Artificial intelligence technologies have reached impressive levels of adoption and are seen as a competitive differentiator. But there comes a time when technology becomes so ubiquitous that it is no longer a competitive differentiator – think of the cloud. Going forward, the organizations that succeed in AI will be the ones that apply human innovation and business acumen on their AI basis.
This is how the challenge is identified in a study published by RELX, which finds that the use of AI technologies, at least in the US, has reached 81% of companies, an increase of 33 percentage points from 48% since a previous RELX study in 2018. They They are also positive about AI delivering the goods – 93% report that AI makes their business more competitive. This ubiquity can be the reason why 95% also report that finding the skills to build their AI systems is a challenge. In addition, these systems can be potentially flawed: 75% worry that AI systems could potentially introduce the risk of workplace bias, and 65% admit that their systems are biased.
So there is still a lot of work to do. It comes down to the people who can make AI happen and make it as fair and accurate as possible.
“While many AI and machine learning implementations fail, in most cases it’s a minor issue with the actual technology and more about the environment around it,” said Harish Doddi, CEO of Datatron. Moving to AI “requires the right skills, resources and systems.”
It requires a well-developed understanding of AI and ML to deliver visible benefits to the company. Although AI and ML have been around for many years, “we are still barely scratching the surface to uncover their true abilities,” says Usman Shuja, general manager of connected buildings for Honeywell. “That said, there are many valuable lessons to be learned from the mistakes of others. While it’s undoubtedly true that AI can add significant value to virtually any department across any business, one of the biggest mistakes is a company can commit to implementing AI in order to implement AI without a clear understanding of the business value they hope to achieve. “
In addition, AI requires skilled change management, Shuja continues. “You can install the most advanced AI solutions available, but if your employees can not or will not change their behavior to adapt to a new way of doing things, you will not see any value.”
Another challenge is bias, as expressed by many leaders in the RELX survey. “Algorithms can easily be biased based on the people who write them and the data they provide, and bias can happen more with ML as it can be built into the base code,” says Shuja. “While large amounts of data can ensure accuracy, it is virtually impossible to have enough data to mimic real-world applications.”
For example, he illustrates, “if I researched recruiting collegiate athletes for my professional lacrosse team and I discovered that most of the players I hear about are Texas Longhorns, it could lead me to conclude that the best “lacrosse players attend the University of Texas. However, this may just be because the algorithm has received too much data from one university, thus creating a bias.”
The way the data is set up, and whoever sets it up, “can inadvertently sneak into the algorithms,” Shuja says. “Companies that do not yet think through these implications need to put this at the forefront of their AI and ML technology efforts to build integrity into their solutions.”
Another problem is that AI and ML models simply become obsolete prematurely, as many companies found out, and continue to figure it out as a result of Covid and supply chain problems. “Having good documentation that shows the model’s life cycle helps, but it’s still inadequate when models become unreliable,” says Doddi, “AI model management helps bring accountability and traceability to machine learning models by getting practitioners to ask questions like” What were the previous versions “like?” and ‘Which input variables enter the model?’ ‘”
Governance is the key. During development, Doddi explains, “ML models are bound by certain assumptions, rules, and expectations. Once implemented in production, results can differ significantly from results in development environments. This is where management is crucial once a model is operationalized. There needs to be a way to keep track of different models and versions. “
In some cases with AI, “less is more,” Shuja says. “AI tends to be most successful when paired with mature, well-formatted data. This is mostly in IT / enterprise data, such as CRM, ERP and marketing. But when we move into areas where the data is less coherent, such as with operational technology data, it is here that achieving AI success becomes a bit more challenging.There is a huge need for scalable AI in an industrial environment, for example by using AI to reduce energy consumption in a building or an industrial plant – an area with great potential for artificial intelligence. One day soon, entire companies – from the factory floor to the boardroom – will be connected; constantly learning and improving from the data it processes. This will be the next big milestone for artificial intelligence in the company. ”