We caught up with Dan Knight for this super exciting blog that covers AI and advances in eCommerce ahead of his talk at CogX later this month!

Why AI Is An Incremental, Not Transformational, Advance For eCommerce

Artificial Intelligence is a polarising subject. Often, people believe it will transform our lives and be the key to the future of business, especially eCommerce. The extreme counterpoint is that AI could take over the world in a matrix-style machine coup. I’d like to argue AI is a powerful but incremental step that follows very similar patterns and characteristics of the millennia of technological advances that preceded it. Like many technologies, it can be used for good or ill.

A Little Bit of History

Around 320,000 years ago, some of our ancestors found a better way of making stone tools. A technique called “prepared core” or Levallois technology, named after the Parisian suburb, where the first examples of its use were found. It meant good quality stone tools could be made reasonably quickly from a blank core of flint.

Artificial Intelligence seems like a far cry from hominin stone knapping. Still, I would argue it is just the most recent step in the progression of human technology that likely started with the humble rock. Before we get into why AI is an important but incremental step, not a transformation advance, I want to set out a few core concepts about how technology advances.

How Technology Helps

The steps of technology advance typically allow us to either:

1. Do something more efficiently
2. Do something to a higher quality
3. Do something that wasn’t possible before.

More often than not, technology provides efficiency. Before Levallois, Neanderthals could still make stone tools. It just took longer and was harder to get the exact shape desired. Before the early internet, we could still send each other information via post or look things up in a library. Before the tractor, people could still dig fields and grow food; long before the nuclear reactor, humans could make things (especially water) hot.

Examples of technology allowing something totally new are a little rarer, but pharmaceuticals are probably a good area—for example, the discovery and then refinement of penicillin.

These two steps in technology over 14 years enabled the use of penicillin, which allows the treatment of infections in a way that wasn’t possible before. Since then, mass production technology and genetic modification of the Penicillium mould have made the process more efficient.

Technology layering

Like the penicillin example, the internet is another group of technologies that build on previous steps. The internet is not one technology; it combines many things. For example, Transmission Control Protocol or Domain Name Servers and Hypertext markup language, etc.

Putting all this together with a vast host of other technologies means computers can communicate much more effectively and extensibly than before. Most modern uses of technology are not one thing created in isolation but are an iteration of existing technology or the combination of multiple independent ideas used together to create something more meaningful/powerful.

Technology isn’t always better

Technology can allow us to do something of a higher quality, e.g. laser eye surgery. It’s easy to think technological advances mean universally positive progress. I would argue technology tends to focus on efficiency and cost saving, and sometimes quality suffers due to technological progress.

To give a few examples:

Concrete: If we compare Roman concrete from 2,000 years ago to modern concrete, we discover Roman concrete is mostly better. It is better at resisting salt water, much more environmentally friendly, emits less toxic gas, and is much less likely to crack because it wasn’t used with steel reinforcement, which reduces the lifespan of concrete. Roman concrete cracks also self-heal thanks to lime clasts present in the mix. There are many concrete Roman structures standing in good condition today, but little, if any, of today’s concrete will be standing in good condition in 200, let alone 2,000 years from now, because technology has been optimised for efficiency.

Wikipedia: If we compare Wikipedia to previous sources of knowledge, like going to a library, Wikipedia comes up short both in terms of accuracy and depth of information. One study estimates it's 80% as accurate as other sources. It’s a fantastic bit of technology and a key pillar of the internet, but it has been optimised for efficiency. Wikipedia is also interesting because it’s used to train many AI models people use either in whole or in part, for example, chat GPT.

Digital Cameras: If we compare digital cameras to film, digital cameras don’t take quite as good quality photos, yet from the early 2000s, digital camera sales overtook cameras with film to the point that now, most camera sales are digital, and many of the remaining film camera sales are either cheap disposable ones or niche professional cameras. On most key metrics around quality, resolution and ease of capture, a film camera outperforms a digital camera. This is another example where an entire industry has persuaded consumers to settle for “good enough” in exchange for convenience.

There are, of course, countless examples of where technology improves capabilities, but it’s a mistake to think that something “more technologically advanced” is always better or better across all factors.

The use of AI in eCommerce

Mindera recently helped a client map out how AI could help their e-commerce business. We found that many of the scenarios produced little value, and those that were valuable were typically harder or more expensive to implement, making them potentially important but certainly not transformational.

AI, like many preceding technological advances, can do many things more efficiently but often to a lower standard. We grouped the use of AI into four sections with clear criteria/framework behind these four areas:

1. Quicker wins
2. Longer-term investments
3. Risky investments
4. Potential distractions.

Quick Wins

The scenario where AI adds the most value in eCommerce is where it is processing unreasonably large sets of data. It’s especially powerful when the data is messy and unstructured, as AI can be trained to navigate unstructured data and process it in a more efficient way than a person (at scale). What makes something a quick win vs. a longer-term investment is the simplicity of the result. If the source data is large and the answer can be some form of binary yes or no answer or score-based answer (e.g. 75%), AI is often a quick win.

For example, AI could find all the V-neck T-shirts in a catalogue of 100,000 products using only the images by first finding if it is a T-shirt and secondly finding out if it’s a V-neck or not. Of course, a person (even a young child) could do this pretty quickly and cheaply, but the power of AI comes when you want to keep re-processing a large data set to find dozens of attributes or the data set becomes millions of images.

Longer-Term Investments

These also need a large data set to be “understood”, often on multiple dimensions. In eCommerce, this could be understanding customers, products and orders, and the answer the AI is seeking might be a personalised recommendation, which is not a simple yes or no. This is a good use of the power of AI but requires a lot of investment of time in building and/or training models to work a certain way. The extra dimensions add complexity and, therefore, effort.

Risky investments

These are pushing the limits of what is possible. For example, one idea in this space is that a user could take a picture of themself and then choose an online item of clothing and reconstruct the two together in a virtual try-on scenario. Whilst it’s possible to do this manually with 3D modelling, it takes hours to build the objects. The idea is AI could do it in a quick and scalable way. This sounds great and could be transformational, but the challenge is getting good accuracy with limited data (say, just a photo or a few photos of a person).

Many start-ups are trying to do this specifically; however, if you ask 100 online shoppers how many use AI to try on clothes before they buy, I expect the answer would be zero or very close to zero.

Potential Distractions

Most AI applications people talk about in the eComm space fall into the category we labelled potential distractions. Of course, AI can write a product description for you. Below are two product descriptions for a dress on Net-a-porter. One is written by AI and had key information about the product, and the other is written by a person.

1. This maxi dress features a swirling striped pattern in a rich, luxurious cashmere-blend fabric. TheFabric.com's Ocon stripe dresses are made to flatter all body types and are perfect for an elegant evening out or a special day at the office. The waistband is made from stretchy cotton, and the hem is finished with an elegant draped skirt;
2. Gabriela Hearst's 'Ocon' maxi dress is an effortless way to inject colour into your wardrobe. It's crocheted from a blend of wool and cashmere in swirling stripes that are curved to enhance your frame. Style yours with a pair of statement earrings and heels.

AI with training and feedback can create better results, but the time saved using AI and either correcting mistakes manually or giving feedback versus just writing a good description is not huge, and so, as we explored earlier, we may end up with “good enough” for a small efficiency gain.

For some eComm businesses, this may be a worthwhile trade-off (e.g. if there are many products and no descriptions and users don’t need accuracy in the description). Still, for many eComm businesses, it’s a distraction.

To round up

The most expensive bread, the most expensive luxury handbags, the best Japanese carbon steel cooking knives and some of the fastest and most expensive cars all use some of the oldest production technology available, being largely or wholly handmade. This is because technology often makes things more efficient but at the cost of a slight reduction in quality.

AI can be used to improve quality, and it can undoubtedly make some processes more efficient, but it’s not trivial to make it do both. As such, it will follow the pattern of many preceding technological advances, changing how some things work and reshaping some roles, improving some areas but not being wholly or rapidly transformational.

The internet has no clear invention year. A message was sent in 1969 using some of the foundation technology of the internet, but you could hardly say the internet was invented then. It took decades to mature, and like the internet, over the coming decades, AI will evolve and gain adoption and reach new heights of success.

If you would like help with your technology strategy, including the use of AI, Data or eCommerce technology, please reach out to us at consulting@mindera.com, and we’d be happy to have a chat!