Selected Work
Different stakes, same discipline: decide where the system acts on its own, where it asks first, and where judgment stays with the person.
Each was the reasonable default. The work showed what it cost.
The experiments below test these same judgment calls at smaller scope.
A continuous score left the lawyer setting a threshold with no basis for choosing one. The concept replaced it with draft / reviewed, where review stays an explicit human act.
See the review statesThe mapping surfaced two different jobs. About a third of sites had a patient-experience team whose mandate was rounding; the rest had service-line leaders fitting it around everything else they owned. One needed depth, the other speed.
See the operating modelsEight panes are scannable in theory and exhausting in practice. The supervision model nests glance, inspect, and intervene, so the supervisor goes only as deep as the moment requires.
See the supervision modelExperiments
Can a research report stay honest after it ships?
Multi-pass quality gatesLive citation checksRe-syncs against source data
Where does an always-on agent actually spend its money?
Token spend trackingAnomaly alertsRegression evals on every commit
Does giving an agent the design system beat pasting a screenshot?
Figma to design-token contextDelivered over MCPScreenshot-diff evals
When should an agent stop and ask for a human decision?
Escalation · review states · handoff patterns
Writing
Underneath the buzzword sits a real design craft: shaping the boundary between what a system does on its own and what it checks with a person first.
Inside
What changes the answer
The design work is deciding where a feature sits on this line, and making that legible to the person on the other side.
Rescheduling an internal focus block is not the same design problem as changing a patient priority or approving access to financial data. The interaction may look similar. The authority is not.
The design work isn't making the agent capable. It's making the boundary visible.
AUTONOMY · APPROVAL · REVERSIBILITY · ESCALATION · AUDITABILITY
FULL ESSAY · IN PROGRESS
Most of the design work that matters happens before any pixel exists. The decks come later. The brief is the deliverable.
On patient rounding, the critical decision was not the eventual screen. It was recognizing that the proposed product boundary did not match how rounding actually happened on the floor. Once that changed, the interface problem changed with it.
The brief is the deliverable
Brief fragment
The earlier the ambiguity, the more expensive it is to carry forward.
A good brief doesn't remove the design work. It makes sure the team is solving the right problem.
FRAMING · PRODUCT BOUNDARIES · ASSUMPTIONS · NON-GOALS · DECISION RECORDS
FULL ESSAY · IN PROGRESS
Enterprise AI is won in the workflow around the model: who reviews what, where the system pauses, what it can access, and what happens when it gets one wrong.
The model is one dependency
Enterprise products also have permissions, policy, exceptions, queues, ownership, audit requirements, and existing systems of record. A better model does not make those disappear.
Model quality determines what is possible. Workflow design determines what is safe and useful enough to deploy.
Swap the model and the product barely moves. Fix the workflow and it ships something people trust.
The questions I look for
The model determines what the system can do. The workflow determines when it should.
PERMISSIONS · APPROVAL · EXCEPTIONS · PROVENANCE · AUDITABILITY · RECOVERY
FULL ESSAY · IN PROGRESS
Compare notes
System acts
Where errors are bounded and recoverable.
Pauses for approval
Where a person sees the evidence before the system commits.
Hands control back
With a clear record of what changed, what triggered it, and why.
Send a note. I reply within two business days.