Redefining the data entry process through an AI-learning partnership with Microsoft.
XDC is a platform built on top of a machine AI capable of learning how to fill out invoices and billing receipts from human input. It outputs simple analytics to ensure quality control, and simplifies the role of data entry employees on what can be a complex process.
This was a project for AvidXchange that I did as part of my time contributing to their product lines. XDC was the backbone that supplied invoices in numerous client product offerings.
Lead Designer
Visual Design, Interaction Design, User Research, Prototyping & Testing, Information Architecture, Ideation Facilitator, UX Educator
Jan 2020 - June 2021
Create a data entry system to reduce overhead times on time to customers and reduce complexity for employees.
AvidXchange spent nearly 2 million per year on manual data entry of client invoices into their digital platform through outsourcing to third party companies. In a cost saving measure and to speed up the invoice intake process we partnered with Microsoft on their new OCR ( optical character recognition ) platform to create a tool that would read invoices and learn from human interaction the correct data it was supposed to fill out. This tool would also cut down on the amount of errors made in the entry process. Our speed from time the invoice was accepted until the customer saw it on their end could be up to two weeks in length.
To help me understand how the data entry process currently worked and, what factors in the design of the system gave them the most trouble. I arranged meetings over Teams and Webex to walk through their day to day activities while using the current manual system. I asked them questions and had them speak aloud to their thoughts while processing an invoice.
Results
Since the outcome of this project was to create a new indexing platform we had the opportunity to address some of these problems and reduce the friction in the data entry process. The AI on the backend could help resolve some of the complexity issues as it was trained on a particular client. This would free our internal indexers to do what they did best: auditing.
To get a rough idea of what the new process of indexing ( data entry ) could be, I helped to facilitate the construction of a simple user journey map with the creation of new personas ( roles ) that would come about from introducing an AI into the process. It was useful in helping the team ( product, stakeholders, development ) work out a roadmap of features and understand what exactly we needed to build.
From the prior investigative research and journey mapping session I was able to sit down and create some of the first mockups to present to the project team. These mockups were a starting point for discussing what we would like the future of the product to be. It also aided our stakeholders in envisioning how their current processes would change with the introduction of the AI into the mix.
Through a very informal A/B test we determined that Option B was easier for our indexing team to understand. The form fields were closer to what they were used to seeing and took them a lot less time to recognize what they needed to and where they could interact with the page. In these earlier iterations I toyed with the idea of having symbols to correspond with the machine's confidence in its answer. The machine would output a percentage on its certainty when filling out a field.
Option B also was an experiment to simply the color coding on the screen for easier readability. Labels were added to the outlines of the corresponding data to aid our coworkers with vision issues in understanding what they were looking at. It was certainly a challenge to come up with a total of 24 different colors that met accessibility contrast standards for our developers!
Improvements suggested in our sessions
Method
Unmoderated User Testing through UserZoom
Target Audience
6 Indexing Employees of AvidXchange
Overview
We wanted to test out the workflows with indexers to ensure its ease of use without needing to refer to knowledge articles. I tested over the course of a week, and spent another week correlating the data to present back to stakeholders, product management, and development leads. As a result of their feedback we were able to modify the designs to incorporate their feedback and present a finished product to development to begin coding out in earnest for the UI.
Taking the feedback to heart I spent some time implementing the changes that came up during user testing. During this time I was also overseeing a new design system version for the company. XDC was targeted as an ideal candidate to switch over. Thus during the process of incorporating user feedback and last minute additions by stakeholders I also had the added task of incorporating the new system version.
As part of the problem solving process from user testing I included the following in the new mockups:
We did a soft launch with less complex customer invoices in April 2021. As a metric for comparison we had both the old system and XDC process the same invoices.
The XDC platform powered by our AI reduced invoice time to customer by a week on average compared to the older system.
Compared to the user testing run in the prototyping phase we saw user satisfaction in the experience increased to 90% compared to the 16% before the alpha version.
With the alpha launch results the company was able to reduce its reliance on offshore data entry by removing one of the three teams we were contracted with.
During this project I learned the valuable lesson of how to juggle the need for accessible features without obstructing the use for everyday uses with the labelling system. There was also value to be added during the ideation phase to present visuals to the stakeholders for new project endeavors where there is no system to show as a reference point.
This was the first time where I had to work with an AI as one of the primary user personas. It was an interesting experience that I would like to revisit in the future. In short, the AI has needs too!
