Teaching AI to Think Like a Lawyer

In practice, tax law often turns on questions that look deceptively simple. Is a worker an employee or a contractor? Does a transaction qualify as business income or a capital gain? For lawyers and accountants, answering them can mean hours of combing through case law, weighing precedent and judgment calls that rarely resolve neatly.

In theory, artificial intelligence should be well suited to that kind of work. “Law is fundamentally a system of prediction,” says Benjamin Alarie, a professor and Osler Chair in Business Law at U of T’s Henry N.R. Jackman Faculty of Law. “If you know the facts, you should be able to predict the legal outcome.”

The catch is that law is heavily nuanced. Interpretation matters. Context matters. Small factual differences can change everything.

Tax law is that fussiness squared. It’s arguably the most complicated legal field – an aspect that Alarie finds intellectually irresistible. It sits at the intersection of politics and philosophy, sustained by a collective leap of faith that keeps taxpayers voluntarily complying with the system. But that same complexity slows decision-making and introduces inconsistency, even for experienced professionals.

Early machine learning wasn’t up to the task. It’s one thing to dethrone human champions on Jeopardy; it’s quite another to quantify the intuition of judges. Still, Alarie suspected that as the technology matured it could fundamentally change how law is practised and taught – even tax law.

In 2014, he approached two faculty colleagues with strong technical expertise – Anthony Niblett and Albert Yoon. Along with Brett Janssen, they formed a company, Blue J, and created a prototype chatbot. As a test, they turned it loose on some of the thornier areas of tax law – such as the seemingly-simple-but-actually-murky difference between an employee and a contractor.

The “answers” to such questions lie in the collective wisdom of case law. By analyzing hundreds of decisions in a particular domain, Blue J was able to extract signal from noise.

Alarie thought: Okay, this thing can scale. And a decade later, with the arrival of large language models and generative AI, it finally has – dramatically.

Rather than focusing narrowly on tax litigators, Alarie and his colleagues targeted accounting firms, whose business represents a far larger market than that of specialists who see only the cases that don’t move smoothly through the system. Last August – a decade after Blue J incorporated – the company raised US$122 million in venture capital, pushing its valuation past US$300 million, according to reports.

As AI has improved and data pools have grown, Blue J’s accuracy has soared, easing the work of the professionals who use it. Alarie likens it to giving a riding mower to a landscaper who had been cutting grass with scissors. “How this thing has changed my job is nothing short of amazing,” says Travis Thompson, a California-based tax lawyer, who calls Blue J “a paradigm shift for the tax industry.” What once took hours now takes minutes. Users report that 99.9 per cent of queries yield accurate results.

By March, the platform had surpassed six million queries. It is now used by three of the world’s four largest accounting firms and by nearly 5,000 firms globally. For Alarie, the rapid uptake points to a larger goal. “Our ambition isn’t to be the best Canadian tax AI,” says Alarie. “It’s to be the global platform for tax reasoning.” Thompson describes Blue J as “absolutely the best starting point for all my tax research.”

“Starting point” is key. Alarie is careful not to position Blue J as a replacement for human judgment. “Not everything should be automated,” he says. “Research, synthesis and initial analysis should be – because humans are slow and inconsistent at it. What should not be automated is the final judgment call, the advice to the client and the ethical and strategic dimensions of a tax position.”

Counterintuitively, Blue J becomes more valuable as a user’s tax expertise increases. It’s not a turnkey solution for the average taxpayer, both Alarie and Thompson stress, because effective use depends on knowing what questions to ask.

“The tax professional’s role has shifted from ‘person who finds the answer’ to ‘person who evaluates the answer and counsels the client,’” Alarie says. “We’re not replacing judgment – we’re compressing the time and cost to exercise it well.”