Case study · AI · Accessibility
AI alt text at retail scale
Partnering with Adobe on Brand Aware Metadata to put LLM-written alt text into production across walmart.com.
<img src="…" alt=""><img src="…" alt="Maroon air fryer with digital controls on a kitchen counter beside a bowl of artichokes and a matching carafe.">Public walmart.com homepage: Jan 2023 (Wayback Machine) vs. today
- 70% → 99.1%
- alt text success rate, manual entry vs. the AI pipeline
- ~200
- workdays saved a year, about nine months of one person's time
- 48K
- assets a year now get alt text automatically
The problem
On a site the size of walmart.com, every image needs alt text. Screen-reader users rely on it to understand a page, and search engines read it too. But alt text had no dependable path through the content pipeline. When people wrote it by hand, a large share was missing or unusable, and nothing caught it before publishing.
What I built
Adobe's Brand Aware Metadata (BAM) lets a large language model look at an image, its existing metadata and a prompt, then write a value for a metadata field. I worked with Adobe as a hands-on design partner: tuning prompts, testing edge cases and giving product feedback. With the Copy team's Senior Manager of AI, I designed the prompt and the workflow around it.
- Generate: the model writes alt text as assets are processed in AEM Assets.
- Check: automated QC fails anything that's missing or over 150 characters.
- Regenerate: failures go back for a new pass and push, so nothing ships without alt text.
- Review: the process went through legal, compliance and accessibility teams before rollout.
Results
It's live across walmart.com product and marketing imagery, about 48,000 assets a year. Manual entry succeeded roughly 70% of the time. The pipeline succeeds 99.1% of the time on the first pass, and the regenerate loop catches the rest, so every published asset carries alt text.
How the savings were measured: the Copy team's Senior Manager of AI calculated them from team time-reporting data and feedback from the teams involved, then presented them to agency leadership. The estimate is about 200 workdays a year, roughly nine months of one person's full-time work. The project was also featured in Walmart's session at Adobe Summit 2026.
You can see the difference yourself: compare the homepage's image tags in the Wayback Machine (early 2023, mostly alt="") with the site today.
What's next
The one manual step left is a person flagging a failure and asking for a regeneration. A small Slack bot that triggers the regen directly would close that loop.