Rethink conventional IT methods and some popular wisdom to effectively integrate AI into the business processes. Consider the following suggestions.
Artificial intelligence (AI) has made its way into the business world, rapidly progressing from a pipe dream to a reality. According to a recent O'Reilly survey, the majority of organizations (85 percent) are either implementing or testing AI, with more than half using AI in development or for research.
These efforts are still in their early stages, despite their rapid expansion. And the aches and pains of maturation are already visible. According to O'Reilly researchers, "companies need to do something to place their AI activities on stable ground." Whether it's controlling for common risk factors like bias in model growth, missing or poorly conditioned data, or the potential for models to degrade in production, adopters will have their work cut out for them as they work to create reliable AI production lines.”
AI may not fit neatly into the same processes and approaches that the IT department has used previously. The best practices and common-sense methods used to evaluate, assess, introduce, and scale non-learning systems can not always extend to learning systems. They can, in some instances, backfire.
1. Take it easy.
There is a hasty rush to AI-ify the company in certain companies, which can be risky if left unchecked. According to Dr. Jerry A. Smith, vice president of data sciences, machine learning, and AI at Cognizant, modern AI has a high IQ but a low EQ. Both are needed for true intelligence.“You're essentially releasing a psychopath into the environment if you collect data and use AI to evaluate and learn from it without emotion and at scale.”
Smith advises IT leaders to take their time to make sure they have human conversations on what they're trying to do with AI early on.
He adds that executives frequently want AI to save them. “However, if they do not provide the proper structure and strategy, it will ultimately affect them.”
2. Prioritize knowledge and community over tools.
“It's not shocking that technology is always where businesses start when seeking to innovate,” says Shawn Rogers, vice president of analytic strategy at TIBCO. “However, ignoring the human and cultural aspects would almost definitely doom you to failure.”
“New skills, as well as a culture that fosters acceptance and action to get to the importance of AL and machine learning (ML) technology, are needed to drive progress in AI,” Rogers adds. Success necessitates a well-balanced strategy.”
3. Make iteration a priority.
Businesses that want to get started with AI should start with a use case in mind. However, according to Ashish Thusoo, CEO and co-founder of open data lake company Qubole, most AI and ML use cases evolve in a very iterative manner over time. According to Thusoo, who previously co-founded Apache Hive and developed the Facebook data platform, “it is crucial that organizations invest in the capability to conduct continuous data engineering and have both SQL and programmatic access to train and deploy models.”
4. DevOps is insufficient
The majority of forward-thinking IT companies have already jumped on the DevOps bandwagon. This is necessary, but not appropriate, for AI adoption. According to George Mathew, client partner for technology services at Fractal Analytics, organizations must add MLOps. “This integration needs to be designed early in the application lifecycle and pursued through all phases,” he says.
Organizations must, for example, consider the retraining of AI models that will occur in development. “This means that additional pipelines would need to be installed to equate the AI models' observations (such as forecasts) with the real numbers obtained from the field a few weeks or months later,” Mathew explains.
5. Get ready for the scale
Early forays into AI usually rely on a few models and a limited set of data. Those efforts, on the other hand, can easily spiral out of control. Managing hundreds of models in development and numerous authoring environments for developing data science teams poses new challenges, according to TIBCO's Rogers.
6. Look for prejudice, but not in the places you expect to find it.
“The easy part of AI is understanding the relationship between inputs and outputs,” says Cognizant's Smith. “A lot of people blame AI for its prejudices. They want to be certain it isn't skewed." You don't want biased AI algorithms making loan decisions, for example.
Human underwriters, on the other hand, may be biased. Smith says, “You have to get your hands around the human intelligence part of this to make sure the individual designing AI systems isn't biased in the modeling.” Bias mitigation cannot begin or end at the data level.
7. Don't delegate authority to data scientists.
The most powerful AI systems are designed to help people. Smith explains, “If you want a structure to serve human beings, it needs to be human-centric.” “You need sociologists and psychologists (people who understand consumer behavior)” (people who know how your customers interact in society). AI is far too important to be left solely in the hands of data scientists.”
8. Get ready to offer an explanation
Explainable AI (XAI) – techniques that enable humans to comprehend, trust, and manage AI – is gaining traction. As a result, according to Mathew of Fractal Analytics, some IT companies will face regulatory audits inquiring about the AI-model training runs, including what data sets were used, how the algorithm was evaluated, and what model metrics were created at each point.
“These elements must be collected and stored for the duration of the models' development – and beyond, in some cases,” Mathew explains. “The solution architect must prepare the architecture to meet these criteria, the project leader must include these steps and deliverables in the project plan, and the data scientists and engineers constructing the application must function within this framework,” says the author.
For AI applications, planning for the work at the outset of the project and having the required resources to accomplish it during the device development lifecycle has become a crucial success factor.