I write to think in public about what organizations get wrong about AI and what the research actually supports.
Clinicians who adopted AI support for tumor detection lost about six percent of their unaided detection ability within three months, a finding relayed in the International AI Safety Report 2026…
In a study of 905 participants, AI personas that were never identified as AI measurably shifted how teams behaved, and no dashboard caught it. We are measuring human-AI teams with instruments built for human teams. Throughput cannot see the two things that actually drive performance: whether the role was redesigned or just loaded, and whether trust is calibrated or assumed.
In a 2,257-person sample, AI threat perceptions and depth of use varied significantly by job level. The people carrying the most expertise are the ones whose use stalls, and they stall for reasons unrelated to not knowing how the software works. That is identity threat at work, and it is a design problem, not a sentiment problem.
Ask a room of knowledge workers whether AI will eliminate their own job in the next two years, and most say no. Ask the same room whether AI will eliminate the job of the person next to them, and the number climbs. Employee sentiment and employer hiring data are pointing in opposite directions, and the size of that gap is what deserves scrutiny.
Picture a procurement agent with two permissions that score identically low-risk: approve invoices under $2,000, and cancel duplicate purchase orders. Then it cancels an order for a component that has already shipped. The dollar amount is small. The decision cannot be undone, and reversibility, not the risk score, is the question nobody asked.
Consider the case of an endoscopist with a decade of training who spends months working alongside an AI tool that flags suspicious lesions before her eye reaches them. Then, for a routine competency check, the tool leaves the room. Her unassisted detection rate falls. Any review workflow where AI does the first-pass catching is a candidate for the same quiet erosion.
For decades, organizations designed work around specialization. AI is exposing the gap between how roles are scoped and how work actually happens. The composite role is the organizational response, and most companies are letting them emerge accidentally. Part 1 of a series with Chris Lloyd-Jones.
Wilson and Daugherty (2018) proposed three mechanisms of collaborative intelligence. The enterprise AI literature can tell us who is using AI and how often. It cannot tell us whether the humans and the systems are getting better together. Co-evolution is the most consequential mechanism, and it is almost entirely unmeasured.
Anthropic published its AI Fluency Index. I have been thinking about these constructs for two years. The study gets something important right: iteration predicts fluency, and polished outputs suppress evaluation. But fluency is necessary, not sufficient. The construct bridging individual fluency and organizational value is collaborative intelligence, and the mechanism producing compounding returns is co-evolution.
Every week I see another AI readiness assessment making the rounds. Score your organization. Get a maturity rating. But readiness toward what? The infrastructure question gets attention. The destination question does not. And without a destination, readiness optimizes for deployment speed rather than strategic value.
We have been assuming that frontier AI requires frontier investment. But the most disruptive force may not be what AI can do, it may be how cheaply it can do it. DeepSeek trained a frontier-competitive reasoning model for $294,000. The compute moat is dissolving. The question is what replaces it.
No one is announcing AI-driven restructuring. But reporting lines are shifting. Junior roles are being quietly absorbed or redefined. Competency models that mattered eighteen months ago are already outdated. The reorg is happening. We are just not calling it that.
I have been building AI agents, a robotic arm, and now an animatronic. And I keep asking myself: why is software not enough? The impulse to embody AI reflects a structural gap in how we design intelligent systems.
The entire conversation right now is about models. Who has the biggest one. Who has the fastest one. But a different question deserves attention: what about the connective tissue? The protocols that determine how AI systems talk to tools, to data, to each other? When Anthropic, OpenAI, and Google converge on shared protocol infrastructure, the implication for every other organization is clear. The protocol layer is not where you differentiate. It is where you participate or get designed around.
Every AI pitch deck I see leads with efficiency. Save time. Cut costs. Do more with less. But efficiency is a ceiling, not a floor. The work redesign literature shows that the most consequential innovations do not optimize existing tasks. They create entirely new categories of work. So what happens when we stop asking what AI makes cheaper and start asking what AI makes newly possible?
I built an AI-generated outlaw country band called Southern Oracle. Three women, a debut album, and a production capacity I lacked but a creative vision entirely my own. What I learned about co-evolution, authorship, and the line between human and machine creativity that no strategy document could have taught me.
I keep hearing the same request: "We need more use cases." But the use-case mentality treats AI like a vending machine. Insert problem, receive solution. The organizations I see pulling ahead are not building use case libraries. They are building adaptive capacity. Job design research tells us that when work is dynamic, prescription fails. What if the goal is not to identify the right 50 applications, but to build a system that surfaces the next 500 on its own?
Adoption is the metric everyone tracks. But what does it actually tell you? That someone logged in. That a button was clicked. I-O research on training transfer has shown for years that exposure alone does not produce behavior change. You need reinforcement, feedback loops, environmental support. My hypothesis is that the organizations scaling AI successfully are measuring something else entirely: learning velocity, or how fast teams integrate new capabilities into how they actually work.
There is a pattern that repeats across industries. An organization deploys a new AI tool. Adoption is uneven. Leadership labels it a "change management problem" and dispatches internal comms. But decades of research on organizational change suggest something different: resistance is not noise. It is signal. When people push back on a system, they are surfacing a mismatch between how work actually happens and how someone assumed it happens. The question worth asking is not how to overcome resistance faster, but whether resistance might be the most honest feedback mechanism an organization has.
On the bifurcation risk in organizational learning capacity, and what happens when a small number of organizations figure out human-AI integration while the rest are still running pilots.
On composite professional identities, the shift from single-role careers, and what career architecture looks like when the unit of work is the project, not the position.
On the gap between controlled AI evaluations and real-world performance, and what complexity, context, and organizational friction do to benchmarks.