Mind map for YouTube video: AI Automation that actually works: $100M, messy data, zero surprises - Tanmai Gopal, Hasura/PromptQL

mindmap_WnTq5Mc5bIU.md

AI Automation that actually works: $100M, messy data, zero surprises - Tanmai Gopal, Hasura/PromptQL

TL;DR This presentation introduces an AI automation solution, PromptQL, designed to tackle the complex and costly problem of appointment scheduling and procedure code selection within a large public healthcare company. Currently, call center operators spend 12-15 minutes per call navigating a nightmarish UI and deciphering intricate, often inconsistent, medical codes and regulations. This inefficiency costs millions, with every 3 minutes saved equating to a $50 million impact. The core issue is an "automation paradox": business users understand the rules but can't code, while developers can code but don't grasp the nuanced business logic. PromptQL proposes enabling non-technical administrators to "vibe code" business logic in natural language. The solution addresses key challenges like translating business language to AI intent, defining SDLC for non-technical users, and ensuring security. It works by training a domain-specific AI model to generate deterministic "company QL" plans that can be tested and deployed by business users, bypassing developers. This approach has already demonstrated a $100M+ impact by streamlining procedure code and appointment selections.

---

Information Mind Map

🧠 AI Automation for Complex Business Logic: PromptQL's Approach

🔑 The Problem: Healthcare Appointment Scheduling & Procedure Codes

• Context: Large public healthcare company specializing in software for radiologists/clinics.
• Current Process: Patients call operators to schedule appointments.
  • Operators collect patient data (age, gender, symptoms, insurance, vendor).
  • Operators must determine the correct procedure code.
  • Operators perform extensive data entry and schedule appointments.
• Inefficiency & Cost:
  • Call duration: 12-15 minutes.
  • Financial Impact: Every 3 minutes reduced saves $50 million.
    • Increased call volume (more appointments scheduled).
    • Reduced capital cost (less operator training/retraining).
• Complexity for Operators:
  • UI Nightmare: "Worst most complicated enterprise UI screen" with 15+ tabs requiring constant switching and data entry.
  • Procedure Code Selection:
    • Highly complex and contextual.
    • Dependent on: patient symptoms, age, gender, medical history (previous visits).
    • Influenced by: State, Federal, and Local regulations.
    • Impacted by: Specific clinic rules (e.g., "clinic doesn't work after 3 PM").
    • Explosive Permutations: Number of combinations is "explosive."
    • Inconsistent Codes: No universal agreement on procedure codes across clinics.
      • Example: Some clinics have 250 codes for mammograms, others 5.

🔑 Organizational Challenges: The "Automation Paradox"

• Key Players:
  • Operator: Takes calls, performs data entry, determines codes. (Jobs targeted for replacement)
  • Developer: Builds complex software to handle edge cases via configurations (e.g., timing block).
  • Administrator (Non-technical): Knows all clinic rules but cannot code.
• Resulting Problems:
  • Configuration Explosion: Developers encode every edge case into configurations.
  • Training Burden: Increased configuration requires more operator training.
  • Uncoded Business Logic: Many rules remain uncoded (e.g., "clinic in Chicago doesn't like to work on a Friday").
  • Negative ROI for New Rules: Encoding new rules becomes more expensive than the benefit derived, leading to offloading training to operators.
• The Automation Paradox:
  • People who understand the rules cannot code the automation.
  • People who can code the automation cannot/will not understand the rules.

🔑 The AI Idea: Natural Language Automation for Non-Technical Users

• Core Proposal: What if non-technical people could write and update algorithms in natural language?
• Goal: Cut developers out of the loop; enable admins to "vibe code in production."

🔑 Key Challenges for AI Implementation

• 1. The Language Problem:
  • Business users have their own obvious domain language.
  • Stock LLMs speak programming languages (React, Rust, JavaScript, TypeScript), not specific business terminology.
  • Translating business algorithms into AI intent is challenging.
  • Risk: Misinterpretation could lead to severe consequences (e.g., "MRI machines catching fire" - hyperbole, but highlights the risk).
• 2. The DevOps Problem:
  • Defining SDLC (Software Development Life Cycle) for non-technical users is complex.
  • Concepts like review, staging, production, fixing, troubleshooting are alien to non-technical users.
• 3. The Security Problem:
  • Allowing non-technical users to write business logic on the fly poses a massive data breach or security leak risk.
  • Requires opinionated solutions to prevent system shutdown.

🔑 PromptQL's Solution Architecture

• Current Developer Workflow: Developers + Tribal Knowledge -> Foundation Model + Tooling -> Programs.
• PromptQL's Proposed Workflow:
  • Non-technical user speaks to a domain-specific model.
  • This model is "taught the language of your domain."
  • Generates a "plan" in a language called company QL (revealed as PromptQL).
  • PromptQL plan is a deterministic artifact (program).
  • This artifact is then programmatically executed.
• Hard Part: Encoding domain-specific practices (procedural semantics, ontologies, entities, specifics) into the model so it generates business-sensible outputs.

🔑 Demo Walkthrough (GitHub Issue Assignment Example)

• Objective: Dynamically reassign GitHub issue supporters based on business rules.
• Steps for Business User:
  1. Initial Request: "Given an issue description like 'data pipelines are not working', find the most relevant file using AI, then find the top contributor."
  2. System Response: Identifies analytics_pipeline.py and a top contributor.
  3. Introduce automations Primitive: Convert the logic into an automation.
     • Define input (e.g., description) and output (e.g., name).
     • System converts, runs tests, fixes issues in background.
  4. Testing & Iteration:
     • Test with more inputs (e.g., "database is down," "pods are not scaling down").
     • Refinement: If results are unsatisfactory (e.g., "Tom doesn't sound like the right guy"), modify rules.
       • Example: "Show me all users and their emails." -> Discover Tom is from an external company.
       • Add rule: "Remove somebody else from an external company."
  5. Deployment: Once satisfied, hit "deploy" button (non-technical action).
• Adding More Complex Rules:
  • Example: "In case there's a generic request, assign to a default person." (Exclusion rule)
  • Example: "In case no relevant files are found, assign to this person."
• User Confidence: Business user gains confidence that logic works, system guarantees execution.

🔑 Security & Impact

• Security:
  • The PromptQL plan runs strictly in user space, not data space.
  • Data layer keeps it "real" and handles multi-tenant authorization rules.
  • Allows "vibe coding" on higher layers without compromising data.
• Impact:
  • The solution for procedure code selection and appointment selection has a $100 million+ impact for the healthcare company, already being realized.

🔑 Future Vision

• Shift in Development: Instead of developers building software, the future is building "vibe coding platforms" unique to each organization.
• Key Takeaway: Empowering non-technical users to directly define and deploy business logic is the next frontier in automation.