PharmacyPlus — AI operations for a retail pharmacy

PharmacyPlus (Imagine Health Pharmacy & Travel Clinic) is a working retail pharmacy in southeast Calgary with about six employees. It’s also the hardest test of my AI-operations model, because pharmacies are regulated, handle protected health information, and cannot tolerate a chatbot improvising about medicine.

The stack

• Manager — seven AI agents on cron covering the back office: staff scheduling, payroll and timesheets, sales and finance analysis, invoicing, and marketing insights, with decisions logged to Airtable
• admin.pharmacyplus.ca — a Next.js 16 + Supabase back-office console where the human work happens: schedules, timesheets, pay runs, invoices
• Storefront — the pharmacyplus.ca Shopify storefront with behavioral analytics
• Watchdog — a VPS monitor with Twilio SMS escalation; if an agent goes silent or a system goes down, my phone knows before the staff does

Regulated-industry guardrails

The autonomy framework here is stricter than my e-commerce version, and the differences are the point:

• PHI never enters agent context. Agents work on operational data — shifts, hours, sales aggregates — never patient information.
• Pharmacist override on anything clinical. No agent writes health claims or patient-facing clinical content. The pharmacist of record owns that lane entirely.
• Same audit spine — tiered approvals and a decisions log with the “no log, no action” rule.

Why this one matters

Anyone can demo an AI agent on a toy problem. Running seven of them against a regulated business with real employees, real payroll, and a college of pharmacists looking over your shoulder forces the discipline that makes AI operations actually deployable: scoped data access, human override lanes, and audit trails by default.