Put human checkpoints in orchestration design
A 73% deflection slide hides the 27% you still owe an explanation for—and whether any of the 73% should have stayed with a person.
Philip Morgan
June 9, 2026 · 6 min read
The vendor slide shows 73% deflection. The board nods. Nobody asks what happened to the 27%, or whether the 73% included cases that should never have been automated. Deflection trap metrics reward the wrong optimization. Orchestration strategy is graded autonomy: explicit human checkpoints per step, tuned to consequence.
Agents are workflows with marketing budgets
"Agent" pitches hide choreography cost: twelve steps, three reviews, two systems, one timeout nobody tested. Strategy names the truth: orchestration is the choreography connecting models, tools, knowledge, and people. If you cannot draw the workflow, you are buying a slide deck.
Healthcare prior auth is the textbook case. Gather documentation, verify coverage, draft submission, human approve, submit, track, escalate denial. The executive decision is which steps require a human before the workflow continues.
Checkpoint types executives should recognize
| Checkpoint | When |
|---|---|
| Approve | irreversible or regulated action (submit auth, send patient message) |
| Edit | draft acceptable with human refinement (note, letter) |
| Takeover | model uncertainty or policy edge case |
| Observe | sample for quality; no user friction |
High-stakes clinical adjacency gets Approve or Takeover. Internal summarization might get Edit with sampling. Patient-facing autonomous write without a checkpoint belongs in Legal review before engineering schedules it.
Planner vs explicit states
Delegates ask: LLM planner or state machine? For regulated flows, explicit states win audit. The model can reason *inside* a state; it should not decide *which* state comes next in production without guardrails. Flexibility lives in language inside each state; control flow stays explicit.
Predictable handoffs when the machine is wrong matter more than one model for every step. The machine will be wrong in interesting ways.
The instant answer counterargument
Patients want correct. Instant wrong answers in healthcare travel farther than delayed right ones. Orchestration strategy trades fake instant for reliable path to human, with queue depth metrics executives can watch on Mondays.
Downstream:
- Interface shows humans the checkpoint
- Gateway logs each transition with correlation IDs
- Training measures override rates
Operations needs queue depth and override rates on Monday morning instead of a deck that rebrands workflows as agents. Name the checkpoints first; marketing language second.