One Tuesday morning an email landed in my inbox with the subject line: “Your AI just killed my $8k campaign.” It was from an ecommerce founder who had uploaded 40 product shots for Instagram ads and gotten back images with jagged edges and floating shadows. I had spent months polishing the ad generator, yet this single message exposed the weakest link.
The Setup That Wasn’t Working
AdLoft had launched with a solid pipeline: upload photo, pick template, generate variations. Early users loved the speed, but power users kept hitting the same wall. Background removal failed on reflective metals, transparent packaging, and complex jewelry. Conversion rates for those categories sat at 1.8 % while the rest of the catalog averaged 4.1 %. Support tickets about cut-outs were climbing 22 % month-over-month. I knew the model needed work, yet I kept prioritizing new templates instead.
The Decision Triggered by One Message
That customer email included side-by-side comparisons showing exactly where our output fell short. I spent the next weekend auditing every failed generation from the previous 30 days. The pattern was clear: 67 % of low-performing ads used products with tricky edges. I decided to pause all new feature work and rebuild the background-removal engine from scratch. Over the following four weeks I swapped in a fine-tuned segmentation model, added manual refinement brushes, and introduced an automatic shadow-matching pass. The change touched roughly half the codebase, including the preprocessing queue and the final export layer.
During testing I also linked the new engine to our background remover for ecommerce so users could download clean cut-outs for other tools. Internal benchmarks showed edge accuracy jumping from 71 % to 94 % on our hardest test set.
Numbers That Told the Real Story
After shipping the update we watched metrics for six weeks. Average ad conversion rate across all users rose from 3.2 % to 4.9 %. Users who had previously churned after one month stayed 38 % longer. The most striking number came from the exact cohort mentioned in that email: reflective-product campaigns improved by 41 % in ROAS. Support tickets about background issues dropped 81 %. One customer’s blunt feedback had saved us months of guessing.
What I Learned About Building in Public
I now keep a dedicated Slack channel where I drop every piece of raw feedback, good or bad. The next time a single email forces a pivot, I want the whole team to see it the same day. That habit alone has shortened our iteration cycle from six weeks to nine days. The product is still evolving, but at least it’s evolving in the direction users actually feel.
I’m Didar, the founder of AdLoft — an AI ad creative platform turning product photos into professional ads. I write here about the stuff I’m building and what’s working.