Note: This post originally appeared as: The More Things Change, The More They Stay the Same: A Lesson on Today’s AI for the Masses on our COO, Carlos Meléndez, Under Development blog at InfoWorld, and is reprinted with permission from IDG.
Before the 1800s and the invention of the power loom, clothing was hand-made at home, by necessity. Once clothing was mass produced in the 1900s thanks to the power loom and the Industrial Revolution, everyone wanted the ready-made styles that were offered. Yet, fast-forward to today, and it is once again in vogue to wear distinct styles. Only the wealthiest among us have custom tailored clothing. The rise of the textile industry only goes to prove that the more things change the more they stay the same.
The same evolutionary cycle can be seen in the software industry. When software was first introduced in the mid 1900s it was custom-developed, the secret weapon for large enterprises looking to automate critical tasks. As the industry matured, however, and software became a commodity, companies turned to best-of-breed software to quickly and cost-effectively gain the benefits that software delivers. And today, once again, while off-the-shelf software is still being used, companies understand that in order to have a competitive advantage, they need to custom develop solutions to meet their own unique business needs and differentiate themselves.
Ripping a Page from History – AI for the Masses
If history proves itself correct, the same evolutionary cycle is unfolding when it comes to AI. Large vendors, such as Amazon, Microsoft and Google, among others, have rolled out Machine Learning as a Service (MLaaS) offerings that provide companies with cloud-based and ready-made data sets, as well as trained algorithms that provide pre-built AI. But the same companies embracing ready-made AI may revert to custom-built solutions as the market becomes too commoditized.
Machine Learning as a Service
The appeal of MLaaS tools is understandable. For most businesses, machine learning can be an extremely complex and costly process, requiring tons of man hours, but with MLaaS you can jump-start an AI initiative without much investment or manpower.
The three leading cloud MLaaS offerings today are Amazon Machine Learning services, Azure Machine Learning, and Google Cloud AI. These solutions promise fast algorithm training and deployment with little to no data science expertise. They can leverage large data-sets such as facial recognition or natural language data, which would be difficult and costly to acquire for any given company on its own. When these models are made publicly available by these vendors and others, anyone can analyze any volume of data without needing to train the algorithms themselves.
Building Vs Buying AI: The Pros and Cons
While MLaaS offerings are gaining traction, companies need to carefully assess their options and not just jump on the prebuilt bandwagon. There are benefits to custom-designed AI that can provide companies with distinct competitive advantage including:
- Cost. While it may be cheaper to use ready-made datasets and algorithms to address a one-off, specific business need, most of the larger vendors operate on a subscription model. While it may only cost a few hundred dollars per prediction each month, costs could spiral out of control in the long term when you have several algorithms to develop and data to train. Additionally, companies have to pay every time they build an API between the AI algorithm and their business application.
- Expertise. Regardless of how easy an MLaaS vendor says it is to be up and running with its prebuilt AI offerings, building an API between your solution and the pre-built algorithm still requires that you have internal IT specialists who can integrate and operate it.
- Data. The data libraries offered by the big MLaaS vendors provide massive amounts of data, which would be difficult for any company to ever acquire on their own, so it often makes sense to leverage those data-sets to jumpstart your AI project. But to get any type of distinct benefit from AI applications, companies need to augment those data-sets with their own. To make a real difference, the data needs to be customized to meet regional, corporate or cultural variances depending on your unique business challenge. In many cases, a Google prediction engine or classification engine can only take you so far – it needs to be trained with your data to be truly effective.
To be fair to the Googles and Microsofts of the world, it would be impossible to create data-driven predictive analytics localized to all specific regions or companies – it would simply require too much data that would not have broad appeal.
There’s no question that there are clear benefits to choosing pre-built MLaaS AI offerings in terms of short-term cost savings and fast ramp-up. You might even argue that for certain types of universal problems, like face recognition and object detection, it makes sense to buy off-the-shelf.
But, as history has proven, when something becomes a commodity, everyone has the same thing and any competitive advantage a company hopes to gain through AI can quickly be lost. In truth, there really can be no such thing as a one-size-fits-all approach when it comes to AI. Every company is distinct, with its own culture, assets and business challenges, and AI initiatives need to reflect this uniqueness in order to be truly effective.