Intuit scrapped its personal AI agent structure twice in 4 months. At VB Transform 2026, its AI VP referred to as that the quick path

Intuit was an early pioneer within the utilization of agentic AI, however its path to success has hardly been a straight line.

At VB Transform 2026, Intuit VP of AI Nhung Ho described how the corporate rebuilt its agent structure twice within the span of about 4 months, first transferring from a fleet of specialist brokers to a central orchestration layer, then abandoning that layer for a abilities and instruments based mostly system as soon as the orchestrator itself began failing below its personal complexity. The full second rebuild took 60 days, with a primary working model in below 20.

The failure mode that compelled the second rewrite was particular. Agents within the orchestrated system handed outcomes to one another in pure language, and every handoff misplaced context the subsequent agent wanted to behave appropriately. 

“If you have 10 agents and they all are passing to each other, every time that pass happens, error compounds,” Ho mentioned.

Why the orchestration layer broke down

Ho mentioned the unique push towards specialist brokers got here from an easy buyer grievance. A fleet of succesful brokers remains to be one thing a buyer has to handle, deciding which agent to make use of for which job. Intuit’s reply was a system that might take a job and route it internally, with out asking the shopper to select an agent themselves.

That orchestration layer held up for about three months, which Ho described solely half joking as roughly a 12 months within the compressed timeline of agent improvement in 2026.

It broke for a structural purpose somewhat than a capability one. Passing outcomes between brokers in pure language meant every downstream agent needed to infer how the upstream agent reached its conclusion, and that inference degraded with every extra hop. A ten agent chain didn’t fail often, it compounded errors by design.

That prognosis is what despatched Intuit again to a abilities and instruments structure.

The 60-day rebuild, and what it took to get engineering buy-in

Rebuilding a manufacturing agent system in 60 days required greater than an architectural choice. Ho mentioned the more durable drawback was inside, convincing each management and the engineers who had constructed the unique brokers that scrapping latest work was the proper name.

The pitch to management relied on proof somewhat than argument. Ho’s staff constructed a demo of the brand new structure utilizing actual buyer queries pulled from manufacturing, then confirmed it performing higher than the prevailing system on the identical duties. 

“The best proof, at least my belief, is what are customers trying to do? And whatever system you build needs to address those problems,” Ho mentioned.

Winning over engineering required a special case. Hundreds of engineers exterior Ho’s core staff had constructed the specialist brokers being retired, and the ask was to take their brokers aside into particular person abilities and instruments as a substitute. 

Ho mentioned the motivating argument was scale. A standalone agent solved one slender drawback, whereas a shared ability or instrument constructed into the brand new structure may serve each buyer who touched that a part of the product. That shift additionally modified what accomplice groups had been answerable for day after day, transferring their focus from constructing brokers to working evals, since evals grew to become the one solution to measure whether or not the brand new structure was truly working.

Bringing a human into the loop, and suggestions at a special scale

The clearest buyer dealing with results of the rebuild is a characteristic that lets a dwell agent dialog pull in a human — although it is at the moment in early testing, dwell to about 1% of Intuit’s buyer base. “We’re going to be scaling it up in the next few weeks,” she mentioned.

Ho mentioned a buyer can usher in an Intuit product help particular person mid dialog, or their very own accountant, or one in all Intuit’s personal bookkeepers, and that particular person joins with the total context of what the agent has already carried out.

Ho drew a direct distinction with how most AI chat merchandise deal with the identical scenario. A common function assistant answering a tax query sometimes ends with a disclaimer to seek the advice of an expert. Intuit’s system is constructed to attach the shopper to that skilled immediately, inside the identical dialog.

That human handoff sits alongside a permissions mannequin constructed for monetary knowledge particularly. Every motion an agent takes on a buyer’s monetary knowledge requires express permission first, although Ho mentioned that requirement can ease over time as clients construct belief within the system. Intuit retains an audit log of all the things an agent does that may be reversed if wanted.

Feedback within the agentic AI period

The rebuild additionally modified how Intuit gathers and makes use of suggestions, a shift Ho mentioned is qualitatively completely different from what got here earlier than. 

“Feedback in the past used to be very, very sparse, and it was also very bimodal,” Ho mentioned. “Either they loved it or they hated it, and usually it tends towards the negative.”

In a chat based mostly system, each dialog features as suggestions, which Ho mentioned moved the corporate from roughly 0.3% of consumers ever giving express suggestions to one thing near 100%.

Ho mentioned she has returned to writing code herself particularly to construct fashions that analyze that suggestions quantity systematically, in search of the place the system is falling quick at a scale no handbook overview course of may sustain with.

That quantity comes with a tone most product groups aren’t used to listening to immediately. Customers inform the agent precisely the place it failed, in plain phrases.

“They straight up tell you, ‘You suck. I hate this. This is not right,'” Ho mentioned. “But they’re also willing to give the systems grace and correct it as well, and so the onus is on all of us to harvest this new piece of feedback and type of feedback, and actually improve the system.”