Why Deloitte Built a Tax AI That Knows When to Say ‘I Don’t Know’

For decades, tax research has followed a stubbornly familiar rhythm. A complex question lands on a consultant’s desk. They open case law databases, dig through circulars, scan judgments, draft a view, send it up the chain, wait for reviews, revise, and finally respond. Roughly 10-20 hours later, a carefully worded answer goes out. 

The process worked, but it never really changed. That lag is what Deloitte decided to challenge.

Earlier this month, Deloitte India launched Tax Pragya, an agentic AI platform for tax research and summarisation that aims to do something radical in a conservative profession. It compresses hours of deep technical tax research into minutes, without compromising trust in accuracy, sources or judgement.

Trained on more than 1.2 million tax cases and over 5,000 Deloitte technical papers, the platform spans income tax, GST and transfer pricing on a single interface.

Building it was no small task. The team spent 18 months building it, following several rounds of internal tests and experiments with hundreds of Deloitte’s own expert employees. 

Sumit Singhania, partner and chief strategy and digital transformation officer, tax at Deloitte South Asia, told AIM that the idea essentially stemmed from internal conversations he had with teams and many of Deloitte’s clients. It was revealed that deep technical tax research is still being done the same way as 10 years ago.

“In the time that we live today with the advancement of technology, we cannot spend the same amount of hours doing technical research,” he said. The ambition was not incremental improvement. Deloitte wanted to compress research time from hours into minutes without compromising accuracy. 

That constraint shaped every design choice.

Under the hood, Tax Pragya runs on a private LLM enriched with a massive, tightly curated knowledge base. Public jurisprudence data from over a million cases is fused with more than 5,000 Deloitte-authored technical papers and internal knowledge assets.

“The intersection of the public database and the Deloitte knowledge base brings a very powerful source for research,” Singhania said. “By adding the power of generative AI to this, the speed has been made such a fast and agile outcome.”

On a similar note, EY India, in collaboration with Taxmann, launched Taxmann.AI, an AI-powered platform tailored for tax and legal professionals, which also offers an AI bot for responses. PwC also features Navigate Tax hub, a generative-AI-based tax colleague, aimed at automating tax work for firms.

Accuracy, Not Novelty, was the Real Challenge

Deloitte aims to be different. Debashish Banerjee, partner and data science and applied AI leader at the company, said that the foundational problem has always been hallucination. 

In tax, fabricated answers are not merely inconvenient. They are unacceptable.

“Whenever clients would use any open source LLM or any other models, there’s a big risk of hallucination, and especially for businesses like tax that we are trying to solve for, it’s like a complete no-no,” Banerjee told AIM. “It’s almost like those aircraft engines where you need to run at a 99.99% accuracy.”

The solution started with data discipline. Every judicial precedent, every law, every internal paper was vetted by functional experts before entering the system.

Model selection came later. Deloitte tested multiple foundational models across hyperscalers through a blinded validation process. This included Microsoft’s Omni, AWS Lambda, Llama, Mistral and Gemini. 

Tax experts evaluated responses without knowing which model produced them. The company then landed on Microsoft OmniParser 4.0 as the best model.

That choice led Deloitte to build on Microsoft Azure, though Banerjee stressed the setup remains flexible as models evolve. The firm has also invested in on-prem GPU infrastructure to experiment with smaller proprietary models in the future.

Refusing to Answer When it Doesn’t Know

What sets Tax Pragya apart is not just what it answers, but what it refuses to answer.

“To the extent, if our model does not know an answer, it will come back and say ‘Sorry, I don’t know the answer,’” Banerjee said. He contrasted this with open-source LLMs, which try to generate an answer irrespective of the question.

That restraint was tested internally for nearly a year. Deloitte treated itself as client zero. “We don’t want to launch something that we haven’t tested on ourselves,” Singhania said. “We gave it into the hands of 4,000 of our tax professionals.”

For 10 months, the platform was stress-tested against the most complex real-world queries Deloitte receives. A dedicated centre of excellence tracked feedback, accuracy and failure cases in production conditions.

The numbers convinced them. Roughly 60% of Deloitte’s tax professionals now use the tool actively. Hundreds of users spend more than 15 minutes on it every week. Hallucination rates dropped to near zero.

This internal confidence paved the way for a public launch.

Unlike conventional tools that merely retrieve documents, Tax Pragya handles contextual queries, performs background research and delivers substantiated answers. An interactive chatbot allows users to explore case law and migrate documents into conversations with a single click.

Becoming AI-First

“Tax Pragya represents a pivotal shift in how tax teams of businesses, large or small, can use responsible AI to access technical insights at speed,” Gokul Chaudhri, president of tax, Deloitte India, said in a statement. 

The rollout itself reflects Deloitte’s cautious ambition. The first wave targets 500 clients through closed-door sessions. A second wave in January expands access to 5,000 clients.

Singhania frames this less as a revenue play and more as an exercise of being AI-first.

“If you don’t do this, we will be left behind,” he said. “This has become table stakes.” That view extends beyond client delivery into talent strategy. Rule-based work, he argues, no longer attracts or retains skilled professionals. “We want to be a tech-first, tech-savvy business.”

Banerjee echoed that pragmatism. Deloitte’s broader agentic AI strategy is built around solving specific business problems, not chasing hype.

That philosophy was sharpened by industry missteps, including recent public cases of AI-generated fake citations; it agreed to issue a $440,000 refund to the Australian government for an AI-assisted report filled with fabricated references and hallucinated citations, which was published in July. 

Singhania did not dismiss those incidents. For Deloitte, he says the lesson reinforced the importance of guardrails, curated data and models trained to say no. 

Looking ahead, Singhania sees Tax Pragya as the beginning, not the end. AI will increasingly layer itself across compliance tools, advisory workflows and decision systems. Over time, Deloitte wants to take this capability beyond large enterprises.

“Our vision also is that this platform should be the MSMEs deep into the Bharat markets,” he said. “At a reasonably affordable price point.”

For Banerjee, the most significant risk is not technology failure but adoption.

“The tax profession is already being disrupted,” he said. “People will have more time to do more analytical and critical strategic work.”

The post Why Deloitte Built a Tax AI That Knows When to Say ‘I Don’t Know’ appeared first on Analytics India Magazine.