The logic underlying Woods’s practices for ensuring employees use time strategically is older than business literature—by about 2,500 years. Answering questions with questions, explaining your reasoning, and holding people to standards all constitute the Socratic method. Socrates believed that questioning generates understanding more reliably than informing someone of facts does, because a person must construct an answer and, in doing so, discovers where their thinking is incomplete. Research supports this: Questioning-based instruction develops not just comprehension but awareness of one’s own reasoning in real time, with documented improvements in critical thinking and problem-solving.

Each of the other practices Woods recommends maps onto its own documented phenomenon, too. Research on learned helplessness shows that when employees are given the solutions to problems, they stop generating their own, with measurable declines in work involvement as a result. Additionally, high expectations produce better performance because leaders who believe in someone’s ability change how they behave toward that person. Finally, research on the protégé effect (the finding that explaining something to another person demands more than just knowing it) suggests that when a leader explains their reasoning, they’re not only developing an employee’s thinking; they’re also developing their own.

Lead the Change—and Keep It Moving

Woods contends that AI adoption fails most often when it’s treated as a technology initiative rather than a leadership one. Technology serves strategy rather than substituting for it—which means leaders need to establish what they’re trying to achieve before selecting the tools they’ll use to get there. When building internal support for an AI initiative, he recommends you begin with a specific problem worth solving, identify a use case likely to produce visible results quickly, and then involve others in designing the solution, rather than presenting them with a finished proposal. He notes that people tend to champion what they helped build.

(Shortform note: When people participate in building a solution, they make a series of small, active investments in it, and those investments quietly reshape how they see themselves. Robert Cialdini explains in Influence that people who make even minor public commitments to a position bring their subsequent beliefs and behavior in line with it, not primarily because of social pressure, but because their contribution changes their self-conception. They become, in their own minds, someone who supports this thing. In Switch, Chip and Dan Heath add that the most effective way to build investment in an organizational change isn’t to argue for it, but to create an experience that helps people feel emotional ownership of it.)

Execution is where most strategies fail. Woods draws an example from the race to the South Pole in 1911: Roald Amundsen’s team covered 20 miles each day regardless of conditions, while Robert Falcon Scott’s team sprinted in good weather and halted when it was harsh. In the end, Amundsen arrived at the pole a month ahead of Scott. For leaders, the lesson is to protect time for strategic priorities even when daily demands tempt you to do otherwise.

Woods recommends translating your plan into 30-day milestones, blocking time to work on them, and using regular one-on-ones with direct reports as coaching conversations. AI can help with each of these—from breaking a plan down into near-term milestones to ideating questions to keep one-on-ones focused on growth.

The Other Half of Amundsen’s Lesson

In Great by Choice, Jim Collins and Morten T. Hansen use the same expedition to draw an additional lesson. They note that the most disciplined organizations set two kinds of targets: a floor (the minimum they’ll achieve even in difficult conditions) and a ceiling (the maximum they’ll allow themselves to do even in favorable conditions). Amundsen’s team capped their daily march at 20 miles even on days they could have covered 25. That restraint preserved the reserves they’d need when conditions turned harsh. Similarly, Collins and Hansen found that companies that push as hard as possible during favorable periods end up too depleted to absorb setbacks when they happen.

The practical implication for Woods’ milestone and calendar-blocking advice runs in both directions: A blocked time slot has an end time as well as a start time, and both are commitments. Stopping when your two hours of strategic work are up—even when the thinking is going well, and even when momentum is on your side—is the discipline that keeps an AI initiative alive across months rather than burning bright for a week and then collapsing under the weight of competing demands.

For leaders who aren’t sure where to begin, Woods proposes a single reframe as the starting point: Replace the question “How do I do this?” with “How might an AI tool help me do this?” This begins a self-reinforcing cycle in which asking the question creates awareness of possible use cases, acting on those cases improves your skill at creating AI prompts, and better prompting produces better results that make the question worth asking for the next problem you face.

The Habit Behind the Habit

Woods presents his reframe question as the starting point of a self-reinforcing cycle. The behavioral science behind this explains both why it works and where it can stall. What Woods describes is essentially what habit researchers call an implementation intention: a decision made in advance about what you’ll do when a specific situation arises—namely, to ask the question “How can AI help with this?”

Research detailed in James Clear’s Atomic Habits shows that tying a new behavior to a reliable situational trigger—“when X happens, I’ll do Y”—makes you more likely to follow through than a resolution to act differently. The reframe question functions as that trigger: It activates automatically when you face a task that you’re not sure how to approach, removing the need to re-decide in the moment whether AI should be one of the options you consider. What the cycle requires to actually become self-reinforcing, however, is a satisfying reward at the end of it; without one, the behavior doesn’t repeat. Applied here, that only happens if the results you get by asking whether an AI tool can help are useful.