7 Crucial Mistakes CTOs Make When Implementing AI

With the potential benefits that AI offers, the world is moving towards an AI-transformed future. Many leading companies and technology giants have already successfully implemented AI strategies, while others are playing catch up. In this uncharted race towards AI transformation, there can be many challenges and obstacles that enterprises face. However, there are a few crucial mistakes that are important to avoid in your AI automation journey.

Not Clearly Defining KPIs or Identifying Goals

The first and most important step of implementing AI in enterprises is to define specific, measurable, achievable, relevant, and time-bound (SMART) goals. While this may seem obvious, the depth required in this step can often be overlooked.

There are many questions and issues that need to be addressed in this step. As an example, just a few of these questions include: Which business problems need to be solved? Is it a problem that can and should be fixed with AI? What are the required functionalities? How will it affect the bottom line of the company? How long will it take to reach ROI? What are the business needs and data constraints? How does this relate to the company’s overall AI strategy and transition? What budget is needed compared to what resources are available?

Analyzing and answering these types of questions are critical to identifying the right approach to implement AI in any organization. Without answering these questions and defining SMART goals, many enterprises might look for a fast and cheap off-the-shelf, one-off AI solution without thinking of long-term ROI or how it will impact the company’s overall AI transformation. This could damage the project itself, business efficiency, and potentially even the company.

Not Properly Scoping Out Projects

The solutions that AI can offer enterprises can be very exciting as the technology provides limitless possibilities for advancement, efficiencies, and accuracy. So, maintaining restraint to not tackle a wide variety of problems at once with AI can be challenging. But it is worthwhile to remember that just like Rome was not built in one day, an enterprise cannot transform its entire business and all its process with just one AI project. AI needs to be implemented gradually and incrementally to achieve a cumulative effect. With AI, there is such a thing as too big to manage. 

When implementing AI, it is important to start small, with a specific use case and invest in the fundamentals to create a good foundation. This is an opportunity to learn what works and what does not, to learn from mistakes and implement better AI solutions in the future, to refine and improve. It is often recommended to start with a single use case that would be considered a low-hanging fruit in order to increase the probability of setting a positive precedent that proves added value.

Implementing AI

Investing in a One-off, Off-the-shelf AI System that is not Sustainable for a Full AI Transformation

When implementing their first AI use case, many companies look to a one-off, off-the-shelf AI solution that only solves a very specific problem instead of a comprehensive solution that solves a complete process. While use-case specific AI can be successful at quickly solving very specific problems, this approach is not sustainable for a long-term, full AI transformation. So even though it is best practice to start with solving a single use case, there still needs to be a long-term approach for how AI will work once it has been expanded to new use cases, otherwise it can actually make reaching AI maturity more challenging. A short-visioned approach can create a fragmented internal environment of data and processes instead of creating a connected ecosystem that works together to achieve improved continuity and synergy. For this reason, it is important to consider the long-term AI transformation goals and find a solution that can help the entire company reach AI maturity. When using such a comprehensive approach, companies should look for an AI platform that can expand with their AI goals, that can both be easily adaptable to solve individual use cases, but also universally applicable so it can address any other business process in the future.

Not Having the Right Infrastructure

AI is incredibly resource intensive. It requires a broad ecosystem and a high-performance computing infrastructure, comprised of networks, GPU-enabled servers, security, and flash storage. If there is a lag in any of these components in the infrastructure, it can damage the performance of AI. One workaround for this can be moving to cloud-based AI instead of on-prem, depending on regulations.

Another type of resource that needs to be built up for AI is human capitol, especially if planning to build AI in house. This type of feat will typically require developing an AI center of excellence, which can be very costly and timely – especially since market supply is not meeting demand. One way to avoid the need of developing an AI center of excellence is by using AI platforms that already have pre-developed AI capabilities and an intuitive user interface that empowers any stakeholder across the organization to take part in the human-machine feedback loop.

Lastly, while data is not typically considered infrastructure, it is a key component that is required for reaching successful production. For AI to properly leverage data, enterprises often need to collect, clean, normalize and prepare their data. However, there are some advanced AI platforms that are robust and strong enough to perform under extreme data scenarios, such as, noisy, messy, and even limited amounts of data.

Implementing AI

Spending Too Much Time in POC environment

The standard AI automation journey is to start in a proof of concept (POC) environment, in which the data is clean and stable. However, the POC environment is drastically different from the often-unstable production environment, where data is messy and noisy. So, after engineers invest a significant amount of time and money stabilizing their AI model in the POC stage, they often realize too late that the model does not properly function in production when faced with real-life business requirements and other constraints. It is in part due to this POC gap that many AI projects never reach production, much less ROI positivity. This failure rate costs businesses a significant amount of money. In 2020 alone, out of $50 billion that was spent on AI, $43 billion was wasted on projects that NEVER reached production. (MITSloan Management Review, BCG, October 2020, Expanding AI’s Impact With Organizational Learning, IDC, August 2020, Worldwide Artificial Intelligence Spending Guide, Venturebeat, July 2019, Why do 87% of data science projects never make it into production?) For this reason, many AI experts are suggesting to avoid the POC environment completely and instead test out AI directly in the production environment.

New call-to-action

Only Developing a Trained Model versus a Full System

There might have been a time when AI solutions could function as just a ML code, but as AI continues to mature and is required to solve more complex business problems, the solution itself also needs to become more robust. A trained model is only one part of the complete puzzle.

A full AI solution requires many additional pieces to ensure that it stays on the rails in the long term. For instance, on-going monitoring that can detect environmental changes, such as out of distribution data, or identify model performance degradation. Another piece that is key is the human-machine collaboration, such as a feedback loop that can help establish trustworthiness. When all of the pieces are assembled into one end-to-end solution, AI projects have a significantly increased probability of achieving long-term performance and reaching positive ROI.

Allowing Sunk Costs to Drive Business Decisions

When AI fails to reach production, there can be a temptation to continue down the same course and throw more and more money, time, and resources at the solution to try to ensure that the previous investment does not go to waste. But as money continues to poor into AI without achieving any ROI, this can end up resulting in organizational hesitancy and fear of the technology.

However, if organizations treat their previous investment into a failed AI solution as a sunk cost and take a step back from the project, they could instead see the failures as lessons learned. This could allow organizations to take a new and fresh look at how AI should be implemented in their enterprises. For instance, this might mean transitioning from a self-build AI solution to using a vendor for a hyper-customizable AI platform. The main goal is to avoid getting stuck with an AI project that is doomed to fail.

With any investment that has potential for a great reward, there is also bound to be a great risk. This is also true with AI. However, these risks can be reduced by avoiding many of these typical mistakes. As the AI industry continues to mature, it is crucial that we collectively learn from the past so we can continue to accelerate the productization of AI.

Need to solve core business problems with customized AI solutions?
See how you can solve individual use cases or achieve a company-wide AI transformation using one platform and gain a competitive advantage.