Concept / Independent product exploration
AI Operations Copilot for Digital Pharmacy
Use an LLM on unstructured operational exceptions. Keep clinical decisions with licensed humans.
Independent exploration. Not deployed at Epiq Scripts. Not a production feature. I have not shipped production AI.
At a glance
- Problem
- Operators reconstruct context, classify exceptions, route work, and write summaries before anyone decides. That coordination is slow and uneven.
- Why AI
- Exception language is unstructured. Rules engines own known, stable conditions. An LLM is only useful where input is messy and output can be constrained.
- Constraint
- No prescribing, no prescription changes, no unsupervised clinical advice, no replacement of pharmacist verification, no autonomous controlled-substance approval.
- System
- Structured suggestion beside the queue: type, owner, confidence, summary, next step. Human accepts, edits, or rejects. Low confidence goes to a human queue, never a silent write.
01
Problem
Insurance failures, fulfillment issues, incomplete intake, shipping problems, and partner-system noise hit operations as mixed text and partial state.
Users are operations staff and technicians on non-clinical exceptions. Pharmacist verification stays a separate, accountable workflow.
02
Why AI — and why not only rules
Deterministic automation should own status transitions, required fields, SLA timers, and hard routing rules.
LLMs belong where the job is interpretation: classify, summarize, suggest a route, draft an internal note, cluster recurring patterns. They do not belong on the write-path without a human gate.
03
System, HITL, and structured output
The copilot reads only the operational fields required for the task. It returns exception type, suggested owner, confidence, short summary, and recommended next step.
Free-form generation is limited to draft internal notes. Never prescription content. Nothing commits to the system of record without an explicit human action.
04
Safety constraints
Clinical independence is a product rule, not a model-tuning preference.
- Must not prescribe or change medications
- Must not give unsupervised clinical recommendations
- Must not replace pharmacist verification
- Must not autonomously approve controlled-substance activity
05
PHI, privacy, and access
Minimize what the model sees. Prefer operational identifiers over clinical narrative when either would suffice.
Role-based access. Treat prompts, retrieved context, and outputs as sensitive. Retain only what audit and evaluation require.
06
Failure modes and audit
Hallucinated context, wrong exception type, overconfident routing, stale state, omitted safety-relevant detail.
Log model/prompt version, input field set, structured output, confidence, human action, and committed state. If routing fails later, the product must show whether AI suggested it and whether a human overrode it.
Response to failure is a lower confidence threshold and forced review — not a longer generated paragraph.
07
Evaluation, metrics, rollout
Offline labeled historical exceptions, structured-output validity checks, then shadow mode against human routing — before any write-path.
- Suggestion acceptance rate
- Routing accuracy and hallucination/error rate
- Time-to-resolution and time saved
- Escalation rate and share requiring correction
- Shadow mode → narrow assist mode → expand only if metrics hold
- Prompt/version control, confidence thresholds, and a kill switch as launch requirements