Building Beyond a ChatGPT wrapper

Owning 0-1 design and product direction for AI-native EdTech

0-1, AI, Consumer, EdTech

TEAM

Solo Product Designer (me)

3 Engineers
1 Product lead

SCOPE OF IMPACT

40M+ Weekly Search Queries
60% of weekly active users
on Quickbooks Online

TIMELINE

3 Months

CONTEXT

THE BIG PICTURE

1 / Move away from a keyword-based search

While a keyword-based search worked well for Transactions, it was insufficient as Search expanded into other entity types like Contacts.

In the GenAI age, querying in natural language is the bar

2 / Company-wide focus on AI

Intuit initiated a number of initiatives surrounding Intuit AI Assist, tinkering with the use of AI in saving users time across multiple Quickbooks workflows.

QuickBooks is a financial management software that helps over 7M+ small businesses streamline accounting, payments, payroll, etc

The Quickbooks Global Search allows customers to quickly look into their transactions, contacts, and records

3 Key Business and Technical developments happened within Search at Intuit:

3 / Introduction of Full-page Search Results

As the entities that Search covered expanded, Quickbooks decided to introduce Full-page Search results, which opened up fresh real-estate for results, recommendations, upsells, etc.

THE ASK

How might we drive users to discover more of Quickbooks through Search?

BUSINESS TRANSLATION

Capitalise on new technical capabilities to

boost engagement and upselling through suggestions

within Search results, nudging more users into ‘action’ on Quickbooks

FREE-FORM RESEARCH

I went around San Jose markets and clicked pictures of good suggestive experiences in-real-life, while also doing an audit of good “suggestions” online

So I proceeded to think through:

What makes a good suggestion?

AHA !

It turns out (validated from existing research at Intuit),

A suggestion is only a “suggestion” if it’s relevant and not pushed too hard.

Much else is an “ad” (we hate ads — they’re annoying)

EMERGING FIRST PRINCIPLES

So with some push-back from PM,
we advocated for some key decisions:

  1. Suggest, but don’t be pushy

    It was important we don’t keep pushing in-product discovery in the garb of it driving revenue but creating a bad search experience. We have to respect that our users already pay a premium to use Quickbooks, and cannot be upsold to at every opportunity.


  2. Never get in the way of results

    The Search dropdown has limited real estate. An accountant user wants to go in, look at the transaction at a glance, and get out. Obstructing that will raise VOC (voice of customer) complaints.

    So when the PM pushed for recommendations to be on the top of the dropdown, we pushed back.

  3. Relevancy is key

    The only factor differentiating a good suggestion from an annoying ad is relevancy. So we need to ensure that till the ML model has learned and trained significantly, we limit the number of places where these suggestions appear

ALIGN, ALIGN, ALIGN

These decisions required a lot of alignment work;

And rounds and rounds of presenting at design crits

DESIGN

Form Explorations

I thought through a bunch of low-fidelity forms and ran a design crit workshop to gain wider feedback and thoughts

“will this scale as our suggestion scope grows?”

“is this too similar to the shortcut icons we use on the dashboard?

“can this be developed in time for v1 release?

Early concept for alignment

“does this form support enough text length across other languages?”

“will we need more buy-in from design systems for this?

We aligned on using cards as the visual element.

This made sense since:

  1. We wanted to limit # of suggestions shown

  2. Lend the suggestions enough visual cue for new (often overwhelmed) customers to notice and feel guided

  3. Unlike chips, they could fit into other places in real-estate

Anatomy of a card

DESIGN HIGHLIGHTS

Intentional delay to save big $$

I implemented an intentional loading skeleton delay for the recommendations to:

  1. Save costs by not running the recommendation engine while the user is mid-typing

  2. Make recommendations feel personalized and ‘calculated’, not preloaded ads

Never more than 3 in the dropdown

People can’t process much beyond 3 options.


This ensures that each suggestions provide genuine and relevant value - and that’s key to these suggestions not being perceived as advertisement / annoying upsells.

People like to see faces

Based on past internal research,
Users were more likely to click on ‘Expert Help’ if they saw an expert face

Accessing expert help in the first 30 days is highly correlated with platform retention.

SHIPPING TIMELINE / IMPACT

30K customers

Oct 2024

>

(Metrics and numbers from testing are confidential)

6.4 Million customers

Q1 2025

LEARNINGS FROM DEV HANDOFF

300K customers

Dec 2024

>

30M+ Weekly Search queries

  • Save Devs HOURS by teaching them key Figma shortcuts

  • Once you have a design, Devs are your next customers

  • Be in their proximity. Sit with them.

a love letter to ‘Search’

REFLECTIONS

Search is SO beautiful.

Working on this project had me think very deeply about Search interfaces.

We’re searching all the time. For a friend in a crowd, our car keys around, a budget meal nearby, or a photo lost to memory. Search interfaces are the ultimate capture of intent - almost like a wishbox.

With much of the future of the internet moving towards AI, it feels so fitting that people will get to express what they’re looking for in free form. Search makes our curiosities tangible. Im so glad I got to work on this, of all the things.

CROSS-FUNCTIONAL FEEDBACK

Jen, who I reported to (Principal Product Designer) and me 
@Figma Config 2024!