Want smarter insights in your inbox? Sign up for our weekly newsletters to get only what matters to enterprise AI, data, and security leaders. Subscribe Now
In the frenzied land rush for generative AI that followed ChatGPT’s debut, the mandate from Intuit’s CEO was clear: ship the company’s largest, most shocking AI-driven launch by Sept. 2023.
Responding with blazing speed, the $200 billion company behind QuickBooks, TurboTax, and Mailchimp, delivered Intuit Assist. It was a classic first attempt: a chat-style assistant bolted onto the side of its applications, designed to prove Intuit was on the cutting edge.
It was supposed to be a game-changer. Instead, it flopped.
“When you take a beautiful, well-designed user interface and you simply plop human-like chat on the side, that doesn’t necessarily make it better,” Alex Balazs, Intuit’s Chief Technology Officer, told VentureBeat.
AI Scaling Hits Its Limits
Power caps, rising token costs, and inference delays are reshaping enterprise AI. Join our exclusive salon to discover how top teams are:
- Turning energy into a strategic advantage
- Architecting efficient inference for real throughput gains
- Unlocking competitive ROI with sustainable AI systems
Secure your spot to stay ahead: https://bit.ly/4mwGngO
The failed launch plunged the company into what Dave Talach, SVP of the QuickBooks team, calls the “trough of disillusionment.” The chatbot took up valuable screen space and created confusion. “There was a blinking cursor. We almost put a cognitive burden on people, like, what can it do? Can I trust it?” Talach recalls. The pressure was palpable; he had to present to Intuit’s Board of Directors to explain what went wrong and what the team had learned.
What followed was not a minor course correction, but a grueling nine-month pivot to “burn the boats” and reinvent how the 40-year-old giant builds products. This is the inside story of how Intuit emerged with a real-world playbook for enterprise AI that other leaders can follow.
How a split-screen observation sparked Intuit’s AI pivot
The pivot away from the chatbot began by observing customers as they did their work. Talach recalls his team’s “big aha moment” when they noticed QuickBooks users manually transcribing invoices with a “split screen”—an email open on one side of their monitor, QuickBooks on the other.
Why force a human to be a copy-paste machine when an AI could ingest data from the email and populate the invoice automatically? This observation sparked a new mission: stop trying to invent new behaviors with chat and instead find and eliminate “manual toil” within existing customer workflows.
Recognizing this bottom-up momentum, CTO Alex Balazs and Marianna Tessel, GM of the business group, made their move. “We need to make a declaration together,” Balazs recalls Tessel saying. The only path forward was a full commitment to an AI-native future. “It’s burning the boats, and it’s only going to be the AI way.”
To execute this, management redeployed a key technology leader, Clarence Huang, from the core tech team and “parachuted” him into the heart of the QuickBooks business. His mission was to scale a “builder-centric mindset” of rapid, customer-focused prototyping.
Embracing this new model also meant dismantling the old one. To empower smaller, faster teams, the company made a difficult decision: it slashed layers of middle management, letting go of 1,800 employees in 2024 in roles no longer aligned with new priorities, while pledging to hire back about 1,800 new employees with skills in engineering, product and other customer-facing roles.
The three-pillar framework that turned AI failure into enterprise success
Intuit’s transformation required a new operating model built on three core changes: empowering its people, re-engineering its processes, and building a technology engine for speed.
Pillar 1: Forge a ‘Builder Culture’
To execute the pivot, Intuit first had to get the right people in the right structure and empower them to work in entirely new ways.
- Aggressive Talent Acquisition: The company hired aggressively to add to its core AI team, bringing it to several hundred today, from just 30 people in 2017 – accelerating over the past two years by poaching top-tier AI leaders from giants like Uber, Twitter and Bytedance.
- New Team Structures: The core of the new model was small, empowered, cross-functional teams. These groups, sometimes including members from up to 10 different units – data science, research, product, design, engineering, and more – focused solely on delivering a specific agentic experience. To enable this, managers ruthlessly prioritized, eliminating any tasks that weren’t among the top three priorities. “That ruthless prioritization… was really, really important,” Huang said.
- Empowered Ways of Working: Within these teams, traditional job descriptions dissolved in what Huang calls a “smearing” of roles. Everyone was expected to talk with customers. Huang kept his own spreadsheet of 30 customer names he called regularly. The transformation was profound, exemplified by data scientist Byron Tang, who stunned colleagues by using new AI “vibe-coding” tools to build a full prototype with a beautiful UI single-handedly. Huang recalls his reaction: “Oh my god… you are the renaissance man. You got it all!”
Pillar 2: High-Velocity Iteration Over Bureaucracy
With the right people in place, Intuit systematically dismantled the processes that slow large companies, replacing them with a system built for speed and customer obsession.
- Prototype-Driven Development: The old way of using spec docs was replaced by a new mantra: a prototype is worth 10,000 words. Teams began shipping functional prototypes to customers almost immediately. “We’ll literally show a working, functioning prototype to the customer… and we’ll vibe code it on the spot,” Huang explains. “The reaction on their faces is just magic.”
- Customer-Centric Design: This rapid feedback loop led to key innovations, including a “Slider of Autonomy,” a concept popularized by developer Andrej Karpathy in June. Intuit noticed that customers feared features that seemed “too magical,” so it gave them control over the level of AI intervention, ranging from full automation to manual review – creating a “smooth onramp” to trusting the agents. For example, in Intuit’s QuickBooks accounting agent, users can click a button to allow the agent to post all transactions it recommends. But if users want to maintain more control, they can use icons to see the entire reasoning chain of the agent for user-friendly explanations.
- Ruthless Bureaucracy Busting: Leadership actively cut red tape. They implemented a “no meetings on Tuesdays” rule on the platform team, banned afternoon meetings for individual contributors in the business unit, and instituted a formal “friction busting” campaign, imposing a seven-day deadline for leaders to unblock any inter-team disagreements. A rule limiting AI rollouts to a small number of customers for experimentation was revised to allow for tests involving up to 1,000 customers at once, up from the original limit of just 10.
Pillar 3: Build an Engine for Speed
Underpinning the entire effort is GenOS, Intuit’s internal AI platform. It flowed from CDO Ashok Srivastava’s desire to democratize AI access across the company.
Instead of a slow, top-down build, the platform evolved at the same speed that the business grew, through a strategy CTO Balazs calls “Fast Follow Harvesting.” As customer-facing teams built agents, they would identify gaps in the platform. A central team then ran in tandem with the customer teams, closing the gaps with new features.
A key feature of GenOS was the Agent Starter Kit, which enabled 900 internal developers to build hundreds of agents within a five-week period. Other features included a runtime orchestration and a governance framework.
Another core component was an LLM router that provides resilience and allows LLM calls to flow to different models depending on which one is best for the given task. Huang recalls getting a late-night call from Srivastava. “He’s like, ‘OpenAI is down. Are you guys okay?’” Because the team was on GenOS, “it just auto-switched to the fallback LLM in the gateway… it was okay.”
This platform allows Intuit to leverage its core differentiator: decades of domain-specific data. By fine-tuning models on a finite set of financial tools and APIs, Intuit’s agents achieve accuracy that general-purpose models can’t. “In all of our internal benchmarks, our stuff just works better for in-domain data,” Huang said.
The payoff: 5 days faster payments and 12 hours saved monthly
The result of this pivot is a suite of AI agents deeply woven into QuickBooks and increasingly across Intuit’s other products. The QuickBooks Payments Agent does things like proactively suggest adding late fees if a customer’s payment history shows they’ve been late in the past. The impact is tangible: Small businesses using the agent get paid, on average, five days faster, are 10 percent more likely to get paid on overdue invoices, and save up to 12 hours a month.
The Customer Agent transforms QuickBooks into a lightweight CRM, scanning connected Gmail accounts for leads, while the Accounting Agent automates transaction categorization and flags anomalies. Today, these “virtual employees,” as Talach calls them, surface their work through tiles in the QuickBooks “business feed,” turning the dashboard into an active, collaborative space. These translate into more holistic offerings for customers, and could help Intuit take market share from competitors who offer similar services, such as HubSpot.
In last week’s quarterly earnings call, CEO Sasan Goodarzi credited the company’s strong results, 16 percent growth for the full year – to its investments in AI. He said the agent launch was already bearing fruit: “We’re seeing strong traction since last month, with customer engagement in the millions and repeat usage rates significantly above our expectations.”
Intuit is now applying this playbook to bigger challenges, recently announcing agents for mid-market companies with up to $100 million in revenue – a significant expansion from Intuit’s traditional base of customers with $5 million or less in revenue. The logic is simple: Bigger customers have more complex workflows, and thus a greater need for AI agents.
For enterprise leaders navigating their own AI transformations, Intuit’s story offers a clear roadmap. The initial stumbles aren’t just common – they may be necessary. The path forward is more than integrating AI magic. It’s about dismantling old ways of working and building a culture, process and platform that lets established companies move with startup speed while following AI-age best practices.
The biggest lesson? Start with the work your customers actually do, not the technology you want to deploy.
