AI-powered photo approvals

Introduction

My role during this project was Lead Product Designer, working with our Platform engineering team to create a proof of concept.

Artificial Intelligence allows us to solve problems in ways we couldn’t previously. This has meant that organisations everywhere are flipping the script and instead of asking, “how might we solve problem X?”,  they’re rightly asking, “what problems might we solve with AI?”

We’ve been doing the same at Sidekicker, but even though the business is AI-curious, it’s far from sold on it. So, rather than talking at length about all the possibilities, we needed to find a way to make it tangible.

Problem

Sidekicker is a two-sided marketplace, connecting Businesses who need temporary staff with Workers who need work. When a Worker is onboarding to our platform, they need to upload a profile photo that meets a set of basic criteria (plain background, no hat or sunglasses, etc.).

Worker onboarding experience, showing where they are asked upload their profile photo.

‍

Our support team then routinely goes through and reviews these photos, approving or rejecting them based on the criteria. If they reject a photo, they need to contact the worker, explain why it was rejected and ask them to resubmit. This might delay the worker getting onto the platform by days, or at worst cause the worker to churn.

This problem stood out to us as a clear candidate for AI to solve: repetitive, rule-based, complex/variable inputs. It is also a problem that aligns with broader business goals of reducing our overheads, and improving the worker experience.

One of many conversations between our support team and workers who have had their profile photo rejected.

Success Metrics

The success metric for this project was ultimately to reduce the operational cost associated with manually approving profile photos (approx. 80 hours/month). But before we could achieve that goal, we needed to produce a proof of concept that could be endorsed by the business.

Process

Obligatory disclaimer: the actual process was of course not as neat and linear as below

Discovery

This was very much a tech-lead project. I worked with our Platform engineering team to assess the viability of the concept in the only way we really could - by throwing photos at it.

What we found was that not only was ChatGPT able to accurately approve and reject the vast, vast majority of profile photos we gave it, it was more accurate than our support team!

In a collection of test photos that have been approved or rejected, ChatGPT correctly rejected 95% of the photos that had been rejected by our support team. It also rejected an additional 12% of the photos that had been approved, but should have been rejected.

‍

Proving the concept using GPT-4.

Design

Once we knew the concept was feasible, the next question was how to use this solution to improve the worker experience. The biggest pain point for workers was when a profile photo was rejected, as they’d then have to go back into the platform and submit a new one. The clear solution to this was of course to provide that feedback as instantly as possible.

I created a high-fidelity concept that adapted the current experience to do just that.

Delivery

We haven’t moved into delivery as this project is still at the proof of concept stage.

Result

The business has endorsed the project and it will be moving from concept to reality sometime during 2024!

As this is still in proof of concept stage, so there are no additional success metrics to share. However, we do estimate that this work will save approximately 80 hours/month of our support staff’s time.

Lesson

Show, don’t tell. I am continuously reminded of the power of visuals in storytelling. As designers, we can wield this super power to more effectively communicate with stakeholders. And when combined with data, it is a powerful way to influence product and business strategy.

‍

_{Photo by Rock'n Roll Monkey on Unsplash}