I call it "coffee and Claude time." Most mornings, before anything else, I open a conversation and start working. Not searching. Not prompting for a quick answer. Working — the way you work with a sharp colleague who has read everything, forgets nothing within a session, and has no stake in telling you what you want to hear.
I have used Claude to build a production AI platform, design scoring architectures, write investor materials, and think through organizational decisions. I have also used it to pressure-test arguments I was about to make in front of a room and to understand concepts I had not encountered before. It is the most useful professional tool I have encountered in thirty years of practice. I say that with full awareness of how strong that claim sounds.
But this article is not an endorsement. It is a practitioner's account of what AI collaboration actually looks like when you take it seriously — including the limitations, the failure modes, and the discipline required to use it well. Those things matter too, and they are rarely discussed honestly.
What it looks like in practice
Most people who say they "use AI" mean they use it to generate a first draft, summarize a document, or answer a factual question. That is a thin use of a deep capability. It is roughly equivalent to using a master carpenter to hang a picture frame.
The sessions I find most valuable are the ones that look like this:
I want to build a scoring methodology that captures the emotional alignment between video content and advertising context. I have a taxonomy of 92 emotions, 189 contexts, and 28 media genres. The question is how to weight the interaction between them in a way that is defensible to a CMO and reflects behavioral science, not intuition. Where do we start?
The weighting question is downstream of a prior question you haven't settled yet: are you trying to maximize emotional resonance for a given ad, or minimize contextual risk for a given brand? Those are different optimization targets and they produce different weighting architectures. Which problem are you solving first?
That exchange is representative of what I value most: not answers, but the right question asked at the right moment. I had been thinking about the weighting problem for weeks. Claude identified in one response that I had not clearly separated two distinct optimization objectives. That clarity changed the direction of the architecture.
This is not magic. It is pattern recognition operating at scale — the ability to recognize that a question I am asking maps to a class of problems with a known structure, and to name that structure before I start solving. A good analytical colleague does the same thing. The difference is availability, patience, and the absence of any incentive to tell me what I want to hear.
The three things AI does well in professional practice
Structured thinking under pressure
When I am under time pressure and my thinking is not yet organized, Claude functions as a thinking scaffold. I can describe a problem in rough, incomplete terms and work toward clarity through the exchange. The process of explaining something to a capable interlocutor — even an artificial one — forces the kind of precision that thinking alone often does not.
This is not the AI solving the problem. It is the conversation creating the conditions for me to solve it. The distinction matters because the output is still mine — the framework, the judgment, the decision. The AI is the instrument, not the author.
Adversarial review
I regularly ask Claude to argue against positions I am about to take — to find the weakest point in my reasoning, the assumption I have not examined, the alternative explanation I have not considered. This is the most valuable single use I have found. Every analytical framework I have built has been strengthened by running it through this kind of pressure test before it meets a real audience.
Most people use AI as a yes-machine — a tool for producing more of what they already think. The organizations that will extract the most value from AI are the ones that build the discipline to use it as an adversary. To ask it: what am I missing? Where am I wrong? What would a skeptic say?
Domain translation
One of the underappreciated capabilities of large language models is their ability to translate a concept fluently across domains. When I am building an analytical framework for an industry I have not worked in directly, I can use Claude to rapidly understand the vocabulary, the conventions, and the known failure modes of that domain. This does not replace domain expertise — I am always careful about that — but it dramatically accelerates the orientation process.
The discipline is knowing where the translation ends and where domain expertise must begin. AI can tell you the vocabulary. It cannot tell you where the bodies are buried.
The three things AI does poorly — and what that reveals
Judgment about what matters
AI is exceptionally good at generating options. It is not good at knowing which option is right given the specific human, organizational, and political context of a real decision. That judgment — the experienced practitioner's ability to read a room, understand what a client can actually absorb, and calibrate the output to what will be acted on rather than what is technically correct — is not something AI currently has or approximates.
This is the last mile problem again. The model produces something. The human decides what to do with it. The gap between those two things is where experience lives, and it is not compressible.
Knowing what it does not know
AI systems, including Claude, can be confidently wrong. Not often, in my experience — Claude is notably better calibrated than most LLMs about acknowledging uncertainty — but the failure mode exists and it is dangerous precisely because the output sounds authoritative. The discipline of verifying AI-generated factual claims against primary sources is not optional. It is the minimum standard for professional use.
I think of this the way I think about a brilliant junior analyst: the output is often excellent, occasionally wrong, and always requires the judgment of someone experienced enough to know the difference. If you do not have that judgment, the AI makes you more dangerous, not less.
Continuity of organizational context
Within a session, Claude holds everything. Across sessions, the context resets. This means the institutional knowledge that accumulates in a long-term human working relationship — the understanding of an organization's culture, constraints, history, and political dynamics — does not transfer automatically. Managing that context gap requires discipline: documenting decisions, maintaining working documents that can be shared at the start of a session, building the habit of explicit context-setting.
Organizations that treat AI as a drop-in replacement for institutional memory will discover this limitation expensively. The right model is AI as accelerant applied to human-held context, not AI as repository of organizational knowledge.
What this tells us about AI
The honest conclusion I have reached after more than a year of daily serious use is this: AI is the most powerful thinking tool I have ever encountered, and it requires more rigorous human oversight than any tool I have previously used. Those two facts are not in tension. They are the same fact.
The power comes from scale, fluency, and the absence of ego. The oversight requirement comes from the same source — a system that produces at scale, with fluency, and without the self-correcting instinct that ego sometimes provides, needs a disciplined human in the loop who knows where the edges are.
I am a better practitioner because of these conversations. I am also more aware than ever that the value of the output depends entirely on the quality of the question, the rigor of the review, and the judgment of the person deciding what to do with the result.
The tool does not change the discipline. The discipline changes what the tool can do.