Top Challenges of Implementing AI Models And How to Overcome Them
AI models can deliver significant benefits once deployed by any organization. But how successful have companies been in leveraging AI so far? Here’s a look at the top challenges organizations face when implementing AI models for production.
An excessive dependence on manual processes can seriously impair the effectiveness of an IT network composed of several siloed apps and databases. Companies are automating various processes to reduce these risks. But what happens when using AI itself adds complexity? Let’s look at several deployment challenges for AI, steps that leadership can take to incorporate AI models into production successfully, and some common AI use cases.
Data scientists spend a lot of time cleaning and munging data before they can begin constructing models, which is probably the most exciting part of their work. You may observe how the prediction performance increases after creating the characteristics and testing several models. The work is not over even if you now have a highly effective model. A critical phase in the workflow is the deployment of your AI models because this is when they start to benefit your business genuinely.
Implementing AI is the next stage after data structuring, which most organizations find quite challenging. Let’s find out why implementing AI is so challenging, what specific skills teams and leaders lack while doing so, and various AI applications.
See More: Why NLP is the Next Frontier in AI for Enterprises
Challenges Implementing AI Models in Production & Its Common Use Cases
What makes AI implementation so difficult?
Mike Loukides, VP of emerging tech content strategy at O’Reilly Media, says, “Most of the focus in developing AI systems has been on building models. As Andrew Ng and others have pointed out, modeling has become much easier with autoML tools. Developing new models is a research topic that most companies smaller than Google don’t need to think about.”
“The problem is and has always been data. Since we started talking about big data, we have said that 80% of the work is in cleaning data. That hasn’t gone away.”
– Mike Loukides, VP of emerging tech content strategy at O’Reilly Media
Loukides adds, “No effort on automation has made it any simpler. It is crucial to get data, ensure it’s good, understand its provenance and consequences of acquiring and utilizing it, clean it, and design pipelines to transport it to your application.”
Samir Agarwal, EVP of global products and channels at ElectrifAi, believes that there are many factors that make AI implementation difficult. “The sheer buzz of AI and every vendor loosely using the term makes AI sound like something that can be sprinkled on or attached to a solution.” Here are a few factors, according to Agarwal, that make the AI implementation difficult:
- The data is often incomplete, inaccessible, unclean, and unstructured, making it difficult to use for ML.
- Not clear on what features should be considered while applying machine learning.
- The art and science of training the machine and backtesting to ensure the machine will produce the best results for data are not seen yet.
- Fitting all of this into the operational flow of existing systems.
- Adoption of something new is uncomfortable.
Loukides highlights another difficulty on the deployment side. “In the last few years, we have made a lot of progress in IT operations. This goes under continuous delivery, DevOps, or something like that. AI doesn’t fit into those models well,” he believes. For example, any good software developer will use a source code repository like git. But what does that mean for AI, where the training data and the model built from that data are as important as the source code, if not more so? “We don’t have good versioning systems like git for data or for models and other artifacts. We will, but we are not there yet. How do you think about continuous delivery when retraining a model might take hours or even days?”
These problems will be solved eventually, but right now, people need to understand that building a successful AI project is a lot more than simply coding up a model, adds Loukides.
See More: Is the Skills Gap Putting AI implementations at Risk? Five Ways the Gap Can be Bridged
What precisely are teams and leaders lacking while deploying AI?
Agarwal mentions that understanding what it means to adopt AI and ML is frequently challenging. “Where does one start? Does one hire data scientists and how many? Does one need to buy a platform? If yes, which one?” There are also questions about whether the data is accessible, and is the business problem clear? He thinks there is a substantial cost and risk of failure, which can impede securing AI’s benefits.
Kjell Carlsson, head of data science strategy and evangelism at Domino Data Lab, says, “While all big tech companies have adopted AI and leverage it more than their non-tech counterparts, they do not leverage AI to more than a fraction of its potential.” All have large parts of their business that have yet to leverage AI — even the most AI-advanced parts of their businesses are still constrained due to a “lack of professional-grade tools, scalable infrastructure, immature processes, and a lack of skilled data scientists.”
On the other hand, Loukides draws attention to the fact that so many individuals in computing seem to have an aversion to mathematics. “Just the other day, I was reading a few online posts that discussed how programmers don’t need to know math. Maybe that’s true for some specialties, but it is definitely not true for artificial intelligence and machine learning. I can agree that you don’t really need to know linear algebra if you are not developing new algorithms. But anybody working with data needs a really good grasp of statistics.”
In addition to having strong mathematics skills, Loukides asserts that you also need to understand how statistics apply to everyday situations. He further states, “If there’s a bump in some statistic derived from a dataset of 100 people, will that bump still be there when you look at 10,000 people? If there’s a 1% year-over-year change in a metric you care about, is that noise, or is it important? Those abilities are crucial for everyone in software development, but they are especially important for AI.”
Giving AI its long-awaited value – the possible use cases
Carlsson believes, “There is no major, successful US enterprise today that is not leveraging AI technologies, but it is mostly hidden in apps such as the voice of the customer analysis, intelligent document processing, customer service automation, intelligent search, recommendation engines, scheduling applications, etc.”
“By the same token, there is no enterprise in the world that has exploited more than a fraction of the business value they could get from using AI technologies.”
– Kjell Carlsson, head of data science strategy and evangelism at Domino Data Lab
Every company needs not one but multiple AI strategies, he adds. “The different methods, data and use cases for different types of AI applications are so distinct that there is little benefit and many pitfalls to creating a single AI strategy.” As per Carlsson, companies with separate strategies, e.g., for natural language understanding (NLU), computer vision, chatbot, and recommendation engines, will be more successful than those that attempt a one size fits all approach.
Another good news is that securing the AI value is not very hard if it is in the right hands. As AI is nothing more than arithmetic, Agarwal advises defining one’s business challenge precisely. “Optimizing revenue is not the same as optimizing profits. Explore possibilities of leveraging pre-built ML/AI-driven solutions. If the vast body of work is done and tested, a variant with very acceptable outcomes can produce value in as few as four to eight weeks, and at a small fraction of the cost of hiring an army of data scientists.”
While talking about the AI use cases, he says that AI can be applied to any business area – “savings on spending and reducing vendor risk in procurement to dynamic pricing, inventory forecast, sales forecast, and better scheduling.”
Loukides outlines other use cases: “Optimizing the use of heating and cooling systems, figuring out whether the dishes in your dishwasher are clean yet, and things like that. There are lots of interesting applications in agriculture: AI systems that can detect plant diseases or tell you when the tomatoes in a greenhouse are ripe.”
Some IT-related use cases include:
- IT preventative maintenance
- Threat recognition
- Root cause investigation
- Noise mitigation and event correlation
- Resolving problems without human intervention
- Estimating the usage beforehand
Conclusion
The idea that your business isn’t ready for AI is not true. No matter how simple, every organization may benefit from the wide variety of approaches and uses available. Additionally, AI is more crucial than ever. You can identify, invent, and react to the quick changes in your business environment with the help of AI models, and those changes will only get quicker. Although using AI to transform your company is challenging, avoiding it is not a viable option.
Can you list some innovative use cases that are gaining momentum? Let us know on LinkedIn, Facebook, and Twitter. We would love to hear from you!
MORE ON AI TRENDS
