Part 1 of this series — on how AI changes executive thinking — is here. I write about AI, behavioral science, and the intersection of technology, finance, banking and human systems.
It’s 11:47 PM on a Sunday and I’m arguing with an AI about the architecture of an octopus’s nervous system.
Not metaphorically. Literally. I’m writing a 7,000-word technical article about AI architecture in banking, and I’ve decided that the central metaphor should come from Peter Godfrey-Smith’s Other Minds — the idea that an octopus distributes intelligence across its arms, each arm capable of independent decision-making while the central brain sets intent. I think this maps to a specific engineering architecture I’ve been designing. The AI thinks the metaphor breaks down at the third layer. We’ve been going back and forth for twenty minutes.
It’s the most productive argument I’ll have all week.
A few months ago I wrote about how AI changed the way I think as an executive — three shifts in strategy, signals, and decision-triage. That article was about cognition. About using AI as an external prefrontal cortex that pressure-tests your assumptions and reveals your blind spots.
This article is about something different. This is about the craft — the concrete, unglamorous, often late-night workflow of actually making things with AI as a collaborator. Strategy documents. Technical articles. Architectural diagrams. Prototype scenarios. Board presentations. Product specifications. The artifacts that turn ideas into influence.
The first article answered: how does AI change what an executive thinks?
This one answers: how does AI change what an executive produces?
The answer is more uncomfortable than you’d expect — because the skill that matters most isn’t prompting. It’s knowing what to throw away.
A few weeks ago I published a technical article — roughly 7,000 words on AI architecture in banking, covering neuroscience, reinforcement learning, behavioral economics, and five different executive audience sections. A reader messaged me: “This must have taken months to research and write.”
It took eleven days.
I need to be precise about what that means, because the wrong interpretation is “AI wrote it fast.” AI didn’t write it. I wrote it. Every sentence is mine. Every analogy, every argument structure, every technical claim — mine. If you fed the article into an AI detector, it would register as human-written, because it is.
What AI did was compress the process — not the output. It collapsed what would have been months of solitary research, false starts, structural dead ends, and multiple complete rewrites into eleven days of intense, iterative, adversarial collaboration. The quality didn’t decrease. If anything, it increased — because the AI forced me to defend every claim, justify every analogy, and rethink every structural choice in real time instead of discovering the problems three drafts later.
But the agentic workflow that produced that result looks nothing like what the AI vendors describe. It’s messier, more argumentative, and involves deleting far more than keeping. Here are the five modes I actually use.
01.The first mode is what I call the skeleton extraction. I never ask AI to write something for me. I ask it to reveal the structure of something I’m about to write.
Here’s the actual process: I start with a thesis I believe in, a set of ideas I’ve been developing for weeks or months, and a rough sense of who the audience is. I dump all of this into a conversation — unstructured, fragmented, full of half-formed arguments and notes from three different notebooks. Then I ask: “Given everything I’ve told you, what’s the architecture of this argument? What are the sections? What order should the logic flow? What would a reader expect to see that I haven’t mentioned?”
The AI returns a structure. It’s usually wrong — not in architecture, but in emphasis. It gives equal weight to everything. It doesn’t know which ideas I’ve been thinking about for a year and which I mentioned offhandedly. It doesn’t know which arguments I have evidence for and which are intuitions I haven’t validated. It doesn’t know which sections will be the ones readers remember and which are necessary but forgettable connective tissue.
But the structure is useful — the way a rough sketch is useful before painting. It shows me the shape of the argument from outside. I can see where the gaps are. I can see which sections I thought were important but actually don’t serve the thesis. I can see which arguments need to come before which other arguments, because the AI — processing my ideas without my emotional attachment to them — naturally sorted them by logical dependency rather than by the order I thought of them.
I keep about 30% of the AI’s proposed structure. I rearrange the rest. I delete sections that the AI thought were important but that I know are digressions. I add sections the AI didn’t suggest because it didn’t have the context — the personal experiences, the cross-cultural references, the specific industry knowledge that make the argument mine rather than generic.
Then I write. From zero. In my own voice. Using the skeleton as a map, not as a draft.
The result is an article that’s structurally tighter than anything I’d have produced by writing linearly from paragraph one to the end — because the structural thinking happened first, separately, adversarially. The AI didn’t write the article. It showed me the article’s bones before I grew the muscle.
02.The second mode is the one that’s most specific to how I think, and it’s the one I’ve never seen anyone else describe: the cross-domain translator.
I think in analogies. Always have. When I encounter a technical problem, my instinct isn’t to solve it directly — it’s to ask “what is this like?” and then search across domains until I find a structural parallel that illuminates the problem from an angle that pure technical analysis can’t reach.
The bazaar I used to visit as a child — the one where the saffron merchant read your gait, calculated your intent from your clothes and your hesitation, and adjusted his pitch in real time — became my metaphor for how a customer intelligence platform should work. An octopus distributing intelligence across its arms became my metaphor for agentic AI architecture. Montaigne’s refusal to pick sides during the French Wars of Religion became my metaphor for how leaders should navigate competing stakeholder factions. The neuroscience of habit formation — cue, routine, reward loops from MIT’s basal ganglia research — became the design framework for a financial coaching product.
None of these analogies arrived fully formed. Here’s the actual workflow:
I start with a direction, not a destination. “I need an analogy for a system where central intent is set but local execution is autonomous.” I feed this to the AI along with the technical architecture I’m trying to describe. Then I ask it to explore: find me parallels in biology, in military strategy, in urban planning, in music, in ecology.
The AI returns a dozen options. Most are superficial — the kind of analogy a TED talk would use: “It’s like a beehive!” No. Too simple. The queen doesn’t set intent the same way. “It’s like a jazz ensemble!” Closer — but jazz improvisation is too unstructured for what I’m describing. “It’s like a Roman legion!” Wrong — legions are too hierarchical. Red Team AI helps too.
Then the octopus appears. Each arm has its own neural cluster. Each arm can taste, touch, and decide independently. The central brain sets intent but doesn’t micromanage. The arms coordinate through shared chemical signals, not through top-down commands.
That’s the one. But I don’t accept it immediately. I pressure-test it. “Where does this analogy break down?” The AI tells me: the octopus’s arms don’t learn from each other the way AI agents in an agent mesh do. Good — I note the limitation. “What about Peter Godfrey-Smith’s work specifically?” The AI brings up Other Minds and the specific research on octopus cognition. I read the source. The analogy holds for the first two layers but needs modification for the third.
Twenty minutes of this. The 11:47 PM argument I described at the opening.
The output: a single analogy that carries the entire 7,000-word article. An analogy that readers remember long after they’ve forgotten the technical details. An analogy that I discovered through AI-assisted exploration across five domains but that only works because I have the domain expertise to judge which parallels are structurally valid and which are merely decorative.
The skill here is taste. AI can generate a hundred analogies. It cannot tell you which one illuminates and which one decorates. That distinction requires having lived in enough domains — technology, banking, Asian or other cultures, Vietnamese business, neuroscience, philosophy — to know when a metaphor is carrying weight and when it’s performing.
AI gave me the octopus. I knew it was the right octopus. That knowing is the part AI can’t do.
03.The third mode is the one that surprises people most when I describe it: the visual architect. Most executives think of AI as a text tool. It writes. It summarizes. It analyzes documents. What they don’t realize — and what I didn’t fully realize until this year — is that AI can build presentation-ready architectural diagrams, interactive prototypes, and visual artifacts that would normally require a UX designer and two weeks of iteration.
Here’s a concrete example. I needed to communicate a complex five-layer AI architecture to a product leadership team. The architecture had six zones, each with multiple components, with data flows between them and a governance layer spanning the top. Traditional approach: I describe it to a designer, the designer produces a draft in Figma, I review and give feedback, they revise, we iterate three or four times over a week. The result is professional but slow.
What I actually did: I described the architecture to the AI in natural language — “five zones stacked vertically, FullStoryAI at the top as the executive layer, Coach and CLO in the middle, intelligence substrate below that, AI Hub below that, integration at the bottom, governance bar across the top.” The AI generated a complete HTML/CSS diagram. I opened it in a browser. The layout was 70% right but the visual hierarchy was wrong — the intelligence substrate zone was too prominent and the executive layer wasn’t prominent enough.
I gave feedback in natural language: “Zone 1 needs more visual weight. The roadmap items should have dashed borders to distinguish them from current capabilities. The intelligence substrate needs to be horizontally organized, not vertically.” The AI revised. We went back and forth six times. Total elapsed time: forty-five minutes. The result was a presentation-quality architectural diagram that I used in a board-level discussion the next day.
A week later, my UX team was busy with a product deadline. I needed a working prototype of an executive command center — not a mockup, a working prototype with clickable flows, real data scenarios, and structured response cards. The kind of prototype that normally takes a UX designer two weeks.
I built it with AI in an afternoon. A full interactive scenario: a bank executive opens the system on a Monday morning, sees a proactive briefing about weekend savings closures, analyzes the competitive threat, compares three response strategies, builds a campaign, approves it, and deploys — all in one conversation flow. Every screen, every card, every data visualization, every approval flow. Working in a browser.
I’m not a designer. I have opinions about design — strong ones — but I don’t write CSS for a living. The AI does the implementation. I do the direction, the critique, and the judgment about what a bank executive actually needs to see and feel when they interact with this system. That’s the split: I can’t build the prototype alone, and the AI can’t design it alone. Together, we produce something that would have required a cross-functional team and a sprint cycle.
This mode has changed what I consider possible within a single executive’s scope. Artifacts that used to require delegating to a team — and accepting the communication loss and timeline delay that delegation entails — can now be produced directly, at full fidelity, in hours. The executive’s ideas reach the audience without the game of telephone that usually distorts them.
The risk, and I want to be honest about it: this capability can make you a bottleneck if you’re not careful. The fact that you can build the diagram yourself doesn’t mean you should build every diagram yourself. The value is in the moments where speed matters more than process — the board meeting tomorrow, the competitive response this week, the prototype needed to align the team before a design sprint begins. For sustained production, you still need the team. The AI doesn’t replace the UX designer. It lets you have a conversation with the UX designer that starts at version 3 instead of version 0.
04.The fourth mode is the one I use most frequently and think about least — which probably means it’s the most deeply embedded in my workflow: the audience refractor.
Any senior executive who communicates across levels — board, C-suite, department heads, engineering teams, customers, regulators — knows that the same idea needs fundamentally different framing for each audience. Not just different vocabulary. Different emphasis, different evidence, different emotional register.
A board member needs: strategic implication, competitive positioning, financial impact, risk assessment. They want to know why this matters for the bank’s market position.
A Chief AI Officer needs: technical architecture, model specifics, data strategy, integration approach. They want to know how it works and whether the engineering is sound.
A Chief Risk Officer needs: governance framework, explainability, audit trails, fair lending analysis. They want to know what happens when a regulator asks questions.
A Head of Retail needs: customer impact, operational workflow changes, time-to-value, team implications. They want to know what changes on Monday morning.
I discovered this mode by accident. I was writing a long technical article and realized it needed to speak to five different C-suite roles simultaneously — each of whom would read the whole article but care deeply about only one section. I asked my Red Team AI: “For each of these five executives, what is the single most compelling argument from this architecture? And what is the single biggest objection each one will raise?”
The responses weren’t perfect. But they forced me to think about my material from five different vantage points in rapid succession — something that would have taken me a week of separate conversations with actual executives to approximate. I discovered that the argument I thought was strongest (the technical architecture) was actually the weakest section for the CEO audience, who cared about the commercial model. And the section I almost cut (the compliance framework) was the most important section for the CRO — the one that would determine whether the entire initiative had organizational permission to proceed.
Now I use this mode routinely. Before any significant communication — article, presentation, memo, email to a stakeholder group — I ask the AI to read it from each audience’s perspective and tell me what’s missing, what’s misemphasized, and what will land versus what will bounce. The AI doesn’t know these specific people. But it knows the archetype — the general concerns, priorities, and resistance patterns of a CRO versus a CDO versus a CEO — well enough to surface gaps that I’d miss because I’m unconsciously writing for the audience I most identify with. You can also train the agents by feeding more of those roles coms to agents.
The skill: knowing which AI-identified audience gaps are real and which are artifacts of the archetype not matching the specific person. Your CRO might be unusually technically literate. Your CEO might care more about culture than commercials. The AI gives you the general map. You adjust for the specific terrain.
05.The fifth mode is the most counterintuitive, and the one I’d urge every executive to adopt immediately: the adversarial self-review.
After producing any significant piece of work with AI assistance — an article, a strategy document, a product specification, a board presentation — I take the output and paste it into a completely new AI session. A session with no memory of having produced the work. A clean slate.
Then I ask: “Read this critically. What claims aren’t supported by evidence? Where is the logic weakest? What would a skeptical expert challenge? Where does the writing feel generic rather than specific? What’s missing?”
The results are consistently surprising. The fresh session — unburdened by the collaborative process that produced the document — finds holes that both I and the original session were blind to. It catches claims that feel compelling in context but don’t hold up under scrutiny. It identifies sections where the writing shifts from specific (good) to general (lazy). It flags arguments that depend on assumptions I didn’t state explicitly.
This works because of a fundamental asymmetry in how AI operates: it’s better at analysis than at generation. The same model that produces a document with subtle weaknesses is remarkably good at finding those weaknesses when reading the document cold. The generation process has a completion bias — it wants to finish the thought, fill the gap, produce fluent text. The analysis process doesn’t have this bias. It’s looking for problems, not producing prose.
I’ve started thinking of this as the double-pass discipline: never publish, present, or send any AI-assisted work without running it through an adversarial second session. The first session is your collaborator. The second session is your editor. They’re the same model, but the context difference — creator versus critic — produces genuinely different outputs.
The meta-lesson: if you’re not using AI to check AI, you’re trusting a single pass on work that deserves scrutiny. You wouldn’t publish a strategy document without having a colleague review it. Why would you treat AI-assisted work with less rigor?
Here’s the part I’ve been avoiding, because it’s the part that matters most and is hardest to describe: the problem of voice.
Everything I’ve described — skeleton extraction, cross-domain translation, visual architecture, audience refraction, adversarial review — is a workflow. Workflows can be taught. The thing that can’t be taught is the reason those workflows produce something worth reading instead of something that merely exists.
AI-generated prose has a tell. It’s not always obvious. The grammar is correct. The structure is sound. The arguments are logical. But there’s a flatness — a performative precision, a confident neutrality, an absence of rhythm that any serious reader can feel even if they can’t name it. AI prose never hesitates. It never digresses productively. It never makes the surprising word choice that reveals a specific human mind behind the text. It’s fluent the way elevator music is musical — technically correct and spiritually empty.
My competitive advantage as someone who writes and communicates for a living isn’t that I use AI. Increasingly, everyone uses AI. My advantage is that you can’t tell. And the skill that enables that — the skill that no vendor demo teaches and no productivity tip captures — is the willingness to throw away most of what the AI produces and rewrite the rest until it sounds like it came from someone who grew up wandering bazaars, built digital banks in Vietnam, reads Norman Doidge at midnight, and argues with AI about octopus nervous systems at 11:47 PM on a Sunday.
AI can simulate voice. It can match your vocabulary, mirror your sentence length, approximate your tone. But it can’t simulate biography. It can’t simulate the specific combination of experiences, frustrations, obsessions, and hard-won beliefs that make a person’s writing unmistakably theirs. It can’t simulate having been wrong about something important and carrying the scar. It can’t simulate the moment in a bazaar when a merchant taught you more about personalization than any algorithm ever could.
The craft of using AI as a senior executive isn’t about getting AI to produce your work. It’s about using AI to produce raw material that you then transform through the lens of a life that no model has lived.
The 70% you delete is the AI’s contribution. The 30% you keep, rewrite, and reshape — that’s yours. That’s the craft. That’s what makes the work worth reading.
I want to close with something I’ve been thinking about since I wrote Part 1. In that article, I asked a question I couldn’t answer: if AI can do so much of the cognitive work that defines executive value, what is the executive for?
After few months of the workflow I’ve described here, I think the answer is clearer than I expected.
The executive is the one who knows which octopus to pick. The one who knows that a financial coaching product should be designed around the neuroscience of habit formation, not around feature checklists. The one who knows that a compliance section — the one almost cut from the article — is the section that gives the entire initiative permission to exist. The one who knows that a board member needs to feel the commercial architecture in their bones, not read about it in a feature list. The one who knows that a prototype built in an afternoon communicates more than a specification built in a sprint.
AI is the most powerful craft tool I’ve ever used. But a tool doesn’t know what to build. A tool doesn’t know who it’s building for. A tool doesn’t know when the building is done.
That’s the executive’s job. Not to build everything. Not to think of everything. But to know, with the accumulated judgment of a specific life lived across specific contexts, what this particular thing needs to be — and to relentlessly close the gap between what the AI produces and what the work demands.
It’s 12:30 AM now. The article is done. The AI made it 30% better than it would have been — not because it wrote the words, but because it argued, structured, visualized, refracted, and reviewed until my thinking was sharper than I could have made it alone.
But the voice is mine. The octopus is mine. The bazaar is mine.
That’s the craft.
