Artificial intelligence is very prevalent in movies and science fiction, from sentient beings that are able to walk, and talk, and live with humans like the characters from Westworld, Star Wars, or Star Trek. In reality, however, we are a far way away from the dream of sentient machines that we see and read about in science fiction. So much of today’s AI systems are doing much more mundane things that aren’t getting the attention or interest of the press and media.
However, interest and investment in AI remains strong, and even if AI is unable to live up to the fantasies of science fiction, vendors are riding the hype wave of AI and promising capabilities that AI systems might not be able to deliver. While vendors are doing their best to deliver these capabilities, the challenge is that adopters and end users sometimes themselves get caught up in the hype as well. This mismatch between what is being promised by the technology vendors and what is actually possible is a common reason why AI projects fail.
The reasons for these failures could be due to a product mismatch, overhyping or overselling on AI capabilities, or selling solutions a customer doesn’t actually need. And, it’s not only inexperienced people making mistakes and falling for this. Even experienced teams suffer AI project failures. If you blindly trust the vendors you may end up with a product that doesn’t actually fit your needs and become an AI failure statistic yourself. Despite all these challenges, organizations are still seeing success with AI systems, but not necessarily from the technology companies that you might be hearing about in the news.
Tens of Thousands of AI vendors selling AI Solutions
Research firm Cognilytica has tracked over 20,000 vendors in the AI market, across dozens of different application areas from natural language processing to AI-enabled hardware and industry-specific applications. In fact, over 70% of the vendors in the market deliver solutions specific to individual industries from finance to healthcare, cybersecurity to agriculture, and beyond. As a result, organizations looking to implement an AI solution have a very large field of options to select among. Weeding through this huge list of potential vendors can be daunting.
One approach organizations are taking is to narrow down vendors into their specialties and find which one has the best solution for your specific needs. This sounds simple enough, but the AI markets are in a state of constant change. What is available today from one vendor might not be available tomorrow as those vendors evolve their offerings, merge with other vendors, or even close shop. The last thing organizations want to do is choose a solution that is not a good fit, be a good fit today but a bad fit tomorrow, or might not even be around in the future. Vendors often change or enhance their offerings very quickly due to company wide pivots, acquisitions and mergers, going public, and rapid company growth. All these factors can have an immediate effect on their product offerings that may lead to a mismatch with their customers needs over time.
Assuming that vendors are producing offerings that can do what they claim, the challenge is that many customers are not even sure about their current or long-term AI needs. Most companies are still in the early stages of AI implementation and have not figured out what their ongoing AI technology needs will be. Furthermore, solutions themselves are often not directly comparable. For example, AutoML solutions provide a great opportunity to reduce the cost and complexity of ML model development by automatically doing algorithm selection, model tuning, data preparation, and more on behalf of users who might not have those capabilities or who are looking to speed up the process. However, not all AutoML products do the same things. Some work with quantified data while others work with image data. If you choose an AutoML vendor who specializes in quantified data when you actually need a solution that works with image data, you will end up with a mismatch.
This lack of experience combined with the constantly changing vendor landscape and evolving products makes vendor selection a really difficult process for organizations looking to implement AI as it continues to evolve. There are also other challenges in vendor selection such as choosing a solution that works in the cloud when you need it to be on premise, not realizing the difficulties in data preparation and cleansing, and other issues that cause AI projects to fail even before they’ve gotten off the ground.
The Problem of Pseudo AI
While many vendors are doing as they claim with machine learning and AI solutions, there are other companies that are using humans to fill in the gaps of what AI is unable to do. Known as “pseudo AI”, this approach applies when a company claims their solution is powered by AI but it is actually being provided by humans. In this way, they are claiming that AI is providing the solution, but in reality it’s outsourced to low-cost humans to perform those tasks. While this might be a suitable solution for some, the challenge comes when the vendor doesn’t disclose that a human might be performing AI tasks.
This is especially an issue when it comes to data privacy regulations, security, and customer trust. If you’re handling sensitive information such as medical files, customer records, or images this could become a huge problem. When a vendor doesn’t disclose the use of humans pretending to be machines pretending to be humans, it is very misleading and can erode trust. Customers are now just beginning to ask vendors about human-in-the-loop for the AI solutions. Customer-led vendor selection questions are now increasingly becoming a mandatory part of best-practice AI methodologies such as CRISP-DM and CPMAI, in which business understanding requirements mean that vendor offerings need to be pared down to the minimum that need to meet immediate business requirements.
Highly Specific AI Solutions, Open Source, and Self-Built Solutions Are Winning
One of the lessons learned from companies succeeding with their AI initiatives is to ensure that the vendor solution is an immediate fit for the AI project. This means taking a step back from committing to vendors too early in the AI project process. The second lesson learned is that organizations need to focus more intensely on the early data-centric needs of their AI projects. If over 80% of AI projects are focused on data collection, preparation, and engineering, then it makes sense to focus most of the time on vendor selection on the data-centric needs, rather than on the more “exciting” aspects of model training, development, and deployment.
In fact, organizations who are AI successful are spending more time with firms in the data labeling and preparation markets than they are with cloud or on-premise AI modeling tools. And there are hundreds of vendors just in the data preparation markets. For example, Cognilytica has tracked over 150 vendors in the data labeling markets alone. These vendors all offer different capabilities depending on the nature of the data to be labeled, technology requirements, data labeling workforce and labor requirements and more. What’s even more surprising is that there are billions of dollars being spent on these solutions, and these companies are collectively worth billions of dollars, with some already in the venture capital unicorn territory.
While companies like Amazon, Google, Microsoft, IBM, Nvidia, and others have the most attention from the media and press, the reality is that most of the successful AI implementations are based on open source technology that all the major cloud and on-premise vendors are using. You might be surprised to hear that when asked what is the most common vendor that AI implementers are using, it’s not a commercial technology vendor at all. Instead, these companies reply in surveys that they’re building their own AI solutions on top of open source technology or in-house capabilities.
In O’Reillys 2021 AI Adoption Survey, when asked about the technology platform on which their companies were implementing AI, respondents replied with solutions such as TensorFlow, scikit-learn, PyTorch, and Keras with the top 4 responses, with the only notable vendor-specific response (AWS SageMaker) trailing in the 5th spot. This doesn’t mean that companies aren’t using vendors for those open source implementations, not at all. This does mean that they aren’t focusing on the vendor names because the selection is toolkit based, and therefore the vendor is not as relevant. This is why the top AI vendor that a specific company might say is the key to their AI implementation is not a vendor you have heard of, but rather might be their own internal tool, a data engineering or preparation tool, or an industry-specific vendor producing tools specific to their domain needs.
