AI for Regulatory Compliance: How Intuit TurboTax Compressed Months of Work into Days

The clock was ticking, and the document was daunting: a 900-page, unstructured behemoth known as the One Big Beautiful Bill (OBBB). With no standardized schema, no official IRS forms available, and a non-negotiable shipping deadline, the TurboTax team faced a critical challenge.

Could artificial intelligence accelerate a months-long implementation cycle into mere days without risking a single calculation error? For Intuit, the answer was not just a “yes,” but a blueprint for the future of AI for regulatory compliance.

What emerged was less of a tax-specific solution and more of a scalable architecture. By blending commercial LLMs with proprietary domain-specific languages and a rigorous testing framework, Intuit has created a template for any development team operating within a highly constrained, regulated environment.

From Chaos to Code: The LLM Assembly Line

Joy Shaw, Intuit’s director of tax, remembers the 2017 Tax Cuts and Jobs Act (TCJA) as a manual marathon. Back then, developers spent months manually tracing law sections and cross-referencing codes to ensure accuracy. There were no shortcuts.

The OBBB presented a higher hurdle. It was structurally more volatile than the TCJA, with the House and Senate versions using disparate language to describe identical provisions. The team had to start building before the government even provided the instructions.

To solve this, Intuit deployed a tiered AI strategy. First, they used general-purpose LLMs, including ChatGPT, to handle the “noisy” part of the process: summarizing versions, reconciling differences between chambers, and filtering provisions to identify exactly which customers were affected.

This initial phase of distillation transformed a process that typically took weeks into one that took hours. But as the work shifted from analysis to implementation, the team hit a ceiling.

Bridging the Gap with Domain-Specific Intelligence

The core of TurboTax does not run on common languages like Python or Java; it utilizes a proprietary domain-specific language (DSL). Because general AI models were never trained on this internal syntax, they could not simply “write” the tax engine’s code.

Intuit turned to Claude to bridge this gap. Claude acted as the translator and dependency mapper, identifying which parts of the existing codebase remained static and which required updates based on the new law.

By isolating these dependencies, developers could ignore the “noise” of the existing millions of lines of code and focus exclusively on the new provisions. This precision prevented the “butterfly effect,” where a change in one tax provision accidentally breaks a calculation elsewhere in the system.

But how does a company maintain a near-zero error threshold when using generative AI? Do you trust the machine, or do you trust the expert?

Engineering Determinism in a Probabilistic World

Generative AI is inherently probabilistic—it guesses the next token. However, tax law requires deterministic results—the same input must always produce the exact same output.

To achieve this, Sarah Aerni, Intuit’s VP of technology for the Consumer Group, emphasized a shift toward verifiably correct architecture. The team built two proprietary tools to ensure shippable quality:

• Auto-Screen Generation: A tool that automatically creates product screens based on law changes, removing the need for manual curation for every single provision.
• Advanced Unit Test Framework: Unlike traditional pass/fail tests, this framework pinpointed the exact code segment causing a failure and provided a generated explanation, allowing for in-context corrections.

Even with these safeguards, the human element remained the final gate. Intuit utilizes LLM-based evaluation tools, but every single output is verified by human tax experts. The AI provides the speed; the humans provide the authority.

Could this hybrid approach eliminate the need for massive compliance teams, or will the “human-in-the-loop” always be the most critical component of the stack?

The RegTech Blueprint: 4 Pillars for Other Industries

While this was a tax problem, the conditions are universal for industries like healthcare, financial services, and legal tech. Any team facing complex regulations and hard deadlines can adopt these four strategies for AI for regulatory compliance:

1. Leverage Commercial LLMs for Analysis

Use general models for the “heavy lifting” of parsing and filtering. This is where speed is gained without introducing significant accuracy risks.

2. Transition to Domain-Aware Tooling

Avoid using general models to write production code in proprietary environments. Use tools that understand your specific syntax and dependencies.

3. Front-Load Evaluation Infrastructure

Build your testing framework before the sprint begins. Move beyond pass/fail results to diagnostic tools that identify the “why” behind a failure.

4. Democratize AI Fluency

AI should not be a tool reserved for the engineering department. Intuit trained staff across all functions, ensuring that tax experts and developers spoke the same AI language.

Is your organization treating AI as a niche engineering tool, or as a cross-functional operational standard?