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

The central challenge in many enterprise automation initiatives is what can be described as the “Automation Paradox”: the domain experts who deeply understand the business rules cannot code the automation, while the developers who can code lack the nuanced, often unwritten, domain knowledge. This disconnect leads to brittle, configuration-heavy systems and offloads significant cognitive burden onto front-line operators, ultimately limiting the potential for automation.

The High-Stakes Problem: Healthcare Appointment Scheduling

In a large public healthcare company specializing in radiology, the process for a patient to schedule an appointment is a prime example of this challenge. An operator on a call with a patient takes 12-15 minutes to navigate a labyrinthine enterprise UI, performing complex data entry across multiple tabs. The core task is to identify the correct medical procedure code.

This is far from straightforward. The correct code depends on a vast number of permutations:

• Patient Data: Symptoms, age, gender, and prior clinical history.
• Regulations: A mix of state, federal, and local rules.
• Clinic-Specific Logic: Individual clinic preferences, such as not operating after 3 PM on certain days.
• Lack of Standardization: Different clinic groups use entirely different sets of procedure codes. One group might have five codes for a mammogram, while another has 250.

The result is an explosion of edge cases and uncoded business logic. The cost of encoding each new rule into the software outweighs the benefit, so the burden is shifted to training human operators, a costly and error-prone solution. Shaving just three minutes off each call is estimated to have a $50 million annual impact, highlighting the immense value locked within this manual process.

The AI Solution: “Vibe Coding” for Domain Experts

The proposed solution is to use AI to bridge the gap in the Automation Paradox. The goal is to enable non-technical administrators—the ones who know the rules—to “vibe code in production,” writing and updating complex business logic using natural language.

However, implementing this presents three critical challenges:

1. The Language Problem: A standard LLM does not speak the nuanced language of the business domain. It understands programming languages (React, Rust, etc.), but it cannot reliably interpret domain-specific intent without significant guidance, risking serious errors.
2. The DevOps Problem: A new software development life cycle (SDLC) is needed for non-technical users. Concepts like review, staging, production, and troubleshooting must be re-imagined for users who are not developers.
3. The Security Problem: Giving non-technical users the power to write business logic on the fly introduces significant risk. A poorly constructed rule could inadvertently cause a massive data breach or security leak.

From Non-Determinism to Reliability with a DSL

The architectural solution moves away from having developers use foundation models to generate code. Instead, it proposes a new paradigm:

1. A non-technical business user interacts with a model that has been specifically taught the language of their domain. This involves encoding the domain’s procedural semantics, ontologies, and entities into the model.
2. This specialized model translates the user’s natural language intent into a deterministic plan written in a domain-specific language (DSL), referred to as AcmeQL (a stand-in for Hasura’s PromptQL).
3. This AcmeQL plan is a safe, deterministic artifact that can be reliably executed by the system.

This approach creates a predictable bridge between the non-deterministic nature of AI and the reliability required for business-critical processes.

A demo illustrated this workflow using the analogous problem of assigning GitHub issues. A business user could:

• Start with a simple prompt: Given an issue description, find the most relevant file using AI and find the top contributor.
• Verify the result and convert the logic into a reusable “automation.”
• Iteratively test and refine the automation by adding more rules in natural language, such as Tom is from a different company, so I can't assign this to Tom or In case no relevant files are found, assign to a default person.
• Deploy the finished automation with a single click.

Crucially, the security is managed at the data layer. The AcmeQL plan runs in a sandboxed “user space,” preventing it from violating multi-tenant authorization rules, which are enforced by the underlying data platform.

By empowering domain experts to directly build and manage automation, this model promises to unlock significant value. For the healthcare partner, this system is on track to realize over $100M in impact by transforming how they handle procedure code selection and appointment scheduling. The ultimate vision is to enable organizations to build their own unique “vibe coding platforms,” turning their domain experts into the builders of their own software.

• Tags:
• Generative AI
• Feature
• Enterprise AI
• AI Automation
• Domain-Specific Language
• PromptQL
• Healthcare Technology
• Operational Efficiency