Intuit Assist, AI help for tax experts on live calls
Tax experts spend their day with customers who have complicated questions and real worry about money. We built Intuit Assist into the tools experts already use, to take on the note-taking, searching and follow-up that fill a shift. Every suggestion stays optional.
- Role
- Product design lead
- Worked with
- Engineering, product leadership and Intuit’s AI organization
- Timeline
- 9 months
- I owned
- Strategy, research, prototyping, visual design, executive reviews and final QA
- Outcome
- NPS up 12 points and average handle time down 15 minutes
Decision 01
Design around the call
- Research
- Moderated sessions watching experts and customers in real time, plus interviews about daily frustrations. Privacy rules were strict, so we worked from anonymized journeys and aggregated insights.
- Situation
- Calls were getting longer as customer issues grew more complex. Notes, summaries and follow-up emails took up a large part of each shift. Many experts knew the tax language but wanted help writing with clarity, empathy and a personal touch.
- Choice
- Experts work across text, email, video and phone. For the first release we focused on support calls and split the work into before, during and after the call, with help at each stage.
- Why
- Each stage already had tasks that pulled the expert’s attention away from the customer.
Decision 02
Keep customer data on screen
- Situation
- Experts needed Intuit Assist within reach, but the screen was already full of customer information they relied on during a call.
- Choice
- A small entry button in the global header, chosen after several rounds of testing.
- Why
- It keeps high-value information visible and the assistant one click away at any point in the call.
Decision 03
Do the prep and the note-taking
- Situation
- Before a call, experts searched records for context. During it, they switched tabs to look things up and took notes by hand.
- Choice
- Intuit Assist summarizes the customer before the expert joins, answers plain-language questions in place, and turns the conversation into structured notes.
- Result
- In early testing, answering questions in place saved an average of twelve minutes per call.
Decision 04
Make every suggestion optional and editable
- Situation
- In testing, experts wanted to review and edit generated content before it reached a customer.
- Choice
- Recommended replies that experts can edit to match their own tone, and follow-up emails drafted from the call with adjustable tone and length. Suggestions are always optional.
- Why
- Editing and approval controls increased trust in the output. After follow-up testing we added more editing controls, plus quick-select options for faster adjustments.
Decision 05
Let experts grade the output
- Situation
- The model needed a way to improve from everyday use.
- Choice
- Thumbs up and down on AI output, with a short list of reasons for a thumbs down: not factually correct, doesn’t make sense, generic, or too long.
- Why
- Expert feedback shapes future improvements to the model.
Results
+12
points of NPS
−15 min
average handle time
88%
of experts engaged with it
−28%
silence or hold time on calls
What I learned
- 1
Control drives adoption.
Early adopters liked the speed. Being able to edit and approve is what made them trust the output, so we added more of it after follow-up testing.
- 2
Writing help mattered as much as answers.
Experts already knew the tax language. Generated summaries helped them write clearer responses to customers.
- 3
Status matters after the call.
Experts asked for clearer signals when a summary was ready.
What experts wanted was a faster way to act on their own judgment. Intuit Assist gave them time back and let them sound more like themselves on calls.
