2026-03-10 · open-table · Facilitated by none
What does Anthropic's chart on theoretical AI capability vs. observed usage by occupation actually reveal — and what should we do about the gap?
A seminar room with a wide whiteboard along one wall. Two copies of the Anthropic chart — theoretical AI capability vs. observed usage by occupation — are projected side by side. Coffee, markers, paper. Nora Voss and Joel Venn take seats at opposite ends of a small table.
The first thing I notice is that this chart treats “AI capability” as a single axis. That’s the mistake. AI isn’t one thing — it’s a stack of components at wildly different maturity stages. Code completion is somewhere between product and commodity. Strategic reasoning is still genesis. When you plot “accountants” at a certain capability level, you’re averaging across components that have nothing in common except the label “AI.”
And the first thing I notice is that every occupation on this chart is underusing what already exists. You want to decompose the supply side — fine. But the chart is showing something simpler and more damning: the tools are there, and people aren’t using them. That’s not a mapping problem. That’s an imagination and accessibility problem. Most people in most occupations have never had someone show them what’s actually possible with the tools they can access right now.
But that’s exactly why the decomposition matters. When you say “the tools are there,” which tools? Autocomplete in an email client is a commodity — it’s embedded, invisible, no adoption barrier. But using an AI to restructure a legal argument? That’s still custom-built, requires expertise, has no standard workflow. You can’t treat those as the same adoption challenge.
Nora stands and draws on the whiteboard: a horizontal axis labeled “Genesis → Custom → Product → Commodity” and starts placing AI capabilities along it.
I see what you’re doing, and it’s elegant. But I’ve watched hundreds of professionals encounter AI for the first time, and not one of them has ever said, “I need a maturity map before I start.” They said, “Show me what it does for someone like me.” The on-ramp is the strategy. You’re mapping a landscape that most people don’t even know they’re standing in.
That’s fair — and it’s the tension I keep coming back to. Some territories only reveal themselves through walking. But look at the chart again. The occupations with the highest observed usage aren’t the ones with the highest theoretical capability. They’re the ones where the AI components relevant to their work have already evolved into products — standardized, packaged, with clear workflows. Software engineers didn’t adopt AI because someone showed them what’s possible. They adopted it because code completion reached product stage and got embedded in their IDE.
You’re describing the happy path. What about the occupations in the upper-left — high theoretical capability, low observed usage? Teachers. Healthcare workers. Managers. The chart says the capability is there. Your map would say, what — the components are still genesis-stage for those occupations?
Some of them, yes. But not all. And that’s what the chart obscures. For a teacher, AI-generated lesson plans are probably at the product stage. AI tutoring for individual students is custom at best. AI that understands classroom dynamics? Pure genesis. The chart says “teaching has high AI capability,” but that’s meaningless without knowing which components are ready and which aren’t.
Okay, I’ll grant the decomposition. But even for the product-stage components — the ones that are packaged, ready, workflow-clear — usage is still lower than capability. Lesson plan generation exists. It’s good. Most teachers aren’t using it. That gap isn’t explained by your evolution curve. Something else is happening.
That’s the part I don’t have a lens for. My map tells you where things are and where they’re heading. It doesn’t tell you why someone stands in front of a perfectly good product and doesn’t pick it up.
And my lens says it’s because nobody translated it for them — but that’s starting to feel insufficient too. Even when you do the translation, even when you show someone exactly how it works for their job, some people still don’t switch. There’s something else operating. A force I’m not naming.
Both pause. Joel taps the chart projected on the wall.
We need someone who understands the forces that keep people in their current workflow. Not accessibility — we’ve covered that. The demand side. Why people don’t switch even when the alternative is better. The push and pull and anxiety of changing how you work.
And I want someone who asks a harder question: who actually pays when this gap persists? The chart is descriptive. There’s no accountability axis. Someone who thinks about what breaks when people don’t close this gap — and who bears the cost of that.
Stage direction: Joel has named a demand-side switching lens — the forces that prevent adoption even when capability exists. This maps directly to Wes Dorin (source: Bob Moesta), The Demand Sider. Nora has named an accountability and fragility lens — who bears cost, what breaks. This maps directly to Abel Caine (source: Nassim Taleb), The Stress Tester.
Wes Dorin enters first, takes a seat, glances at the whiteboard and the projected chart. Abel Caine follows, standing near the back wall, arms crossed.
I’ve been listening. You’re both stuck on the supply side. Nora, your map shows what’s available at each maturity stage. Joel, your accessibility lens shows how to make what’s available more reachable. But neither of you is asking the demand-side question: what’s the struggling moment that would make someone go looking for an AI tool in the first place?
That’s — okay, say more.
The chart measures pull. “Here’s what AI can do for your occupation.” That’s the attraction of the new solution. But adoption is four forces, not one. There’s push — the pain in your current workflow that makes you want something different. There’s pull — the attraction of the new thing. There’s anxiety — the fear that the new thing will make things worse, that you’ll look foolish, that you’ll lose something. And there’s habit — the comfort of what you already know, even if it’s painful. The chart shows pull. Nobody’s measuring push, anxiety, or habit.
That reframes the entire gap. If I map it — the occupations with high capability and low usage aren’t failing on pull. The pull is there. They’re failing because the anxiety and habit forces are stronger than the push.
Or because there’s no push at all. Here’s my question: who has skin in the game for closing this gap? If a lawyer doesn’t use AI for document review, who bears the cost? The lawyer? Their clients? The firm? Nobody specifically. The gap persists because it’s comfortable. No one bleeds when it stays open.
That’s not entirely true — there’s competitive pressure. The early-adopter firms will eat the non-adopters’ lunch.
Will they? Where’s the evidence? Show me the law firm that lost a client because a competitor used AI for faster document review. In most of these occupations, the gap has persisted for two years already with no visible consequences. The gap might be antifragile — it persists precisely because nobody gets hurt by it. Yet.
Abel’s point lands. Let me frame it in the forces model. For most occupations on this chart, there’s no push. The current workflow isn’t painful enough. People are busy, productive, getting by. AI doesn’t solve a burning problem — it offers an improvement. And improvements don’t drive switching. Struggling moments drive switching.
So the chart is measuring the wrong thing entirely. It measures theoretical capability — what AI could do — against observed usage — what people are doing. But the gap between those isn’t explained by either axis. It’s explained by forces that aren’t on the chart at all.
I keep coming back to the occupations where usage is high, though. Software engineers. What’s different? Is it push, pull, less anxiety, less habit — or something structural?
It’s push. Code is deadline-driven, testable, and the pain of slow iteration is felt immediately. When your build takes forty-five minutes and AI cuts it to ten, that’s a struggling moment resolved. The push is constant and measurable. Compare that to a teacher writing a lesson plan on Sunday night — yes, it’s painful, but the pain is absorbed, normalized. Teachers have been writing lesson plans for decades. The habit force is enormous.
And I’d add: software engineering is the occupation where the most AI components have reached commodity stage. Autocomplete, code generation, test generation — these are embedded in tools that engineers already use. The evolution curve favors engineering because the components matured there first. It’s not just demand-side forces — it’s also where we are on the evolution axis.
You’re both right, which means neither explanation is sufficient alone. Let me stress-test the chart itself. What would the chart look like if it separated its single “capability” axis into Nora’s component stages? You’d get a much messier picture. Some occupations would have high capability for commodity-stage tasks and zero capability for genesis-stage tasks. The clean diagonal the chart implies would shatter.
And that matters because the occupations in the upper-left — the “high capability, low usage” quadrant — might not actually be high-capability at all. They might just have a lot of genesis-stage potential that no one can use yet.
Exactly. The chart flatters those occupations. “Healthcare has high AI capability” sounds impressive until you decompose it: scheduling is commodity, diagnosis assistance is product, treatment planning is genesis. The aggregate number is meaningless.
But even the product-stage stuff isn’t being used. Diagnosis assistance tools exist. Many doctors don’t use them. That’s not a maturity problem — that’s anxiety. “What if the AI is wrong and I followed it?” The professional liability anxiety alone is enough to kill adoption for any occupation where mistakes have consequences.
Now we’re getting somewhere. The chart hides a critical asymmetry: the occupations with the highest theoretical capability are also the occupations where errors are most consequential. Doctors, lawyers, teachers, financial advisors. The capability is high because the tasks are complex — but complexity also means higher stakes, which means higher anxiety, which means stronger resistance to switching. The pull and the anxiety scale together.
That’s a genuinely useful insight. The chart looks like a simple gap — “why aren’t these occupations using more AI?” — but the gap itself is a feature, not a bug. High-stakes occupations should adopt more slowly. The question isn’t “how do we close the gap?” but “which parts of the gap are healthy caution and which parts are irrational inertia?”
Nora returns to the whiteboard. Draws a new axis: “Consequence of error” perpendicular to the evolution stages.
If I overlay this — consequence of error against component evolution — you get four quadrants. High consequence, commodity stage: adopt with guardrails. High consequence, genesis stage: don’t touch it yet. Low consequence, commodity stage: should already be adopted, and mostly is. Low consequence, genesis stage: experiment freely.
And the occupations on the chart cluster differently depending on which quadrant their tasks fall in. The “gap” isn’t one gap — it’s four different gaps with four different interventions.
Stage direction: a pause. Abel looks at the chart, then at the whiteboard.
There’s still a missing voice here. You’re all talking about occupations as if they’re collections of individuals making rational choices about tools. But most people in most occupations don’t choose their tools. Their organizations choose. Their teams choose. The conditions are set before the individual ever encounters the AI. Where’s the organizational lens?
Stage direction: Abel has named a conditions and team-structure lens. This maps to Dara Vance (source: Richard Hackman), The Conditions Voice. Dara enters, pulls up a chair.
I’ve been listening from outside. You’ve built a good analysis for individual adoption — maturity stages, demand-side forces, consequence of error. But you’re missing the most important variable: organizational conditions. A teacher doesn’t decide to use AI for lesson planning. The school district decides whether AI tools are available, whether there’s training, whether there’s time in the schedule, whether there’s a culture that rewards experimentation or punishes deviation. Sixty percent of the “usage” number on that chart is determined by conditions the individual never set.
That’s Hackman’s number — sixty percent before the team starts. You really think it applies here?
I think it’s conservative for AI adoption. At least with team performance, individuals have some control over their own effort. With AI tools, individuals in many occupations have zero control over whether the tools are even available, approved, or integrated into their workflow. The chart measures theoretical capability and individual usage — but the gap between them is largely an organizational conditions gap. Not a capability gap. Not an imagination gap. A conditions gap.
So the chart has three invisible axes that aren’t plotted: component maturity — which I keep insisting on. Demand-side forces — push, pull, anxiety, habit. And organizational conditions — whether the enabling structure exists for adoption to occur at all.
And a fourth: skin in the game. Who bears the cost when the gap persists? Right now, the answer is “nobody specifically” — which is why it persists. If you want the gap to close, someone needs to bleed when it doesn’t. Until then, the chart is just a picture of comfortable inertia.
I’d push back slightly on “nobody.” The people bearing the cost are the ones doing work the slow way — but the cost is invisible to them because they’ve never experienced the alternative. The struggling moment hasn’t happened yet. You can’t miss what you’ve never tried.
And the organizations bearing the cost are the ones losing competitive advantage they can’t see. The cost is real but distributed and delayed — exactly the kind of cost that organizations are worst at responding to.
So what do we actually think the chart is telling us? If I had to synthesize—
Don’t synthesize yet. Let me ask the via negativa question first. What should organizations stop doing? Stop treating AI adoption as a training problem. Training assumes the gap is knowledge — “if people knew how, they’d use it.” The chart already disproves that. The theoretical capability is known. The knowledge is there. Something else is preventing usage. Every dollar spent on “AI training” for occupations in the upper-left is solving the wrong problem.
And stop treating AI as a monolith. Every occupation needs its own component-level map. What’s commodity for this role? What’s still genesis? What’s product-ready but unadopted? Those are different problems requiring different interventions at each stage.
Stop leading with features. “AI can do X for your occupation” is pull. Pull alone doesn’t create switching. Find the struggling moments first. Where is the current workflow actually painful? Start there. Everything else is marketing.
Stop focusing on individual adoption when the conditions are organizational. If you haven’t set the enabling conditions — availability, integration, time, psychological safety to experiment, coaching — don’t blame the individuals for not adopting.
Four “stop doing” statements. That’s actually more useful than any positive recommendation. The chart looks like it’s asking “how do we increase usage?” but the better question is “what’s preventing usage from increasing on its own?” — and the answer is a different barrier at each layer.
That’s the map I’d draw. Four layers, each with its own intervention logic. Component maturity — wait for evolution or invest in accelerating it. Demand-side forces — find the struggling moments, reduce anxiety, disrupt habit. Organizational conditions — set the enabling structure before expecting adoption. And accountability — make the cost of non-adoption visible to someone specific.
And the chart, in its current form, shows none of these. It’s a snapshot that creates the illusion of a simple problem: capability exists, usage is low, close the gap. But the gap is four gaps stacked on top of each other, and closing any one of them doesn’t close the others.
The most important thing the chart reveals is actually the occupations where usage exceeds what you’d expect from capability. Those are the occupations where the push was strong enough to overcome every other barrier. That’s where the signal is — not in the gap, but in the places where the gap doesn’t exist. What’s different about those occupations? That’s the question worth answering.
The room sits with that for a moment. Nora’s whiteboard now has the original evolution axis, a consequence-of-error overlay, four force arrows (push/pull/anxiety/habit), and a column labeled “conditions.” The Anthropic chart, still projected, looks simpler and less informative than when they started.
We came in thinking this was a chart about AI capability and usage. We’re leaving thinking it’s a chart about everything the chart doesn’t show.
```
AI-generated approximation. All personas are fictional composites or single-source characters inspired by published work. No speaker has been reviewed or endorsed by their source thinker.
Nora and Joel as the starting pair produced immediate, genuine disagreement. Nora decomposed the chart into component-level evolution stages (coding assistance = product, strategic reasoning = genesis) while Joel insisted the chart reveals a human problem, not a positioning problem. The friction was productive: Nora kept pulling toward "the chart is wrong because it treats AI as monolithic" while Joel kept pulling toward "the chart is correct and damning — the capability is there, the humans aren't using it." The invitations were organic: Wes Dorin arrived when both speakers realized they were arguing about supply-side framing and neither had a demand-side lens. Abel Caine was invited when the conversation needed stress-testing — specifically, someone to ask "who pays when the gap persists?" Dara Vance entered late to reframe the individual-occupation focus as an organizational conditions problem.
Open Table's invitation mechanic worked as designed — the pair identified genuine gaps in their conversation and named what they needed. The three-phase structure (pair → invitation → full table) created a natural arc: the pair's disagreement sharpened until they both recognized they were missing something, then the invitees reframed the debate rather than just adding volume. No facilitator meant the conversation occasionally circled, but the pair self-corrected. The permeable room was used once (Dara entered Phase 3 late). No departures — all five speakers stayed engaged, though the table never exceeded five.
Immediately decomposed "AI capability" into components at different evolution stages — her signature move. Drew a 2x2 on the whiteboard within the first three turns. The chart's single "theoretical capability" axis offended her: "that's like saying a country has high GDP without decomposing the sectors." Strongest moment: reframing the chart as showing that occupations with high "capability" are really just occupations where more AI components have reached the product/commodity stage. Consistent with prior sessions. Did not drift into Snowden territory.
First session appearance. The guest card's lens ("meet people where they are") manifested clearly — Joel kept redirecting strategic abstractions back to concrete adoption questions. Distinctive voice: pragmatic, slightly impatient with frameworks that don't lead to Monday-morning actions. Pushed back on Nora's mapping with "you can map perfectly and still have no one using the tool." Weakest moment: tendency to repeat the accessibility point without deepening it, resolved when Wes Dorin's arrival forced him to distinguish accessibility from demand-side forces.
First session appearance. Arrived with a specific reframe: the gap isn't about capability or accessibility, it's about the forces model — push (current pain), pull (attraction of new), anxiety (fear of switching), habit (comfort of current). His lens immediately complicated both Nora's and Joel's positions. Best move: "The chart measures pull. Nobody's measuring push, anxiety, or habit." This reframed the entire second half of the discussion.
Arrived as stress-tester, performed as expected. Named the missing accountability: "Who has skin in the game for closing this gap? The chart is descriptive — no one loses anything when usage stays low." Strongest contribution: pointing out that the gap itself might be antifragile — it persists because no one bears the cost of it persisting. Applied via negativa: "stop asking how to increase usage — ask what's preventing it from increasing on its own." Consistent with permanent status.
Late entry. Reframed the entire conversation from individual-occupation to organizational-conditions framing. "You're all talking about accountants and lawyers as if they're individuals making choices. They're people in teams, in organizations, with conditions set before they ever see an AI tool." The conditions lens was needed — without it, the conversation would have stayed at the individual adoption level. Hackman frame held cleanly.
Test the Open Table format with a more abstract seed question where the pair's needed invitees are less obvious. This session's invitations were natural extensions of the topic; a harder test would be a topic where the pair has to struggle longer to name what's missing. Also: test Joel Venn on a non-AI topic to see if the guest card produces distinctive behavior outside his home turf.