Released in The Interline’s AI Report 2024, this executive interview with Theodo is one of a fourteen-part series that sees The Interline quiz executives from companies who have either introduced new AI solutions or added meaningful new AI capabilities into their existing platforms.
For more on artificial intelligence in fashion, download the full AI Report 2023 completely free of charge and ungated.
Key Takeaways:
- While generative AI has gained significant attention and can be highly versatile, it’s not always the best solution for every problem. Specialized models can be more energy-efficient and better suited for specific business needs.
- A structured approach, starting with a Proof of Concept (POC) to frame the use case and build a baseline, followed by developing the product for widespread use, can significantly shorten the time from initial idea to real value in AI projects.
What’s your working definition of AI? Does it differ from the public understanding, which is currently dominated by large language models and generative text-to-image models? And how does that definition manifest itself in your solution(s) and services for fashion?
Our understanding of AI closely follows Luke Julia’s perspective. He said “artificial intelligence” is a vague and inaccurate term. Instead, what people are generally referring to is machine learning, which involves systems learning from data to autonomously perform tasks.
To get more technical, machine learning encompasses deep learning, a subset focused on creating more complex models based on neural networks. Within deep learning, we have specific models like Large Language Models (LLMs) and transformers, which have gained significant attention recently.
Generative AI is a subset of deep learning, and as we all know really caught the spotlight in 2022 with models like ChatGPT for text and DALL·E for vision. What really sets these new AI models apart is their accessibility. Unlike previous AI models, which were part of larger products, these Generative AI models are products themselves and are accessible to any type of user no matter their knowledge level of AI.
In our solutions and services for the fashion industry, understanding how AI works is crucial. We typically combine multiple models to address our clients’ business problems effectively. While clients often approach us with requests for generative AI products, we dive deeper into their problems and find that generative AI might not always be the best solution. These models are often the most energy-intensive and may not be well-suited for all use cases. Instead, we find that more specialised models are often more energy-efficient and better suited to solving our clients’ business problems.
One of our recent projects was developing an AI personal shopper for a luxury brand. The innovation department of a major French luxury group wanted to enhance the conversion rate of their e-commerce site by replicating the experience of a personal shopper in-store. We developed a conversational product recommendation tool based on LLM in less than two months. The user enters a recommendation request and can then chat with their AI personal shopper. The deployment of the first version took just two weeks, resulting in a 42% click-through rate on products.
As software consultants, you tackle a different side of product development – working with brand and retail businesses to create solutions and build out their technology ecosystems in ways that are logical, viable, and driven by real value. How does that map to the AI space right now, when there are so many conflicting visions for what is possible, and so little clarity around what results to expect?
In the current AI landscape, there are many conflicting visions about what’s possible, and there’s often little clarity around the expected results. Historically, clients would present us with a business problem, and we would find the most appropriate solution, which sometimes involved using machine learning models, and other times, more traditional algorithms.
However, a significant shift in the market is the growing trend of companies wanting to implement generative AI solutions without fully understanding its applicability to their specific problems. This trend is often driven by pressure from financial markets or executive leadership.
To address this, we’ve adopted a product-focused approach to identify which problems are well-suited for generative AI. While in the past year, most engagements were Proof of Concepts (POCs), clients are now starting to understand where generative AI is and isn’t applicable. As a result, the projects clients bring to us are becoming more relevant and are increasingly driven by real value.
When it comes to the results to expect from AI solutions, it heavily depends on the proprietary data available—how structured, clean, and diverse it is. Evaluating the quality of generative AI outputs can be challenging, unlike with classic AI use cases where quality assurance was relatively straightforward.
While generative AI tools can streamline processes, it’s important to understand that today they don’t fully replace human intelligence. Instead, they should be seen as valuable assistants.
To truly extract value from AI solutions, it’s important not to see them just as extensions of current processes but to rethink business models and understand what users truly need. Just as success came from rethinking business models during the shift from paper-based to digital processes, embracing AI requires a similar innovative approach.
Theodo has been active in AI, with a dedicated engineering team, for the past eight years. In that time you’ve worked extensively with non-generative AI – across deep learning, computer vision and other approaches. Do you believe fashion’s current focus on generative AI runs the risk of becoming too narrow, and leading brands to miss out on the potential of those other avenues?
While the fashion industry’s current focus on generative AI is understandable, I do agree there is a risk of it becoming too narrow. Generative AI has certainly brought AI into the spotlight, but it’s not always the best solution for every problem.
However, I see increased attention on generative AI as extremely positive! It’s sparked interest in AI among businesses. Many of our customers now come to us considering AI solutions, even if they’re not sure exactly what they need. That’s where we come in. We’re here to help them navigate the complex world of AI and find the right solution for their specific needs.
That being said, GenAI is on the rise for good reasons. It is incredibly versatile and can be applied to a wide range of use cases. For example, while deep learning models have traditionally been used for classifying images, we’ve found that generative AI can achieve similar results. It’s not the primary function of these models, but it allows for quick Proof of Concepts without needing a ton of data to train a model. It’s the best way to try something new with AI without having to dive too deep into it and risk investing too much.
For a long time, Product Lifecycle Management (PLM) has been positioned as the central hub for all data pertaining to product design and development, giving brands the fabled “single source of truth”. In practice, that vision is often left incomplete – not necessarily because of software, but because complex workflows and long-established processes are difficult to change. Today, though, that brand-specific data could be what separates tailored applications of AI from generic, off-the-shelf tools and capabilities. How do you think about addressing that balance and bringing together all the data needed to make the most of AI, from all the different tools people use day-to-day, in one place?
Your PLM has the potential to hold so much data that could be leveraged with AI, but companies struggle to consistently put data in their PLM. It requires changing the ways of working of large teams. Theodo helps turn PLMs and other asset management platforms into trustworthy sources of truth.
PLM holds the data you need to create AI applications that actually set you apart and give you a competitive advantage. Using commercial AI models trained on public data sets, you end up creating the same things as the other users of these commercial models.
We help companies get data into their PLM without having to change the way people work. Designers keep working in the tools they love (Adobe Illustrator, Browzwear, CLO etc) and we extract product information from there and feed it into the PLM.
With the current pace of AI development, the ability to quickly trial new potential applications approaches feels vital. When it comes to deciding on the right AI strategy, or the right model, and then to getting one or more concurrent proofs of concept off the ground, what do you see as the right approach to shortening the time from initial idea to real value?
In our experience, the key to shortening the time from initial idea to real value with AI projects lies in a structured approach.
Firstly, we begin with a Proof of Concept (POC) stage, which consists of two essential steps: framing the use case and building the baseline.
During the ‘Framing the Use Case’ step, we dive deep into the project, defining its value and complexity. We extensively research the industry landscape, identify potential technical challenges, and outline what it would take to create a baseline version of the product.
Next, during the ‘Building the Baseline’ phase, we create the first usable version of the product to demonstrate its value. This is where we take controlled risks, leverage the expertise of our team, and continuously improve performance.
What’s advantageous about this stage is that it typically spans 2-8 weeks, allowing us to build various POCs to determine which one is the most effective based on the performance of the baseline.
Once we’ve identified the highest priority value, we move on to the build and industrialization stage. Here, we proceed with developing and refining the product for widespread use. By starting the build only after assessing the baseline performance, we ensure that the product can deliver value in as little as one month.
This structured approach significantly shortens the time to realise real value. By following this methodology, our clients feel more confidence switching from an initial idea to a product that delivers tangible results.
What do you see as the near-term future of AI – both within your solution(s) and services, and in general? Do you believe it will be a transformative class of technologies the way people expect?
In the near-term future, I believe AI will play a transformative role, both within our solutions and services at Theodo, and in the broader technological landscape.
At Theodo, we see AI as an integral part of the future of software development. Incorporating AI into our development process enables us to deliver value faster for our clients, streamlining processes, and enhancing the efficiency of our solutions.
Currently, there’s a significant hype surrounding Generative AI (GenAI), and while this hype may eventually subside, the truly useful applications of AI will remain. I anticipate that as the hype dies down, people will become more discerning, better understanding what they want and what actually works.
I like to draw a parallel with past technological shifts, such as the invention of the telephone or the anticipation of flying cars in the 1950s. While AI may not revolutionise society in the same way the internet did, it will undoubtedly have a transformative effect. Like previous technological advancements, AI will both eliminate certain functions and create new ones, reshaping the way we work and interact with technology.