Situational Leadership for AI: More Like a Capable Colleague than a Fancy Formula

Treating LLMs like vending machines guarantees mediocre results—try situational leadership for better results

---

You wouldn’t hand a new hire a laptop, point vaguely at the codebase, and say “make it better.” You wouldn’t expect a brilliant data scientist to magically produce great marketing copy. You wouldn’t give the same instructions to a seasoned architect and a junior developer and expect identical results.

And yet—this is exactly how many (most?) organizations interact with large language models.

The executives who’ve spent decades learning to match tasks to talent, provide appropriate context, and adapt their leadership style to individual strengths suddenly forget everything when the “employee” runs on GPUs instead of caffeine. They pump generic prompts into their LLMs like quarters into a vending machine, then complain when the results disappoint.

Here’s the uncomfortable truth: the skills that make you effective at managing capable knowledge workers are precisely the skills that make you effective at getting value from LLMs. And if you’re bad at one, you’re probably bad at the other.

The Situational Leadership Connection

In the 1970s, Paul Hersey and Ken Blanchard developed Situational Leadership Theory—the radical notion that effective leaders (or parents) adapt their style to the person and task at hand. A directive approach works for someone new to a task. A supportive approach works for someone who knows what to do but lacks confidence. Delegation works for someone competent and committed.

The model seems obvious in retrospect. Yet organizations still promote technical experts into management roles and watch them fail by treating every report identically—usually by treating everyone the way they themselves would want to be treated (we all see the world through our own lens of needs and desires and strengths and past experiences—the best managers come to understand the lens through which their reports are perceiving the world).

Identical treatment for all gets shoddy results with teams. It also gives shoddy results with AI.

Different LLMs have different capabilities, different training, different strengths, different weaknesses. Claude excels at nuanced reasoning and following complex instructions. GPT-4 has particular strengths in some business settings. Grok has some of the best information on trending cultural topics. Smaller models trade capability for speed. Specialized models dominate narrow domains. Treating them interchangeably—or worse, treating them like omniscient oracles—is the management equivalent of giving identical assignments to your entire team regardless of skill set.

The situational leader asks: What does this specific person need from me to succeed at this specific task?

The effective AI user asks: What does this specific model need from me to succeed at this specific task?

The parallel is not metaphorical. It’s operational.

Context: the Briefing You Forgot to Give

Picture a new employee’s first day. A good manager doesn’t just assign tasks—they provide context. Here’s what we’re building. Here’s why it matters. Here’s how your work connects to the larger mission. Here’s who does what and who can help with specific things as needed. Here are the constraints we’re operating under. Here are the decisions that have already been made and why. Who. What. When. Where. Why. How.

Bad managers skip this. They assume context is obvious. They issue instructions and get frustrated when results miss the mark. “I asked for X, why did you give me Y?” Often it’s because they never explained what X meant in the organization, with its constraints, for its customers.

LLMs suffer the same failure mode, magnified.

When you prompt a model with “write me a marketing email,” you’ve given it essentially no context. What product? What audience? What tone? What’s the goal—awareness, conversion, retention? What’s worked before? What constraints exist? The model will generate something plausible and generic, because generic was all you gave it to work with.

Good managers are explicit about context. Effective AI users are explicit about context. The quality of the output is directly proportional to the quality of the input—not the quantity of words, but the relevance of the information provided.

This extends beyond single prompts. LLMs have session context—what’s been discussed, what’s been decided, what’s been tried. Just as a good manager maintains continuity across conversations with their team (”remember, we decided last week that...”), effective AI users maintain continuity across interactions.

The model that helped you architect a system yesterday can help you implement it today—if you’ve preserved the context of those decisions. Wipe the slate clean and start over, and you’re back to onboarding a new hire who knows nothing about your project.

Note that the level of granularity (what level of detail/specificity do you provide?) also matters. Provide too much, and you get what you ask for, with less opportunity for the AI to surprise you with some helpful additions. Provide too little, and the AI may drift away from your needs, creating something tangential but not quite right.

Task-Model Fit Is Just Role Clarity by Another Name

Smart managers don’t assign accounting work to engineers or graphical work to people who think in spreadsheets. They match tasks to talents, projects to strengths, challenges to capabilities.

This may strike you as more-basic-than-needed management advice. It’s also more rare than one might think for AI usage.

Organizations often standardize on a single model for all use cases—the enterprise equivalent of hiring one type of person for every role. They use the same massive model to answer simple questions that a smaller model could handle faster and cheaper. They ask generalist models to perform specialist tasks. They wonder why results vary wildly.

Different models for different jobs:

- Quick, factual queries don’t need your most powerful (and expensive) model. A faster, smaller model often performs equivalently for routine tasks—just as you don’t need your most senior engineer to reset a password.

- Complex reasoning benefits from models trained to think step-by-step. Asking a model optimized for speed to navigate nuanced trade-offs is like asking your fastest coder to lead a design review. You’ll see this implemented in many models as a “thinking mode” of one sort or another.

- Domain-specific work often benefits from fine-tuned or specialized models. General practitioners are valuable; specialists exist for a reason. If you don’t have specialized models, ask your generalized model to adopt a persona that would be adept in the given domain, and you’ll find your tokens better spent -- the predictive model will provide answers weighted toward the desired domain, providing inherently better answers. Tune both the persona and the questions to the specific domain.

- Creative tasks may warrant different models than analytical tasks. The person who writes brilliant copy isn’t necessarily the person who debugs distributed systems.

The effective AI user develops model literacy the way effective managers develop people literacy. They learn what each model does well, where each struggles, and how to route work accordingly.

Strengths and Weaknesses Aren’t Bugs—They’re Features

Every employee has strengths to leverage and weaknesses to manage around. The brilliant architect who can’t stand presenting. The charismatic salesperson who crumbles under detail work. The reliable executor who struggles with ambiguity.

Good managers realize that people are not uniformly excellent. They design teams and workflows that amplify strengths and compensate for weaknesses.

LLMs are similar, but many organizations keep pretending they aren’t.

Current-generation models tend to struggle with:

• Precise numerical reasoning (your accountant should check the math)
• Real-time information (they’re trained on historical data)
• Guaranteed factual accuracy (they can hallucinate convincingly)
• Following extremely long context with equal attention throughout
• Knowing what they don’t know

Current-generation models tend to excel at:

• Pattern recognition across large bodies of text
• Generating plausible first drafts quickly
• Explaining complex concepts at various levels
• Synthesizing information from multiple sources
• Brainstorming and ideation
• Reformatting and restructuring content

The effective approach isn’t to wish away weaknesses—it’s to design workflows that play to strengths. Use the model for the first draft; have a human verify the facts. Use the model for synthesis; have a domain expert validate the conclusions. Use the model for ideation; have the team evaluate feasibility.

This is management, not magic.

The Feedback Loop That Everyone Ignores

Managing knowledge workers isn’t a one-way broadcast. It’s a dialogue. You give direction. They execute. You provide feedback. They adjust. You refine requirements and improve processes. They deliver again, and may provide feedback of their own. Good managers create tight feedback loops that quickly converge on the right outcome.

Most AI users treat prompts as one-shot commands, then complain when the first response isn’t perfect.

Would you judge an employee’s capability based on their first draft of their first project on their first day? Then why judge a model based on a single cold-start response?

Effective AI interaction is iterative. The first response is diagnostic—what did the model understand, what did it miss, where did it go wrong? The second prompt corrects course. The third refines. By the fourth or fifth exchange, you’ve co-created something neither you nor the model would have produced alone.

This is how you work with smart humans. Also how you work with smart machines. Note that this is a reason that some recommend less detail up front for some tasks — the AI will generate more concepts for consideration that can expand your idea when it is allowed more wiggle-room. Different model strengths will affect this — ChatGPT has gotten better and better at providing exactly what you ask for, while Claude has gotten better and better at expanding on your inputs with additional ideas. Knowing this might cause you to choose one or the other for a specific task, or shape how you write your prompt.

The executive who says “I tried ChatGPT once and it gave me garbage” is the executive who says “I told my team what I wanted once and they didn’t deliver exactly what I envisioned.” Both statements reveal more about the executive than about the team or tool.

The Uncomfortable Implication

Here’s where this gets interesting: if AI effectiveness requires the same skills as people management, then organizations’ AI results will correlate with their management culture.

Companies are expecting young non-managers to become expert AI users without helping them understand how to manage. I believe this is the primary reason for the statistics showing much more productive use of AI by older users than by younger users...the inverse of historical technology usage patterns. Companies that are trimming hiring of entry-level folks to do more of the work with AI may have a hard time if they don’t create a path to management, with management training, for their smaller classes of entry-level employees. They will also need to build retention mechanisms to keep a higher percentage of entry-level folks longer...or they’ll wind up with hollowed-out organizations.

Companies with clear communication norms and management training will get clearer AI outputs. Organizations that provide context will get contextually appropriate responses. Teams practiced in iterative refinement will iterate effectively with AI. Leaders who match tasks to capabilities will match tasks to models.

And organizations with terrible management practices? They’ll get terrible AI results and blame the technology.

AI is a mirror. It reflects the quality of the instructions it receives, the clarity of the context provided, the thoughtfulness of the task assignment. If your AI outputs are mediocre, the first place to look is the inputs.

This isn’t to say AI tools are perfect—they aren’t. Models have real limitations, and those limitations matter. But the gap between mediocre and excellent AI usage often has more to do with the human side than the machine side.

What This Means for Leaders

If you’re a new manager struggling to get value from AI tools, you might actually be struggling with management fundamentals that AI is simply making visible. The fix isn’t better prompting tips—it’s developing the skills you’d need to manage capable knowledge workers effectively. If you’re using the AI to write code, a great management model to think about is that of the product owner. It’s a very special set of skills, as some might say. And NOT AT ALL the set of skills you have trained in as a traditional software engineer.

If you’re a senior leader wondering why AI investments aren’t paying off, look at your organization’s management culture. Do your managers provide context effectively? Operate at the right level of granularity? Match tasks to capabilities? Create feedback loops? Adapt their approach to different situations and individuals?

If not, AI tools will underperform—because they require the same inputs that humans require, just in different formats.

The organizations that will excel with AI are the ones that already excel at knowledge work management. They understand that capable resources—human or artificial—need direction, not just commands. Context, not just tasks. Iteration, not just deadlines.

The rest will keep pumping formulaic prompts into their systems and wondering why the results disappoint.