We all know AI success stories in business. To be honest, we’ve written quite a few ourselves. Do AI well, and the increase in productivity may be off the charts. Do it wrong and your AI implementation may never even get off the ground. According to Forbes, between 60% and 80% of AI projects end in failure (Schelzer 2022). Only 54% transition from a prototype to a production-ready AI application (Gartner 2022).
Despite increasing AI budgets and the belief of many managers that AI investments will be key to their growth objectives, few companies actually manage to pull it off. This leads to the 64,000-dollar question (quite literally): what separates success from failure?
Figure 1: Not many companies manage to pull off AI projects.
In the course of more than a decade of implementing scalable AI for businesses, we at Robovision have pinpointed two major factors that determine whether or not an AI initiative is likely to succeed. At the start of every project, we ask ourselves two questions:
- Does the company approach AI in a multidisciplinary way?
- Does the company manage the AI lifecycle in a systematic way?
Let’s take a look at these two problems that doom many AI projects. Then we’ll show you how we specially built our Robovision AI platform from the ground up to address these issues.
Failure Factor: Lack of Multidisciplinary Approach to AI
AI is first and foremost a team effort. AI initiatives have a much higher chance of succeeding when companies take a multidisciplinary approach. An approach where multiple roles from different departments work together to contribute their expertise and knowledge to the AI project. This is because every AI project is bound to involve two wildly different parties: data scientists and domain experts.
A data scientist relies on algorithms, mathematics, statistics, design, engineering skills to derive meaningful and actionable insights to model the data in such a way that they have a positive business impact. Obviously, their technical knowledge is fundamental towards building a working AI application.
However, the model is only as good as the data it is fed. That is why input from the second group, those with domain-specific knowledge like process operators and quality managers, is equally if not more critical to the project. Domain experts are not simply the end users of AI applications, they add invaluable insight during development as data scientists often lack the real-world context behind the data. Only domain experts know exactly what the problem is, why it needs solving, and how it should be solved.
Yet all too often, there is a lack of communication between these two parties due to the corporate silos.
Involving both parties is hard. Domain experts cannot be expected to delve deep into arcane code, while programmers are not supposed to understand the intricacies of the company product and the complexities of operational processes. With nothing to bridge this mutual knowledge gap, each department just pursues their own goals. The only point of contact between these silos often comes in the form of intermittent meetings, leading to slow and poor communication, unanswered questions, and buried feedback points.
It should therefore not come as a surprise that AI projects often result in an AI application that is misaligned with business needs, underperforms, or simply fails to do what the domain experts want it to do.
But even if the application somehow ends up working as intended, companies will soon run into another problem: maintenance or AI lifecycle management..
Failure Factor: Lack of AI Lifecycle Management
A common misconception is to think of AI as a one-off investment instead of what it truly is: a long-term commitment. Training an AI model is just the beginning. Maintaining them and monitoring their performance is actually the hard part. Far from being static, the AI models in real-life AI applications have a ‘lifecycle’. They degrade in accuracy over time if they are not retrained regularly. Major reasons for this performance decline include product variability, production drift and production variability.
Managing Product Variability
Product variability relates to variations in the product. For example, agriculture companies regularly introduce new plant varieties on the conveyor belt, while production lines in manufacturing plants change their product offerings over time. This causes problems for AI-enabled applications such as planting robots or quality control systems if they are never trained on these new varieties.
Managing Production Drift & Production Variability
Contrary to product drift, production drift and production variability are caused by the surroundings. The term production drift refers to degrading AI performance caused by changes in the environment, such as ambient lighting, temperature, humidity, or changes in the raw material types used in production. Similarly, production variability happens when the same AI model is deployed in two different production plants. Environmental deviations in plant A and plant B can lead to noticeable differences in AI accuracy, despite everything else staying the same.
In a nutshell, if you do not retrain AI models on a regular basis, it becomes a nightmare to maintain and scale AI across factories. Having teams of data scientists manually retrain the models each time is both inefficient and unviable in the long term as your fleet of AI applications grows. That is why companies need a streamlined AI lifecycle management system—ideally something that can be used by the product operators themselves. This requires a user-friendly platform they can use without coding knowledge. That way, data scientists get out of the picture as soon as the model is trained and move on to the next project. If you depend on experienced data analysts to constantly retrain the models, you will have to go through the hassle of hiring external consultants to keep your AI applications ‘well-oiled’, again and again.
Figure 2: Companies need a streamlined AI lifecycle management system to keep AI applications ‘well-oiled’, again and again.
How It Can Be Done Differently
We have witnessed these two issues—the lack of a vision on a multidisciplinary approach and lifecycle management—play out over and over again in our AI projects over the years. So when we transitioned from an AI consultant to a product company, we designed our flagship Robovision AI platform with two main principles in mind:
- The platform should engage all parties in the process;
- Once properly trained and set-up, our clients should be able to maintain it themselves.
Success Factor: Engaging all Parties
Our core belief is that AI can only work if all parties are engaged in the process. That is why our Robovision AI platform requires no-coding skills to use and it is built from the ground up around the principle of collaborative intelligence.
“Our Robovision AI platform requires no coding skills”
Thanks to this, data scientists and domain experts can join forces to develop AI in a single environment: data scientists perform the initial set-up, the product experts maintain the AI models afterwards. This way, all parties in the project are able to focus on what they know best while also transferring their knowledge within the company.
Figure 3: Data scientists and domain experts can join forces to develop AI in a single environment.
Success Factor: Making AI Maintenance a Breeze
Our Robovision AI platform’s streamlined end-to-end AI life cycle also solves the issue of AI scalability due to maintenance challenges. Our AI platform makes every step as smooth and intuitive as possible, from data management to model deployment, such that the operators themselves can retrain and redeploy AI models in just a few clicks. This opens the door for companies to maintain and scale large numbers of their AI applications without being bottlenecked by external parties. Our platform takes over once the data scientists are gone, enabling the product experts to keep AI projects running by themselves.
“[…]operators themselves can retrain and redeploy AI models in just a few clicks.”
Thanks to our Robovision AI platform, our decade-long experience and the many success stories under our belt, we believe that we are uniquely positioned to not only make your AI project succeed, but to keep it succeeding.
It is that vision that ensures our clients their AI project is not amongst the 87% that fail. Rather, it is the solution that helps them scale.
Learn more about Robovision AI platform.
- Schmelzer, R., 2022. The One Practice that is Separating the AI Success from the Failures [online], 14 August 2022. Available from: https://www.forbes.com/sites/cognitiveworld/2022/08/14/the-one-practice-that-is-separating-the-ai-successes-from-the-failures/?sh=15e85cf017cb [Accessed 3 October 2022]
- Gartner. 2022. Gartner Survey Reveals 80% of Executives Think Automation Can Be Applied to Any Business Decision [online], 22 August 2022. Available from: https://www.gartner.com/en/newsroom/press-releases/2022-08-22-gartner-survey-reveals-80-percent-of-executives-think-automation-can-be-applied-to-any-business-decision [Accessed 03 October 2022]