AI has the power to transform digital platforms, allowing them to scale up and take on new markets far more efficiently and cost-effectively than ever before. However, unlocking its potential is impossible without humans in the loop, says Invisible Technologies’ Partner and Vice President of Operations, Jean-Paul Biondi.
“AI indeed has the capacity to elevate online businesses to unprecedented levels, but it's not a solitary journey,” he explains. “The real power of AI unfolds when it's paired with human expertise and oversight.”
The Challenge of Unstructured Data
When digital platforms move into a new market or launch a new product, they typically need to take onboard a significant amount of data. This often includes product descriptions, vendor and pricing information, inventory data, compliance data, and customer reviews and feedback.
“We often work with digital platforms, such as marketplaces, that are looking to scale rapidly,” Biondi says. “However, by their nature, these businesses are very decentralized, with thousands of products often coming from thousands of different providers or sources, and none of it standardized.”
Biondi says that, while AI can be used to take on and analyze this data, AI algorithms typically rely on patterns they identify in their training.
This means they’re typically great at dealing with data when it is presented in a consistent format, but can struggle where it is inconsistent - particularly where there are a lot of outliers or anomalies to which AI can’t apply what it has learned.
As a real-life example of this, Biondi cites Invisible’s work with a well-known food delivery app that needed to scale quickly during COVID-19.
“The platform found itself in a position where it needed to onboard tens of thousands of restaurants around the world in a very short space of time,” Biondi explains.
“As you can imagine, restaurants around the world don’t always have a lot in common when it comes to how they display or sell their products. Their menus were all in different formats and there was no consistency; some even had handwritten annotations.”
“There was no way AI could take on and do this work alone. The error rate would have been extraordinarily high.”
According to Adam Haney, Invisible’s VP of Engineering, the solution lies in mapping out every single process involved in onboarding clients, and allocating some tasks to humans, while training AI to perform the rest.
“By breaking down every single process into its components, we could make a call on which tasks could be better - and more efficiently - completed by AI, and which were better performed by humans. We could also analyze what level of human intervention was needed.”
A Hybrid Solution
Haney says it’s not only the nature of the data that can pose a problem for AI. Sometimes interpreting that data can challenge its limits too. This is another area in which human knowledge and intuition are still very much needed.
“No matter how well it is trained initially, an AI Model might not, for example, understand the cultural nuances of a particular product or statement or be able to capture the subtleties and variations of something in the market,” Adam Haney says.
This means that, even though it may process data accurately, it may fail to grasp the broader context or implications, which can lead to misinterpretations or overly literal interpretations, particularly in cases where human communication is involved.
“Without significant human intervention, it’s likely to misinterpret product descriptions or customer interactions. It could also fail to understand the sentiment of customer reviews or to effectively localize content.”
“This can have a real impact on customer engagement and satisfaction, especially where an app is being rolled out or used across cultures or regions,” he says.
The Nature of Generative AI
Another reason humans are still needed in the loop goes to the fundamental - and revolutionary - nature of AI.
To date, most technologies have been ‘deterministic’. That means they operate like a calculator, whereby if you enter certain inputs it will give you an accurate answer. However, AI is that it is ‘probabilistic’, which means it provides ‘likely’ rather than accurate answers and incorporates some lament of randomness in its approach.
“With a lot of AI models you can give exactly the same prompt but get very different answers on the other side,” Haney explains. “The biggest issues can come where there is insufficient data or no data at all, in which case it often produces its own based on what it is trained to believe is most likely.”
“To some extent, this replicates human behavior but it also means, just like in a human-based workplace, you are going to have to have some degree of oversight,” he says.
‘The Last Mile’ of Data
Jean-Paul Biondi says that many online businesses would be familiar with the concept of the ‘last mile’ when it comes to deliveries, which still largely involves human-based interactions. However, in many ways, both the strengths and limitations of AI mean that the same concept can also be applied to data.
In other words, while AI can efficiently process and analyze vast amounts of data in the same way logistics systems manage the bulk of the transportation journey, the final stages of data processing and interpretation also benefit from human intervention.
"Much like the last mile in delivery services requires a human touch to navigate the final complexities, AI in data handling encounters its own 'last mile' challenges,” he says.
“AI can process the bulk of data efficiently, but the nuances, the context-specific interpretations at the end of this journey, often demand human insight. It's in this final stretch where the most crucial decisions are made, and where human oversight becomes indispensable."
Haney also cautions marketplaces to view AI as a partner rather than a panacea.
“While AI offers significant capabilities, many businesses, especially those not at the scale of large tech companies like Google, would benefit from partnering with a firm that can provide both AI solutions and the necessary human oversight,” Haney explains.
“This approach is especially relevant for companies that may not have the extensive resources or specialized expertise to develop or manage AI solutions internally,” he concludes.