Quoting tool

Case study

Quoting any roof in the world in <5 minutes

Product designer

2 Months

Web

A roof quote took 22 minutes, leaving customers waiting on the phone and putting deals at risk. Remi was producing more than 20,000 quotes a month, so every minute mattered. I redesigned the quoting workflow to bring quote time below five minutes, shifting agents from building quotes by hand to reviewing generated ones.

Aerial view of residential roofs used in the quoting tool

62% reduction in friction events

Reduced inputs, clicks and other unnecessary actions by 62%

3.5% increase in accuracy

This equates to roughly $350 per project, or nearly $4 million in reduced float

Average time reduced from 22 min to 5 min

This improves cost per quote, projected headcount needs, and customer satisfaction

The Problem

Customers were ready to buy. Quotes took 22 minutes.

Customers

Customers waited on the phone with a salesperson for more than 20 minutes to receive a quote, putting the deal at risk before they even had a price.

Quotes per month

20,000+

At more than 20,000 quotes a month, slow quoting meant customer wait time repeated and risked worsening at scale.

Accuracy cost

$4,000,000+

Measurement errors of 2–3% were already resulting in millions of dollars per year in discrepancies. Faster quotes also had to be accurate.

Existing Product

The tool was a Frankenstein.

Years of quick fixes had left agents navigating a clunky, inaccurate tool full of repetitive manual work. Customers waited while agents assembled each quote.

Users struggled with tedious and repetitive work

Research

Understanding the purpose of every movement and input

Quoting agents use the tool all day, every day. This allowed for a rich dataset of session replays, user interviews, and user working sessions.

I watched user sessions, documented every friction event (drag, click, and input), and timed users through each part of the quoting process: intake, property research, measuring, material inputs, and final review.

Heatmap of user touchpoints in the quoting interface — Heatmap of user touchpoints

Spreadsheet mapping every input in the quoting tool — Mapping every input and its use

The Core Bet Broken Into 3 Parts

Creating quotes→Audit generated quotes

Researching the tool surfaced a clear opportunity. Most of the data a quote needed already existed in the world. We just had to import it, contextualize it, and shift the user’s job from creating quotes from scratch to quality-assuring generated ones.

257 seconds

Clear hierarchy and flow

Interactions such as editing, multi-select, snapping, splitting, layering, and labeling are all necessary for nearly every quote. Making them intuitive was crucial.

494 seconds

Intuitive editing of generated models

We made the bet that being able to import the measurements and edit the errors, rather than draw from scratch, would be faster. This meant diving headfirst into interactions such as drawing, multi-select, snapping, splitting, and shortcuts.

303 seconds

AI jobsite insights

Many roofs came with 50-100 images and documents showing roof details, engineering reports, and photos of the worksite, often with conflicting information. Sifting through them and understanding what was correct took 25% of quote time.

Starting from scratch

Together with the company’s CXO, we decided that such a substantial change to the user’s role called for a fresh start.

Design and build a beta of each bet

It didn’t have to work in production, just enough to test quote times and accuracy.

Validate one variable at a time

Have a small portion of the quoting team use the new feature to see the change in time and accuracy.

Build iteratively

Assess and tweak as needed.

Bet one

Clear hierarchy and flow

Quoting multiple skilled construction trades anywhere in the US poses an incredible challenge. Users not only had to endure tedium, but remember hundreds of variable rules for each quote.

Every quoting agent had to remember a variety of tiny decisions

By working with finance and operations, we were able to find over 20 inputs that could be removed due to being redundant, deterministic in nature, or just inconsequential.

Simplified side panels aligned to the quoting workflow — Simplified side panels that followed the user’s natural quoting flow

Before, all information lived at the same level. This meant for each part of the roof (a slope, edge, peak, etc) users had to select 10+ items. Not only was it inefficient, it caused a level of noise that affected accuracy. We needed to drastically simplify without granularity being degraded.

Proper information hierarchy

Trade → Structure → Section → Lines & Polygons. This eliminated the user’s most repetitive interactions.

Answer the big questions first

The higher the user was in the hierarchy, the more important (cost-sensitive) the information asked.

Bet Two

Intuitive editing of generated models

Our R&D team had made astounding progress on 3D model generation based on publicly available data. This allowed us to instantly measure most roofs with a tiny margin of error. Still, editing and auditing these models was difficult enough that adoption remained low.

Polygon interactions

By auditing every interaction, it became clear that basic tools got 80% of the way there, but edge cases were so intense they heavily skewed the median. Because of this, we added tools like history, layering, point-snapping, angle-snapping, and keyboard shortcuts.

Keyboard shortcut menu over the quoting interface — Keyboard shortcut menu

Visualizing unusual roof areas was also a challenge. Users were switching tabs to compare references side by side. We added an adjustable modal within the tool to make comparison as simple as possible.

In-view references: LiDAR & Google

Bet Three

AI that users can trust at a glance

Users spent multiple minutes scanning the documents accompanying each quote request, including images, engineering reports, and sales documents. We found that users really only needed to answer the same ~8 questions every time.

The pipeline first processes every file. If the model finds content that matches a tag, it saves that content for extraction.

Imported files

LLM

Tagged files

Then, an LLM runs structured data extraction for each tag, pulling relevant context from the tagged image and determining which information is most likely to be correct. I added the LLM's reasoning as well as a quick view of all the files to ensure the user had confidence in the decisions the model was making.

Consensus view showing extracted pitch values and the supporting plan sheet — Generating consensus from tagged images

Results

From 22 minutes to under five.

The redesigned workflow brought quote time from 22 minutes to under five, shortening the wait for customers and helping salespeople deliver a price sooner. The team is continuing to refine the workflow toward a 90-second goal.

Curveball

Within 90 days, the tool was in use in countries across the world.

By resisting the urge to react too quickly, we were able to zoom out, define the actual challenges, and align on a solution that was more impactful and cost-effective.

Up Next

Estimated start week

↗