How Lars the Tax Bot Helped the WSJ Realize AI Could Make the Front Page

Tess Jeffers explaining the WSJ AI strategy at the International Journalism Festival in Perugia in April

When Tess Jeffers first pitched an AI-generated story for The Wall Street Journal's front page, she knew she was crossing a line that most newsrooms hadn't dared cross. Not because the technology wasn't ready - but because the stakes at The Wall Street Journal of getting it wrong are uniquely high. The experiment worked.

Key points summaries refined by human editors are now routinely sitting alongside traditional reporting on the digital front page that reaches 4 million subscribers daily. This stage represents the culmination of 12 months of careful experimentation that has transformed WSJ from AI-cautious to AI-confident.

WSJ's journey began with a familiar newsroom frustration: drowning in routine coverage that demanded accuracy but offered little room for creativity. Market updates, CPI reports, earnings summaries - the kind of stories that readers need but journalists don't exactly dream of writing.

The WSJ was spending too much valuable reporting time on straightforward “What happened” stories, but didn’t have enough time to dig deeper into the more advanced kind of “Why did this happen?” analytic reporting.

The math was brutal: cover more markets and events, or go deeper on fewer stories. Traditional newsroom logic said pick one.

Tess Jeffers, WSJ's Director of Newsroom Data and AI, saw a third option. What if AI could handle the "what" so reporters could focus entirely on the "why"?

The stakes: In a subscription-driven model where WSJ competes on authoritative analysis, any AI misstep could undermine reader trust built over decades.

Rather than diving headfirst into automation, WSJ started with what Jeffers calls "tiny experiments" - small pilots, tucked away and designed to prove value without risking credibility in front of a massive audience. The chosen niche was a bot that provided help for tax filers during this year’s tax season.

Enter Lars, the tax bot that became WSJ's breakout AI success.

The setup: Launched in March 2025, Lars uses Google's LLM Gemini 2.0 and a Retrieval-Augmented Generation model (RAG) trained on 1,300 WSJ tax articles, IRS publications, and tax guides. Unlike its predecessor Joannabot, Lars stays laser-focused on tax questions. (Joannabot is an AI chatbot that WSJ senior personal tech columnist Joanna Stern built to handle questions about the latest iPhone models. It has an unfortunate tendency to sometimes wander off-topic.)

Early skepticism was intentional: WSJ Editors and reporters “red teamed” Lars and deliberately looked for factual inaccuracies, harmful content, or whether it was going off topic and not being useful. When Lars passed those tests, accurate, reliable tax advice was well within reach.

The main effect: Rather than replacing tax reporters, Lars assisted the newsroom in better serving readers who have hundreds of highly specific and personal questions – far too many for an individual reporter to answer.

The side effect: In addition to Lars serving genuine audience needs, reporters could use their time tackling deeper investigative pieces about tax policy and corporate strategies.

What makes this different: WSJ doesn't use AI to write stories faster or to increase their output with a diminished standard. They use it to write more stories at the same quality level, expanding coverage without expanding headcount.

The workflow integration: AI tools are embedded directly in WSJ's content management system using FastAPI and Svelte frameworks, so there's no awkward switching between platforms. Reporters can access chatbots, automated drafts, and AI suggestions without leaving their normal workspace. The WSJ is primarily using Google Vertex as its base model for product features like bulletpoint summaries. On top of Vertex, in-house built custom prompts give the team the desired output.

The transformation has been more cultural than technical. As Jeffers puts it, the newsroom has evolved from having "AI antagonists and enthusiasts" to mostly "cautiously curious" journalists who see AI as a force multiplier.

Unexpected benefits: The AI evaluation process has actually improved human journalism. Weekly reviews of chatbot outputs have sharpened editors' eyes for accuracy and bias in all content, not just AI-generated pieces.

The WSJ is already planning AI-powered accessibility tools, including auto-captioning and automated alt text generation. They're exploring AI agents for subscriber retention and expanding their Korean-language AI translation service.

But perhaps most telling is what they're not doing: WSJ won't touch investigative reporting with AI. "Human judgment remains irreplaceable" when the stakes are highest, Jeffers notes.

AI-generated key point summaries are transparently disclosed as such. A Click on the “Read More” button reveals how the WSJ uses AI tools.

The tension they're navigating: Readers increasingly distrust AI-generated content, but they still want comprehensive coverage. WSJ's solution? Transparency about AI use combined with doubling down on human bylines for high-stakes stories.

The WSJ is betting big on what they call "answer engine optimization" - creating unique content that AI search tools can't easily replicate. Their focus on evergreen analysis and deep context suggests a future where AI handles commoditized news while humans own interpretation and insight.

“We’re not really interested in publishing material generated wholesale by AI. What convinces subscribers to pay for us is our original reporting, our analysis, our expertise. AI that can be created by anyone probably isn’t a subscription driver”, Jeffers pointed out on a panel at the International Journalism Festival in Perugia in April.

"Our north star is using AI to make journalism more impactful, not just more efficient", says Jeffers. After 12 months of disciplined experimentation, WSJ has found a way to do both - scaling coverage while deepening analysis, automating routine tasks while amplifying human insight.

The question for your newsroom isn't whether to adopt AI, but whether you can match WSJ's discipline in drawing the lines between what AI should and shouldn't do.

For Journalists Using AI: The WSJ approach proves that AI adoption doesn't have to be all-or-nothing. Start with "low-stakes" content where accuracy is crucial but creativity is secondary. Use the time savings to go deeper on analysis and investigation.

For Newsroom Leaders: WSJ's three-category staff breakdown - 10% antagonists, 10% enthusiasts, 80% cautiously curious - probably mirrors your newsroom. Focus training on that curious majority with "AI 101" sessions that demystify the technology with practical use cases rather than oversell it.

Note: I updated this newsletter on June 2 to correct some factual errors and expand on the aspect of key note summaries.

My toolbox for producing this newsletter: