Recruiting · Production

Intuit uses an AI assistant with agentic rules to generate and scale interview exercises, with mandatory human review before deployment

The problem

Traditional technical interviews at Intuit did not reflect how engineers actually work with AI tools, and manually creating diverse, high-quality interview exercises required hours of effort per exercise.

Workflow diagram · grounded in source

Natural language exercise request

trigger

“A content creator can say, "Create a new AI exercise about comparing baseball team statistics," and receive a complete exercise — candidate instructions, assessor guide, and metadata in minutes”

AI generates exercise from rules

ai_action

“The system relies on deep and comprehensive agentic rules. Dozens of documents specifying exercise structure, candidate and role characteristics, quality guidelines, API categories, and assessment philosophy define the exact shape expect…”

Exercise routed to pending folder

routing

“Newly-generated exercises go into a pending folder. Humans must review the AI output before approval.”

Mandatory human approval

human_review

“AI-generated exercises can be good, but they can also hallucinate, make unrealistic choices or miss nuance, for example, with problems that are too easy or narrow, language that's not welcoming, or enhancements that don't give advanced c…”

Approved exercise enters library

output

“We've maintained dozens of exercises across both tracks, with rotation that prioritizes least-recently-used. Candidates rarely see the same exercise twice.”

Rule updates from learning

feedback_loop

“As we learn what works, we update the rules. The next generated exercise reflects that learning.”

Reported outcome

Intuit now generates complete interview exercises in minutes using an AI assistant constrained by agentic rules, maintaining dozens of exercises with consistent quality and a mandatory human review gate before any exercise reaches candidates.

Reported metrics

Traditional exercise creation timehours

AI exercise creation timeminutes

Exercise library sizedozens of exercises

Reported stack

AI assistant

Source

https://medium.com/intuit-engineering/ai-generated-interview-questions-leaning-into-and-complementing-ai-in-software-development-66da27d31620

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Intuit now generates complete interview exercises in minutes using an AI assistant constrained by agentic rules, maintaining dozens of exercises with consistent quality and a mandatory human review gate before any exe…

What tools did this team use?

AI assistant.

What results were reported?

Traditional exercise creation time: hours; AI exercise creation time: minutes; Exercise library size: dozens of exercises (source-reported, not independently verified).

How is this recruiting AI workflow structured?

Natural language exercise request → AI generates exercise from rules → Exercise routed to pending folder → Mandatory human approval → Approved exercise enters library → Rule updates from learning.