Edge cases show where AI ends and human judgment begins: Intuit’s Prashanth Seshadri

AI is rapidly changing financial services, cutting hours of manual work into minutes and reshaping how companies manage money. Financial software company Intuit is betting on AI to transform tax filing and bookkeeping. Known for TurboTax and QuickBooks, the company is integrating AI into its products, focusing on practical tasks like automating invoices, streamlining tax returns, and improving cash flow.

In a conversation with TechCircle, Prashanth Seshadri, Director of Business Intelligence at Intuit, discusses the company’s AI strategy, the role of generative models, and why human expertise remains central to automation.

Edited Excerpts: 

Intuit says AI is driving its recent growth. What is the overall vision behind the company’s AI strategy?

Intuit has been working with AI for more than six years, well before generative AI and large language models became widely used. We have built AI models across multiple product lines and focused on applying them to practical use cases.

Our current focus is on "done-for-you" experiences, where we use agentic AI to handle end-to-end tasks that customers would otherwise do manually. These tasks, which might normally take five to six hours, can now be completed in five to seven minutes. Customers still review and validate AI suggestions before finalizing, ensuring the results fit their needs. This approach reduces manual effort and allows customers to spend more time running their business instead of managing tasks like bookkeeping.

What factors made your tax return innovation possible, and how did it reduce the time customers spend on the process?

TurboTax is widely used in the US and Canada for filing taxes. When people begin filing, they must update all financial information for the year. If TurboTax is connected to accounts such as payroll or banking, the software already knows income details from salary and previous filings. It can then run tax calculations automatically, unless major life events, such as marriage or having a child, require additional updates.

This process replaces manual, time-consuming work. Using past data and patterns from similar taxpayers, the system can file taxes efficiently while aiming to maximize the user’s tax benefit. Users only need to enter changes from the year, such as new income or buying a house, while the rest is calculated automatically.

QuickBooks applies the same approach to financial management. Its automated, AI-powered workflows now handle accounts receivable (AR) and accounts payable (AP). For example, when a company issues invoices with payment due in 30 or 45 days, many customers either forget or delay payment. AI agents track patterns in customer behavior, some pay early, some on time, others consistently late. The system adjusts reminders accordingly, deciding when to send them, in what form, and with what message. This improves cash flow by 10% to 50%.

How do you measure success in AI-driven experiences beyond time savings, such as through trust, satisfaction, or accuracy?

We continuously track precision and recall numbers, along with financial outcomes tied to customer use of our products. For example, we look at whether a customer’s monthly cash flow increases after adopting a product. If a customer was generating $100 per month and that amount grows to $120, we capture that improvement as part of our measurement framework.

We use several metrics like this to evaluate customer benefit. The starting point is always the customer outcome we want to achieve, not the technology itself. If we see that a customer could benefit from an increase in working capital or monthly cash flow, we may set a goal of improving it by 50%. Once that measure is clearly defined, we then determine which technology or solution can help us reach it.

This order is important. We do not create technology in isolation and then try to match it to a customer problem afterward. Instead, the customer business metric comes first, and technology is selected or developed only once we understand the specific value it needs to deliver.

How does generative AI improve decision-making compared to traditional AI or machine learning in your company?

When we think about decision-making, we often end up providing recommendations. These recommendations can take different forms: they might be driven by rules, by traditional AI systems, or by large language models (LLMs). In some way, these kinds of systems have existed for a while.

What has changed is the role of LLMs. Their introduction has accelerated the ways in which we can generate and deliver recommendations to customers. Take a simple example: some customers always pay on time. In those cases, there is no need to send them reminders. On the other hand, there are customers who consistently pay late. For them, reminders are necessary.

The question then becomes: how do we arrive at those decisions? That is where the models come in. They operate in the background, collecting and analyzing data from across the installed base. From this data, they determine which customers tend to pay on time, which customers usually pay late, and what steps might encourage late-paying customers to make payments earlier.

This is the foundation on which we have built our models. The arrival of LLMs makes this process easier. Instead of relying only on traditional mathematical or statistical models, LLMs allow us to handle these decision-making tasks more directly and with less effort in building specialized models from scratch.

Financial data is complex. From both an industry and your company’s perspective, what are the main technical challenges in training and deploying AI models on this type of data?

We have built and use our own platforms to manage data and model development. These include GenOS, a data platform, and a governance platform. Each of these is designed to ensure that privacy, security, and legal requirements are addressed at every stage. By operating on this structured foundation, we can develop models without having to worry about data breaches, data security, or compliance issues, because the necessary safeguards are built into the platform itself.

This approach means that even when working with sensitive information, such as financial data or personal identifiable data, developers and other users who should not have access to it never see it. The system is structured so that the layers of complexity around privacy, responsible AI, and legal compliance are managed within the platform. As a result, the work of building and deploying models happens on a secure, paved path where risks related to data handling are contained by design.

Where does generative AI offer a clear advantage over traditional AI in financial workflows?

AI is no longer limited to financial workflows. It applies to every workflow. In the past, building AI models required experts with advanced degrees in mathematics, computer science, or statistics to design and train them. With generative AI and large language models, this has changed. The technology has become accessible. For example, you can upload a recording and ask the system to generate an article. Within minutes, you have a draft. By refining prompts, you can shape the output to highlight specific points. This is not limited to accounting or finance. The shift requires us to adopt these tools to achieve faster turnaround times than before.

When it comes to automation, there are often edge cases. How do you manage situations where AI may struggle and human review is needed?

Intuit is aware that large language models can generate incorrect results, so the company has built guardrails to limit errors as much as possible. At the same time, human expertise is built into the process. At any stage in a workflow, users can bring in a human expert with a single click. In TurboTax this service is called TurboTax Live, and in QuickBooks it is called QuickBooks Live.

For example, when a chart of accounts needs to be merged, the AI system can recommend which accounts should be combined. The recommendation is presented to the user, but if the user is unsure or does not want to rely only on the AI, they can request a human expert to join the session and confirm or adjust the recommendation.

This creates an interaction where an AI agent can handle much of the work, but a human can step in whenever needed, and the reverse can happen as well. Over time, as more automation is introduced, this combined approach of artificial intelligence and human intelligence will continue to be developed. The intention is to create hybrid models that maintain accuracy and build confidence in the system.