Mariya Lytvynyuk
The Problem Nobody Wanted to Talk About
One of our clients, a large ecommerce vendor with a sprawling catalog, had a bottleneck nobody liked to bring up in meetings.
Thousands of product images needed editing, moderation, and uploading to the catalog. And the team was doing all of it by hand: every single image, one at a time. The backlog kept growing faster than anyone could clear it, and the work was quietly eating months of skilled people's time.
The Challenge
On paper, the task sounded simple: get product images edited and into the catalog. In practice, every image moved through a chain of manual steps, and each step needed a person.
Manual editing: Every image had to be opened, edited, and exported by hand before it was fit for the storefront.
Manual moderation: Someone had to review each image for quality and compliance before it could go live.
Manual upload and matching: Approved files were uploaded one by one and matched to the right product in the catalog by hand.
None of these steps was hard on its own. Multiplied across thousands of images, they added up to a backlog measured in months, and a team that spent its days on repetitive work instead of the projects that actually needed them. And it was sitting on the part of the page customers reach for first: Baymard's usability testing found that exploring the product images is the first thing 56% of shoppers do when they land on a product page.
The Solution: An Automated Image Pipeline
We built them an automated pipeline that takes an image from raw file to live catalog entry, with no manual catalog management at the end of it.
1. Automated processing and editing: Images are processed and edited automatically, removing the slowest, most repetitive part of the job.
2. Human moderation where it's needed: After automated editing, the images that genuinely need a human eye are routed for moderation. People stay in the loop for judgment calls, not for the routine work a machine can handle.
3. Direct upload to SFTP: Once an image is ready, it's uploaded straight to SFTP: no manual exporting or shuffling of files.
4. Filename-based product matching: Each image is matched to the correct product automatically, by filename, so it lands on the right catalog entry without anyone managing the mapping.
The result is a system that runs in the background. Raw images go in; edited, correctly placed product images come out, and the catalog stays current without anyone babysitting it.
The Results
More than six months of backlogged work disappeared in weeks. But the number that mattered most to the client wasn't the time saved: it was what the time was freed up for.
That distinction shows up in the research too. When Anthropic asked tens of thousands of people how AI had changed their work, the most common answer was not that they did the same things faster. 48% described an expansion in what they could take on at all, against 40% who pointed to speed.
The team that had been buried in image processing finally had room for the work that kept getting postponed because nobody could get to it. The bottleneck stopped being a topic nobody wanted to raise, and became something that simply ran on its own.
What AI-Powered Ecommerce Looks Like in Practice
This is what AI-powered ecommerce actually looks like, and it's rarely the dramatic version from the headlines. It's a system that quietly handles the repetitive work in the background so your team can focus on the things that move the business forward. The same principle sits behind our AI sales agent: take the repeatable part, leave the judgement to people.
The best automation tends to be invisible. No one celebrates the image pipeline every morning: it just works, and the backlog that used to define the team's week is simply gone.
The Question Worth Asking
Most teams have at least one process like this: a manual, repetitive task everyone tolerates because it's always been done that way. It rarely makes it onto a roadmap, because it doesn't feel like a project: it just feels like the job.
If you have a manual process that's quietly eating your team's time, it's worth asking how much it's really costing you, not just in hours, but in the work that never gets done because those hours are already spoken for. More often than not, it's more than you think, and far more fixable than it looks. There are more examples of what that looks like in our case studies.
Frequently Asked Questions
It's a system that takes product images from raw file to live catalog entry without manual handling at each step. Images are processed and edited automatically, routed for human moderation only where it's genuinely needed, uploaded directly to storage (in this case via SFTP), and matched to the correct product automatically, so the catalog stays current without anyone managing it by hand.
No, and that's by design. The pipeline automates the repetitive work (editing, exporting, uploading, matching) but keeps people in the loop for the judgment calls. Images that genuinely need a human eye are routed for moderation; everything routine is handled automatically. The goal is to free the team from busywork, not to remove human oversight.
In this case, by filename. Each image is named in a way that maps to a specific product, and the pipeline uses that to place it on the correct catalog entry, with no manual mapping or catalog management required. The exact matching rule can be adapted to whatever naming or identifier scheme a store already uses.
For this client, more than six months of backlogged manual work was eliminated in a matter of weeks. The savings come from removing the per-image manual steps that don't scale, like editing, uploading, and matching thousands of images one at a time, and letting the team spend that reclaimed time on higher-value work instead.
Large catalogs feel the pain first, because the manual work multiplies with every product. But the same principle applies to any repetitive, high-volume process. If a task is manual, predictable, and eating meaningful hours every week, it's usually a strong candidate for automation regardless of catalog size.