When the One Large Stunning Invoice arrived as a 900-page unstructured doc — with no standardized schema, no printed IRS types, and a tough transport deadline — Intuit's TurboTax workforce had a query: might AI compress a months-long implementation into days with out sacrificing accuracy?
What they constructed to do it’s much less a tax story than a template, a workflow combining business AI instruments, a proprietary domain-specific language and a customized unit take a look at framework that any domain-constrained improvement workforce can be taught from.
Pleasure Shaw, director of tax at Intuit, has spent greater than 30 years on the firm and lived by way of each the Tax Cuts and Jobs Act and the OBBB. "There was numerous noise within the legislation itself and we have been in a position to pull out the tax implications, slender it right down to the person tax provisions, slender it right down to our clients," Shaw instructed VentureBeat. "That sort of distillation was actually quick utilizing the instruments, after which enabled us to start out coding even earlier than we received types and directions in."
How the OBBB raised the bar
When the Tax Cuts and Jobs Act handed in 2017, the TurboTax workforce labored by way of the laws with out AI help. It took months, and the accuracy necessities left no room for shortcuts.
"We used to should undergo the legislation and we'd code sections that reference different legislation code sections and try to determine it out on our personal," Shaw mentioned.
The OBBB arrived with the identical accuracy necessities however a distinct profile. At 900-plus pages, it was structurally extra complicated than the TCJA. It got here as an unstructured doc with no standardized schema. The Home and Senate variations used completely different language to explain the identical provisions. And the workforce needed to start implementation earlier than the IRS had printed official types or directions.
The query was whether or not AI instruments might compress the timeline with out compromising the output. The reply required a particular sequence and tooling that didn’t exist but.
From unstructured doc to domain-specific code
The OBBB was nonetheless shifting by way of Congress when the TurboTax workforce started engaged on it. Utilizing massive language fashions, the workforce summarized the Home model, then the Senate model after which reconciled the variations. Each chambers referenced the identical underlying tax code sections, a constant anchor level that allow the fashions draw comparisons throughout structurally inconsistent paperwork.
By signing day, the workforce had already filtered provisions to these affecting TurboTax clients, narrowed to particular tax conditions and buyer profiles. Parsing, reconciliation and provision filtering moved from weeks to hours.
These duties have been dealt with by ChatGPT and general-purpose LLMs. However these instruments hit a tough restrict when the work shifted from evaluation to implementation. TurboTax doesn’t run on a regular programming language. Its tax calculation engine is constructed on a proprietary domain-specific language maintained internally at Intuit. Any mannequin producing code for that codebase has to translate authorized textual content into syntax it was by no means educated on, and determine how new provisions work together with a long time of present code with out breaking what already works.
Claude turned the first instrument for that translation and dependency-mapping work. Shaw mentioned it might determine what modified and what didn’t, letting builders focus solely on the brand new provisions.
"It's in a position to combine with the issues that don't change and determine the dependencies on what did change," she mentioned. "That sped up the method of improvement and enabled us to focus solely on these issues that did change."
Constructing tooling matched to a near-zero error threshold
Common-purpose LLMs received the workforce to working code. Getting that code to shippable required two proprietary instruments constructed throughout the OBBB cycle.
The primary auto-generated TurboTax product screens straight from the legislation modifications. Beforehand, builders curated these screens individually for every provision. The brand new instrument dealt with the bulk mechanically, with guide customization solely the place wanted.
The second was a purpose-built unit take a look at framework. Intuit had all the time run automated checks, however the earlier system produced solely go/fail outcomes. When a take a look at failed, builders needed to manually open the underlying tax return information file to hint the trigger.
"The automation would let you know go, fail, you would need to dig into the precise tax information file to see what might need been flawed," Shaw mentioned. The brand new framework identifies the precise code phase accountable, generates an evidence and permits the correction to be made contained in the framework itself.
Shaw mentioned accuracy for a shopper tax product must be near 100%. Sarah Aerni, Intuit's VP of know-how for the Shopper Group, mentioned the structure has to provide deterministic outcomes.
"Having the kinds of capabilities round determinism and verifiably right by way of checks — that's what results in that kind of confidence," Aerni mentioned.
The tooling handles the pace. However Intuit additionally makes use of LLM-based analysis instruments to validate AI-generated output, and even these require a human tax knowledgeable to evaluate whether or not the result’s right. "It comes right down to having human experience to have the ability to validate and confirm absolutely anything," Aerni mentioned.
4 elements any regulated-industry workforce can use
The OBBB was a tax drawback, however the underlying situations are usually not distinctive to tax. Healthcare, monetary companies, authorized tech and authorities contracting groups frequently face the identical mixture: complicated regulatory paperwork, laborious deadlines, proprietary codebases, and near-zero error tolerance.
Based mostly on Intuit's implementation, 4 parts of the workflow are transferable to different domain-constrained improvement environments:
Use business LLMs for doc evaluation. Common-purpose fashions deal with parsing, reconciliation and provision filtering properly. That’s the place they add pace with out creating accuracy danger.
Shift to domain-aware tooling when evaluation turns into implementation. Common-purpose fashions producing code right into a proprietary surroundings with out understanding it can produce output that can not be trusted at scale.
Construct analysis infrastructure earlier than the deadline, not throughout the dash. Generic automated testing produces go/fail outputs. Area-specific take a look at tooling that identifies failures and allows in-context fixes is what makes AI-generated code shippable.
Deploy AI instruments throughout the entire group, not simply engineering. Shaw mentioned Intuit educated and monitored utilization throughout all features. AI fluency was distributed throughout the group relatively than concentrated in early adopters.
"We proceed to lean into the AI and human intelligence alternative right here, in order that our clients get what they want out of the experiences that we construct," Aerni mentioned.
