Posted on Friday, 8 August 2025 by Léonie Watson in Strategy, User experience

Tags: Artificial Intelligence, Assistive Technology

Imagine being in a department store that sells clothes from multiple brands and having a personal shopping assistant to help you select the clothes you want to buy. As a blind person, that's about the only way it's possible to go clothes shopping, independently at least, but few stores offer such a service, so you resort to shopping online.

Except that retail websites are rarely accessible enough to make purchasing clothes much more than an exercise in “buy six things and return five". This isn't necessarily because the websites themselves are inaccessible, in the sense of being able to search for products, find out a little about them, add them to your basket or checkout, though that's not uncommon. It's that even the most accessible retail websites don't provide enough information to let you confidently make a choice, unless you can see the product images that is.

Let's take an example from a UK high street retailer:

Long sleeve Knitted jumper Crew neck It comes in black 100% Cotton.

You might think that's an adequate description, but if I were to ask you if the jumper is plain or ribbed, or if it falls to the hips or thighs in length, or if it's lightweight or chunky, could you tell me?

A different UK retailer does a bit better:

Crafted from soft pure merino wool, this jumper is a luxurious addition to your knitwear collection. It's crafted to a relaxed fit for endless layering options. A flattering v-neckline keeps things stylish, while cosy long sleeves with ribbed trims add a neat finish to this timeless piece.

But even so, what does relaxed mean, is it a deep V-neck, is the rib trim just on the sleeves or around the neck and/or bottom too?

Enter AI

Artificial Intelligence (AI) tools that can generate image descriptions are helpful in this context. My screen reader now has a feature that lets me request an image description from ChatGPT or Claude without having to switch contexts to do it.

There's no denying the fact that the people responsible for these websites should do a much better job of providing useful text descriptions, but we've been having that conversation for 30 years now, so forgive me if my patience is running a little thin!

There's also no denying the egregious theft of data and consumption of natural resources by AI platforms, and I've made my feelings on that clear elsewhere.

Using AI for image descriptions comes with a risk of hallucinations, but the consequences of falling for a hallucinated description don't put me at any more of a disadvantage than having no text descriptions at all. It's essentially a judgement call.

It's also not a great experience. It depends on the retailer putting the images where my screen reader can find them, or those images having enough of a text description to help me narrow down the choice of which images to query using AI, and on the content of the images themselves of course. Even under the best of circumstances, all of this takes time. A lot of time in fact.

Enter agentic AI

So, what if there was a different way? What if I could ask a digital personal shopping assistant to help me find clothes based on my preferred parameters? What if I could ask the assistant for product descriptions, then have a conversation with it to find out more, before letting the assistant know which one I'd like to purchase?

You may not know that this is already possible. It's still nascent technology, but it exists, and it's being used today to improve the accessibility and usability of the retail shopping experience.

An example is Innosearch, an AI ecommerce platform available in the USA, Canada, and the UK. It lets you shop from 500,000 online stores including those where I found the products mentioned above. Talking of which, it turns out that the long sleeved crew-neck jumper I found is for men... a minor detail that isn't obvious (if you use a screen reader) on the retailer's own product page!

In addition to presenting all products in a consistent and more accessible way than the websites they originated from, Innosearch automatically pulls out information that I'd otherwise have to actively look for - details only available in images for example. That isn't the interesting thing though. Innosearch has an agentic AI called CoBrowse, that goes beyond the behaviour of a typical AI chatbot.

If you're not familiar with the term "agentic", it refers to something that has agency, meaning it's capable of acting in its own right. Agentic AI is therefore a form of AI that can take actions on your behalf, as well as answer questions like a chatbot.

In this context, it means you can ask CoBrowse to search for a particular product. When I say "ask", I mean just that - you can enable your microphone and converse with it if typing isn't convenient. From there, it's a lot like having an actual personal shopping assistant in a bricks and mortar store. You can ask it to filter out results based on specific product features, exclude certain stores from the results, and once you've found a product, you can choose a size and ask it to add it to your basket, to open your basket, then add a coupon code or change the quantity of the last item to 3. One thing it can't do is complete checkout on your behalf, which is a reasonable limitation for security reasons.

Exit websites?

What all this really has me wondering, is why bother with a website? I don't mean the Innosearch website, I mean any retail website and, really, any website at all?

And, if you think this is a solution in need of a problem, then I'll wager you don't depend on the keyboard or use an assistive technology, that you've never been overwhelmed by the sheer quantity of information on a website, or been made nauseous or actually sick by constantly moving content, that you have good enough technology and connectivity to download vast quantities of JavaScript and other junk content without it affecting your ability to browse. I could go on, but you get the idea.

We humans love convenience and we dislike effort. We preferred the Graphical user Interface with its visual cues and affordances because it was more convenient than memorising countless command line instructions; we prefer the tap/swipe interaction of touch screen devices because it takes less effort than using a mouse/trackpad; and we generally prefer to have a chat because (for many of us) it requires no effort at all - and what could be more convenient than that!

No, I'm not going to predict the end of the web as we know it, but I do think there's a high probability that the way we use the web will change significantly - it's already changing in fact.

In July 2025, OpenAI revealed that ChatGPt users send about 2.5 billion prompts a day. That pales in comparison to the estimated 14 billion search requests Google handles each day, but when you consider that in December 2024 OpenAI reported the number as about 1 billion prompts a day, it's hard to deny that AI use is on the rise.

A recent report found that Google's AI search is reducing traffic to websites, in some cases by as much as 50%. Presumably AI platforms like Innosearch are doing the same thing, albeit at a far lesser scale, but what happens as more people discover that shopping with an agentic AI is much more convenient, or when Google uses AI to field all 14 billion of those search requests?

Whether we like it or not, and I confess I haven't yet made-up my mind, AI is almost certainly here to stay. The last time I remember a hype-cycle like this, it was the late 90s and what became known as the Dot-com Bubble. Yes, that bubble burst and the same will probably happen with AI, but what seems likely to remain is an agentic web of some kind.

And accessibility?

We're already being approached by organisations that are pivoting towards the agentic web and asking us what it means for accessibility. If they replace their websites with agentics that generate the information someone wants in the format they prefer, will that help or hinder accessibility? More to the point, how do we determine with reasonable confidence that the generated content is accessible? Anyone who's used the likes of Meta, Gemini, Claude, or ChatGPt for more than 10 minutes will know that you don't get the same response twice even when you repeat the exact same prompt, so where does that leave the concept of representative sampling, and how can we be sure our digital products meet our legal obligations?