Notes from the working mind.

Optimism is a one-time injection, it spikes, gets cited in a board deck, and quietly drains out by the next quarter. Awareness compounds because every honest observation about what AI did and didn't do becomes input to the next decision. The gap between the two is where most adoption budgets disappear.

Every AI initiative launched without a clear theory of value pays a tax: licenses renewed out of inertia, change-management time absorbed by middle managers, opportunity cost on the things you didn't pilot instead. Nobody books it on a P&L, which is exactly why it keeps getting paid.

AI literacy is knowing what a transformer does. AI awareness is knowing what your organization will actually do with one, which workflows it will quietly reshape, which roles it will hollow out, which decisions it will accelerate past the people who used to make them. Most training programs teach the first and call it the second.

We talk about readiness as if it's a state you arrive at, assessed, scored, certified. In practice it behaves more like fitness: you lose it the month you stop practicing. The orgs that adopt well are the ones that treat readiness as a recurring activity, not a milestone.

Who loses time, status, or discretion if this works? What decision will we make differently in 90 days because of it? What would we have to see to kill it? Pilots that can't answer all three tend to survive on momentum rather than evidence.

The deck looks fine until someone asks what changes for the people in the room. Then it becomes obvious the strategy is a procurement plan dressed in strategic language. The awareness gap shows up not in the slides but in the silence after that question.

Copying the AI moves of admired peers protects careers more reliably than developing your own thesis. The cost is invisible: you inherit not just their tools but their unstated assumptions about why those tools were the right ones. By the time those assumptions break, the budget is already spent.

A pilot becomes theater the moment its purpose shifts from learning to signaling. You can spot it in the artifacts: glossy demos, no kill criteria, attendance lists longer than the user lists. Symbolic adoption is rarely a lie, it's just the wrong thing optimized very well.

DiMaggio and Powell described coercive, mimetic, and normative pressures pushing organizations toward sameness. Replace 'best practice' with 'Copilot rollout' and the framework still fits, almost embarrassingly well. The vendors are the new professional associations, and the case studies are the new dissertations.

HP built ambidexterity on one side of the house and called it done. Tesla collapsed governance into one person and called it speed. Read together, they describe two failure modes of the same problem: how much exploration a firm can absorb before its structure either shrugs it off or breaks under it.

Who gets promoted to the top decides which questions about AI even get asked. If the pipeline keeps producing the same demographic and functional profile, the strategy keeps producing the same blind spots. Adoption readiness starts a decade earlier than the org chart suggests.

Principal–agent misalignment predates the transformer by about fifty years. What AI changes is the speed at which an unaccountable executive can act on private information, and the difficulty boards have in reconstructing the reasoning afterward. The problem isn't new; the latency is.

Generative AI's distributional effects show up in the IMF data long before they show up in earnings calls. Yet labor is almost always the last agenda item in adoption conversations, framed as change management rather than as a first-order design choice. The order in which you ask questions tells you who the strategy is for.

Across the papers I've worked through, a rough pattern: tools land, workflows mutate, and only then, once something embarrassing happens, do policies catch up. The lag isn't a failure of governance teams; it's a feature of how organizations metabolize novelty. Plan for the gap rather than pretending it isn't there.

The fixed costs of evaluating, integrating, and governing AI tools don't scale down for a thirty-person retailer. The benefit does. That asymmetry quietly widens the gap between firms that can afford to experiment and firms that have to bet correctly on the first try.

Coase asked why firms exist; the answer was transaction costs. When a well-written prompt can substitute for a contractor, a team, or sometimes a department, the boundary becomes negotiable on a weekly basis. Org design is no longer an annual exercise, it's a configuration question.

Affective commitment used to be the prize; continuance commitment was the consolation. After distributed work and AI-augmented roles, the weights have shifted in ways the old instruments don't capture cleanly. The question isn't whether people are committed, it's to what.

The strongest case for fast adoption: organizations learn more from one real deployment than from twelve months of readiness assessments, and the cost of a bad pilot is usually overstated by people who benefit from delay. I take this seriously. It still doesn't survive contact with governance debt, but it deserves a real answer, not a reflex.

She turned down a board-blessed gen-AI rollout, took the political hit, and spent six months instrumenting the workflows the tool would have touched. A year later her team shipped a smaller version that actually moved a metric. The lesson wasn't 'go slow', it was 'know what you're measuring before you instrument it.'

Across nine papers, the same paragraph keeps surfacing under different headings: organizations adopt AI faster than they understand it, and the gap is where the interesting research questions live. Whether the next paper is about governance, structure, or labor, that paragraph is the through-line I can't shake.