01 — Cargo
The assistant added a shift’s worth of work.
Every AI product pitch includes the same promise: faster notes, smarter summaries, fewer clicks. In care delivery workflows — community health, nursing handoffs, member support — the lived experience is often the opposite. The assistant generates documentation the clinician must verify. The alert fires before context loads. The “helpful” draft is wrong in a way that is expensive to fix because the human is still liable.
That is not a model-quality problem alone. It is a product shape problem: copilots that create cargo.
Cargo is work that exists only because the tool exists — new fields, new review steps, new reconciliation between what the human knew and what the system asserted. Cargo does not show up in a demo. It shows up on the third Tuesday of go-live, when the team stops fighting the model and starts working around it.
02 — Field Research
Field research in week one changed the wedge.
The obvious validation exercise is to test whether users “like” the assistant. The more useful exercise is to watch a shift without mentioning AI — to inventory what already takes time, what already fails, what already gets punted to sticky notes and texts.
In one validation-stage engagement (healthcare AI, incubation context), week-one field contact did not confirm the initial concept. It replaced it. The team had been oriented toward a voice-and-RAG copilot surface; the field said the urgent problem was burden on community health workers — steps between what they already did and what the system recorded. The product direction pivoted toward removal, not addition.
That pivot is the essay’s point: validation is not applause for the demo. It is whether the wedge reduces real work for the role that owns the outcome.
03 — Help
What “help” should mean in consequential systems
Consequential systems — care, security operations, public programs — share a design constraint: wrong automation is worse than slow humans. Help should mean:
- Fewer handoffs, not more summaries of handoffs
- Explicit human ownership at decision points, not implied approval of model output
- Interfaces that show why the system suggested something, not only what it suggested
When those constraints are ignored, copilots become cargo factories. The organization still hires the same number of people; they are just tired in a new way.
04 — Design Moves
Design moves that actually remove cargo
These are not universal rules — they are patterns that showed up when the goal changed from “ship an assistant” to “remove steps”:
Map the workflow before the model — if the diagram needs AI to make sense, the wedge is probably wrong.
Measure burden, not engagement — minutes saved is a vanity metric if verification minutes rose.
Let the first release delete a step — one fewer tap, one fewer duplicate entry; resist the feature that “showcases” the model.
Name the liable role in the UI — who must confirm, who can override, what happens when the model is silent.
05 — Open Bets
Voice modality is still an open bet.
Whether voice is the right primary modality for CHW contexts in all seasons and acoustic environments — still open. Whether community-contributed availability-style patterns (see Greenridge) translate to clinical settings — probably not literally, but the trust mechanics might.