AI Automation that actually works: $100M, messy data, zero surprises - Tanmai Gopal, Hasura/PromptQL
Problem Context: Healthcare Appointment Scheduling 00:00
The project involved automating appointment scheduling at a large public healthcare company serving radiologists and clinics.
Calls from patients to schedule appointments typically last 12-15 minutes; reducing call time by even 3 minutes can save $50 million, increase appointment volumes, and cut operator training costs.
The operator's workflow is highly complex, involving navigation through multi-tab UIs and figuring out the right medical procedure code based on multiple, dynamic factors.
Variables influencing procedure codes include patient demographics, clinic history, local and federal regulations, clinic preferences, and inconsistent code lists across clinics.
No universal set of procedure codes exists; clinics vary widely (some have 5, some 250 codes for the same procedure type).
Encoding every business rule as configuration leads to exponential config complexity and heavy operator training burdens.
Non-technical administrators understand the business rules but can't code; developers can code but don't know the rules, resulting in an "automation paradox."
Training and maintaining automation logic via developer intervention is expensive and unsustainable for the business.
AI-Powered Solution: Natural Language Automation 07:48
The central idea is enabling non-technical staff to write and modify business rules in natural language, translated by AI to executable logic.
The focus is on removing developers from the automation loop, allowing admins to "vibe code in production."
Major challenges addressed include translating business-user language to code-friendly logic, defining software development lifecycle (SDLC) processes for non-technical users, and ensuring security when business users can deploy logic directly.
A new approach involves creating a domain-specific intermediate language ("company QL" or "acmeql") that represents deterministically executable plans.
The AI model is trained on the specific domain's semantics, entities, and rules to generate this language from natural business instructions.
A demo showed how business logic for issue assignment in GitHub (an analog to the healthcare use case) can be created, tested, and deployed by a non-technical user using natural instructions.
Users iteratively add business rules, test logic live, and make corrections without direct coding, simply specifying inputs and expected outputs.
Admins can deploy new automation logic simply by pressing a button; the system handles underlying complexities.
The platform maintains strict data layer boundaries to prevent business user logic from creating security breaches or cross-tenant data access.
All user-generated plans run in user space, not in the sensitive data layer.
The anticipated impact includes over $100 million in realized value from procedural code and appointment selection automation in the healthcare company.
The speaker concludes that the industry should move from developers building apps to building domain-specific "vibe coding" platforms tailored to each organization.
PromptQL invites participants to learn more about their solutions at their booth.