H&R Block answers tax questions using gen AI

Tax preparation company H&R Block is no stranger to AI and machine learning (ML), having leveraged the technologies across its business for years. But now it’s diving headfirst into gen AI as it sees the potential to transform nearly every aspect of its business, from customer-facing applications to internal functions like engineering, marketing, and legal.

“We’ve had what I’d consider more traditional AI and machine learning for over a decade, with thousands of models we’ve built and applied, mostly in service to tax by analyzing the millions of tax returns we have,” says H&R Block CIO Alan Lowden. “We’ve deployed it through many of our products for use cases from fraud detection, to personalization, to our intelligent tax assistant, which helps identify tax savings for our customers.”

This tax season, H&R Block has worked closely with long-time partner Microsoft to apply gen AI to help 69 million customers do their own taxes. The company has long had a strong partnership with the vendor, which has helped H&R Block shift about 90% of its infrastructure to the public cloud. So last year, when H&R Block placed its bets on gen AI, Lowden says Microsoft was its first choice from a partnership standpoint, but also because of its alliance with, and investment in, OpenAI, the AI research organization behind ChatGPT and Dall-E. Microsoft also selected H&R Block to participate in its AI 100, a group of companies prioritizing the development and deployment of solutions using Azure OpenAI service.

Then in December, H&R Block unveiled AI Tax Assist, a gen AI experience intended to help individuals, the self-employed, and small business owners streamline the process of preparing their own taxes. AI Tax Assist leverages the Azure OpenAI Service to answer any tax-related questions, from information on tax forms to guidance on tax rules and changes to laws.

Lowden explains that AI Tax Assist goes beyond standard online help that requires users to be prescriptive in how they search for information by allowing customers to ask questions using natural language, though, he adds, human help is always available.

“We believe these solutions are best when there’s a human in the loop, which is why wherever we deploy any generative AI solutions, we have options to bail out to humans,” he says.

Proceeding with caution

While H&R Block’s leadership and board were enticed by the possibilities of gen AI, Lowden notes he had to address some concerns before they fully bought into the project, especially with regard to safety and data privacy.

“No one wants to implement AI that has a high risk of bias,” he says. “We were careful to lay out governance and controls around those things. There were also some concerns about accuracy and using a public corpus of data, of course, so we used our own data that’s built in-house.”

Lowden’s team started working on AI Tax Assist in earnest in August last year.

“In this company, if we’re embedding something in core tax, it has to be very well-baked by December,” he notes. “It was a pretty short timeframe from ideation to delivery.”

Given the speed required, Lowden established a specialized team for the project to encourage a culture of experimentation and “moving fast to learn fast.”

“You throw away the structures you have everywhere else; you get a small tiger team with your best talent, lay out clear objectives, and get out of the way,” he says. “In this case, while we have the same roles involved that many of our product teams have, such as product, experience design, engineering, and data science, we worked differently by keeping the team small and isolated from all the operational stuff that gets in the way.”

Ensuring data quality

For the data corpus, Lowden’s team turned to the H&R Block Tax Institute, the company’s tax research arm, which consists of a team of tax attorneys, CPAs, enrolled agents, and more. The Tax Institute studies and analyzes the constantly shifting landscape of federal and state tax laws and publishes articles on how to deal with them.

“One of the challenging things we found was in getting the content right, the source documents to feed the LLM,” Lowden says. “The quality of the content is everything.”

While the Tax Institute provided an invaluable body of data with which to train AI Tax Assist, Lowden says the most important thing he and his team learned from their first big foray into gen AI was about creating the right content management strategy.

“Knowing what I know now, I would’ve focused on source content much earlier in the project,” he says. “It’s incredibly important for domain experts to update their content for accuracy and to develop mature content management processes where information is tagged so it can be more easily managed and governed.”

Three layers of content integrity

Another big part of ensuring the integrity of the content was testing, which consisted of three layers. The first was safety and data privacy testing.

“Microsoft was really helpful in that exercise, using their tools and having their team go beat up on it as their own kind of generative AI penetration test,” Lowden says.

The second layer was about accuracy and content engineering to ensure AI Tax Assist was delivering the most accurate responses possible.

The third was guardrails. “There’s a lot of questions you could ask that we don’t necessarily want to answer, like if it has nothing to do with tax in this case,” Lowden adds.

Through mid-February, Lowden says AI Tax Assist saved the company about 3,400 hours in labor year-over-year based on the reduction in chat volumes. He notes that the company also piloted a customer experience that routes callers through a voice-based gen AI solution, and as of mid-February, that virtual assistant contained 52% of all calls. With fewer than half of calling customers dependent on a live agent to answer their questions, Lowden says the firm saved more than 2,700 hours of workload, freeing up human agents to focus on the customers who did need human help.

Plans in the works

Based on their success with AI Tax Assist, the company has a number of additional gen AI projects in the R&D pipeline.

“We’re looking across engineering, marketing, legal, you name it,” he says. “In every function, they’re going to benefit from these solutions. We’re in the early days but the pilots have been really promising.”

Lowden notes that working with gen AI requires getting comfortable with being uncomfortable.

“These solutions are quite different than traditional product development projects,” he says. “We don’t write all the code. Hallucinations are part of what base models produce, so more thought has to go into test harnesses and building other models to help mitigate the risk. You have to lead the team and be okay with trying different approaches.”

Even compared with disruptive technologies of the past, gen AI has been evolving at a furious pace. Lowden says it’s easy to get discomfited by the idea that anytime you launch a new offering, OpenAI or Azure OpenAI services will release powerful new capabilities. Embracing these technologies requires a willingness to pivot midstream, knowing that such changes bring risk but also potentially high rewards.

“Don’t be afraid,” he adds. “I think it’s really important for any CIO or CTO to lean in because this is the future of work. We’re all going to be using AI assistants of some kind in our jobs. The more we can democratize gen AI and get our teams and employees playing around with it and getting comfortable with the tools, the more they’re going to be inspired and come up with ideas on how to benefit their own work and customers.”