From POC to impactful AI for accessibility tool
When your website is among the 25 most visited in the world, you take customer experience seriously. Our Fortune 100 client cared deeply about accessibility compliance — and the ability to reach everyone in its community.
In the early days of the GenAI revolution, the client asked us to help identify use cases that could deliver the most reliable, positive customer impact. After running through a series of discussions and iterative proofs of concepts (POCs), iSoftStone’s AI-Powered ARIA Label and Alt Text Generator emerged as a leader. It soon became the first use of AI on the client’s website and its content team’s first taste of enterprise AI in their day-to-day work.
We initially collaborated with the client to implement a POC solution. Then, as AI platforms and services evolved, we transformed it into a fully integrated Adobe Experience Manager (AEM) toolkit. Now our client’s content team is empowered with inclusive design as part of their page draft process, and they’re creating a higher quality of accessible, compliant pages with a few clicks.
Here’s how we made it all happen.
The challenges of accessibility compliance at scale
In general, ARIA labels and alt text are fundamentals of accessibility compliance. However, they can be time-intensive to produce and can easily get lost in revision cycles and the momentum to publish, especially as web pages grow ever more complex. For example, WebAIM’s 2024 report of the top 1 million home pages shows that 54.5% are missing some alt text for images and 44.6% contain empty links.
With a website running to thousands of active pages, even a well-resourced company like our client isn’t exempt from these pressures. Before implementing a POC of iSoftStone’s AI-Powered ARIA Label and Alt Text Generator, employees had to write and manually review copy for images and links. It was a process that took weeks, with many hours generating and refining drafts and multiple hand-offs between teams. Then, once labels or alt text were final, they still needed to be translated for the website’s 65+ languages.
In other words, it was exactly the high-volume, large-scale challenge that GenAI is ideal for solving.
Says Elijah McCoy, the iSoftStone software development engineer on the project, “It was interesting even for the client’s technology team. Nobody there had done any AI work yet. This was the first, so they wanted to see if we could integrate it into something there was already a clear human-led process for.”
We did a lot of investigation into what AEM could author or support, and then we built out our own tooling to make it possible.
Elijah McCoy
Software Development Engineer | iSoftStone
A non-technical team gets their first taste of AI empowerment
We began by working with the content team to capture their processes, identify their requirements, and fully understand the accessibility guidelines the company used in addition to WCAG 2 standards. Then we got to work fine-tuning the AI model.
Our POC generator initially required page authors to input bulk image thumbnails (for alt text) and CTA/link copy (for ARIA labels) via a spreadsheet or CSV file. The AI returned drafts in spreadsheet form. An offshore resource was then on hand to verify the results and refine or regenerate them. The page authors then copied them into the CMS.
The POC soon proved the value of the AI for accessibility use case. Our client was pleased with the generator’s 95% precision rate, but with AI software tools rapidly evolving leaders wanted to see how much further the solution could go. As passionate advocates of inclusive design principles, they were curious about enabling page authors to truly own accessibility, and the technology team was fascinated by the challenge of pushing the technical boundaries of AEM.
So, the company challenged iSoftStone again. Could we introduce more automation? Could we build a version of the AI-Powered ARIA label and Alt Text Generator within the page authoring interface in AEM itself?
Says McCoy, “We did a lot of investigation into what AEM could author or support, and then we built out our own tooling to make it possible.”
McCoy and the iSoftStone team re-engineered the generator and brought accessibility directly into the page production process as a custom AEM toolkit.
ARIA labels and alt text generated in minutes
Now, with the help of AI, our client’s page authors are creating final drafts for alt text and ARIA labels by simply clicking a button in an AEM pop-up window.
That means complete ARIA labels and alt text for entire pages in just a few minutes. No more handoffs, no more long turnarounds.
To generate alt text, Azure AI Vision reads the pictures on the page. That output is combined with other information or text known about the image (for example, the size of the image or other supporting text) in a tightly engineered LLM prompt. The results? Minimal hallucinations, tightly written and accurate alt text at scale.
For ARIA labels, we architected and built the capability for the tool to crawl the relevant page for context and supply it directly to the AI model. The prompt includes specifics like headings and summaries to ensure an accurate output.
Page authors can easily regenerate copy for a specific asset with one click and then apply drafts across the whole page with another. If an asset is featured in multiple locations throughout the website, the same accessibility content will be connected to it. If an image is marked as decorative, the designation can be automatically applied across the entire site.
The time and budget savings have been exponential.
Our client is moving forward with confidence into other enterprise AI solutions. Says McCoy, “Almost everything I’m working on now [for this client] has some kind of AI component in it. It’s pretty interesting!”