It’s humbling to admit you don’t know something—maybe doubly so for an AI trained to be an all-knowing answer machine. Put on the spot, these systems just seem to prefer fudging facts.
Needless to say, that’s not a trait that small businesses want in a bookkeeper. But Intuit CTO Alex Balazs tells us it’s hammered that habit out of QuickBooks’s AI products.
“We spend a lot of time and energy making sure that the agents err on the side of saying, ‘I don’t know,’” Balazs said at a recent Manhattan event.
The financial software giant behind services like TurboTax, Credit Karma, and Mailchimp has a new cast of AI helpers on offer to its small and medium-sized business clients.
One is a payments agent that will manage and predict the lateness of invoice collection and draw up reminder emails in a chosen tone. An accounting agent categorizes expenses, and a customer agent prioritizes leads, drafts introductory chats, and schedules meetings. Other tools cover payroll, finance, and project management.
Balazs said the goal is to make the businesses feel as if they can delegate these tasks in the same way they would to a human worker.
“We instructed [the agents] to say, ‘Here’s where you’re starting; this is what the outcome is; here are the things you have access to; go complete the task for the customer. And oh, by the way, keep track of everything you’re doing and show them all your work at the end so they have complete control,’” Balazs said.
The safeguards: That last part is especially important for businesses entrusting AI with core tasks where precision is paramount. QuickBooks agents are designed to show their work so that both the evaluation system built into Intuit’s AI platform and human proofers can double-check their decisions, Balazs said.
Rather than confidently plowing ahead with best guesses, agents are also trained to reveal the levels of certainty behind a given choice.
“If [the accounting agent] is doing categorization, and it is certain of the category, it will categorize. In other instances, it’ll say, ‘I’m 80% sure this is the right category,’ ‘I’m 90% sure.’ In some cases, it will say, ‘I’m 20% sure; you should delegate this to a human,’” Balazs said.
The model stack: Intuit works with around 60 AI models–—half are publicly available LLMs and the other half are proprietary ones the company has trained itself, Balazs said. An in-house generative AI operating system (GenOS) platform selects the right model with a feature that translates prompts across models.
“The way the whole experience basically works is the engineer says, ‘I’m solving this use case. I think I’m going to generally solve it this way. I’m gonna get an agent to do X, Y, Z.’ You write it in natural language. [GenOS] will help you pick, well, you should use OpenAI o4, the secondary one should be this Anthropic model, and the third one would be this local one that we train,” Balazs said. “You basically then implement it. You run a test, and it will run it against all three models, and we have this thing called GenEval that comes in and evaluates the results.”
What’s different: Intuit is far from the only company that’s offering these kinds of AI agents that can sub in around the office. Companies from Salesforce and Microsoft to Zoom and Adobe are plying businesses with autonomous systems that can supposedly handle all manner of tasks.
Balazs said the range of the services Intuit has built out—with its early 2020s acquisitions of Credit Karma and Mailchimp—and all the data that comes with that sets Intuit apart from its competitors.
“There are very few companies that actually are in all the businesses that we’re in,” he said. “You could go to Salesforce and, theoretically on Agentforce, create agents. But those aren’t logical businesses that they’re in. So we are actually in these businesses.”
What’s next: As the company thinks about how to expand its AI offerings, it’s looking for ways that agents can work across that sprawling platform, Balazs said—maybe automatically preparing taxes in TurboTax for a QuickBooks user.
Intuit is also beginning to work with companies like Salesforce to make agents cross-compatible, according to Balazs. The company is one of dozens participating in Google’s Agent2Agent (A2A) Protocol project, which aims to make agents more interoperable.
Balazs said he thinks that standards like these will play a big role in customers trusting agents in the future.
“I do think that in the future, it is an absolute necessity that the agents across multiple [platforms] work well together. That they know how to communicate. They know what consent looks like. They know what trust looks like. The customer’s in complete control,” Balazs said. “The reason that we’re involved with these standards bodies is because we know without these standards, there’s going to be a couple of early, colossal failures. The customers won’t trust any of them.”