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

Key Concepts:

• Automation Paradox: The disconnect between those who understand business rules (non-technical users) and those who can code automation (developers).
• AcmeQL (or Company QL): A domain-specific language (DSL) generated by an AI model, tailored to a specific company's business logic, enabling non-technical users to define and update algorithms.
• Language Problem: The challenge of translating business users' natural language into a format that AI models can understand and execute correctly.
• Vibe Coding: A term used to describe non-technical users writing and updating algorithms in natural language.
• Deterministic Artifact: The AcmeQL plan is a program that can be executed predictably and reliably.

1. The Problem: Complex Appointment Scheduling in Healthcare

• A large public healthcare company specializing in radiology software faces challenges with appointment scheduling.
• Operators spend 12-15 minutes per call, determining the correct procedure code based on patient information (age, gender, symptoms, insurance), clinic rules, and regulations.
• Reducing call time by 3 minutes can result in a $50 million impact due to increased call volume and reduced training costs.
• The UI for scheduling is described as extremely complex, with numerous tabs and data entry points.
• Determining the correct procedure code is complicated by factors like patient history, state/federal/local regulations, and clinic-specific rules (e.g., no appointments after 3 PM).
• Different clinics use different sets of procedure codes, making standardization difficult. Some have 250 codes for mammograms, while others have only five.

2. The Current System and Its Limitations

• The current system involves three key players: operators, developers, and administrators.
• Developers build complex software to handle various edge cases, often converting rules into configurations.
• Administrators (non-technical) possess the knowledge of clinic-specific rules but lack the ability to implement them directly.
• This leads to a "config explosion" and a training burden on operators, who must learn to navigate the complex configurations.
• Many business rules remain uncoded because the cost of encoding them outweighs the benefits.
• This situation results in the "automation paradox": those who understand the rules can't code the automation, and those who can code the automation don't understand or want to deal with the rules.

3. The AI-Powered Solution: Empowering Non-Technical Users

• The proposed solution involves enabling non-technical users to write and update algorithms in natural language, effectively cutting out the developer middleman.
• The goal is to allow administrators to "vibe code" in production.
• The solution addresses three key challenges: the language problem, the DevOps problem, and the security problem.

4. Addressing the Language Problem with Domain-Specific AI

• Instead of using a general-purpose LLM, the solution uses a model trained on the specific language of the healthcare domain.
• This model generates "AcmeQL" (or "Company QL"), a domain-specific language that represents a deterministic plan for execution.
• The hard part is encoding procedural semantics, ontologies, entities, and specifics into the model so it can generate code that makes sense to business users.

5. Demo: Automating GitHub Issue Assignment (Analogous Example)

• A demo is presented using a GitHub issue assignment scenario to illustrate the concept.
• A business user starts by providing a natural language description of the desired logic: "Given an issue description like 'data pipelines are not working,' find the most relevant file using AI and then find the top contributor."
• The system identifies the relevant file (e.g., "analytics_pipeline.py") and the top contributor.
• The user can then convert this interaction into an "automation" by specifying the input (description) and output (name) fields.
• The system runs tests, fixes errors, and suggests a user for assignment.
• The user can test the automation with more inputs and outputs, refine the rules (e.g., exclude external contributors), and deploy the automation with a single click.

6. Security and Deployment

• The data layer remains secure, with multi-tenant authorization rules enforced.
• The AcmeQL plan runs strictly in user space, preventing unauthorized data access.
• The impact of this approach on procedure code selection and appointment scheduling is estimated at $100 million or more.

7. The Future: Vibe Coding Platforms

• The speaker believes the future lies in building "vibe coding platforms" tailored to specific organizations, rather than relying solely on developers to build software.
• These platforms will empower non-technical users to directly contribute to and maintain business logic.

8. Notable Quotes:

• "The automation paradox is the people who understand the rules can't code the automation and the people who can code the automation uh can't understand the rules."
• "Instead of developers building software I think uh I think we need to start building the vibe coding platforms that are unique to our organization."

9. Conclusion:

The presentation argues for a shift towards empowering non-technical users to directly participate in automation by creating domain-specific AI models and "vibe coding platforms." This approach addresses the "automation paradox," reduces development bottlenecks, and unlocks significant business value by enabling faster iteration and more accurate representation of complex business rules. The key is to abstract away the technical complexities and provide a secure, user-friendly environment where business users can express their logic in natural language and deploy it with confidence.