Accenture’s report reveals that 3 out of 4 C-suite executives believe that if they don’t scale artificial intelligence in the next five years, they risk going out of business entirely. While you are making efforts to deploy AI, you should be aware of the challenges of this journey.
There are essentially three key challenges associated with AI adoption:
1. Data quality and quantity:
Technologies such as machine learning and artificial intelligence (AI) have the potential to help businesses make better use of these massive volumes of data but these techniques depend on the quality & quantity of data provided. Sourcing consistent and accurate data has been a difficult task since there are no commonly accepted and widely adopted standards of data definitions and governance in enterprises.
“Enterprises are pursuing a range of AI initiatives and modernizing data infrastructure tops the list. But current data practices are an issue, as several companies haven’t attained a high level of sophistication with crucial data-related aspects. In fact, many organizations are stopped in their tracks when pursuing AI programs because the data is not good enough, hence predictions and insights would also be unreliable. Many companies tend to postpone their AI journey in favour of a data journey first before starting the AI leg,” said Saurabh Kumar, Partner, Deloitte India.
A Deloitte study of AI adopters finds businesses face challenges in critical aspects of data management: preparing and cleaning data, integrating data from diverse sources, training AI models, and ensuring data governance.
Solution
Kumar suggests businesses to have a strong data management, data quality and governance framework to ensure that all the data generated in the organization is captured, processed and stored effectively. He further insists on involving all users including the AI practitioners early in the process of architecture design so that the requirements are considered while designing and no patch work required later as an afterthought.
“A right blend of cloud and traditional data warehouse set up will help organizations achieve optimal performance. Focus on Future integrating and scaling- Ensure that your vendors and partners design the data layer with a forward-looking approach which will ensure that the integration of data from a variety of new sources is not a challenge later,” said Kumar.
2. Right People and Talent
The second top challenge while adoption AI could be not having the right team or people to work with. Right talent is the key to success for any initiatives and the same is true with AI as well. Of course AI is a far more complex skill to build and certainly there is a demand and supply gap in the marketplace.
AI comprises a range of new technologies which covers advanced analytics with the ability to predict outcomes, Robotic Process Automation (RPA), Natural Language Processing (NLP), creation of customer friendly BOTS and deep learning. The sheer vastness of the technology makes it difficult to find the right people for both the creation and implementation of an end-to-end AI journey across an organization. Also, AI takes time to evolve and requires constant creative and material investment till it matures and starts providing a level of acceptable accuracy. So not only do we need expert technological resources, but creative brains as well who can innovate the use-cases for the technology.
Solution
Mehmood Mansoori, President – Shared Services & Online Business, HDFC ERGO General Insurance suggests starting AI journey in the organisation with the help of internal champions and to train them in the application of available AI technologies.
“Be willing to challenge the status quo and more importantly build a culture where business teams can think about the use of AI in day-to-day operations. Once you have built this culture, you have more champions beyond the innovation group to motivate the rest of the organisation to walk the same path,” he said.
“Also, it is important to create a conducive environment, which starts right from an office setup, flexi-hours for youngsters who may not be early risers but like to burn the midnight oil and have the zest to prove themselves. This is exactly what we did; our office no longer looks like a traditional insurance office, but is rather the envy for many international digital R&D setups and this is just the beginning. At HDFC ERGO, we are committed to challenge the traditional practices and attract talent from any industry, as long as they are analytical, creative and excited about AI like we are,” added Mansoori.
Companies will have to invest in the right talent, train internal resources with the right aptitude and the know-how in related technologies, add people to the team who are creative and think unconventionally while solving business problems, engage with the business team across the organization and generate excitement about these technologies.
Also read: Here’s how you can set up a successful AI team
3. Eliminating Bias
Biases in any of these use cases impair the possibility of making the right decisions. Businesses might get a false assurance that they are seeing gains using biased models, but models with biases are not really solving the learning problem they are supposed to solve.
“As a common example, consider a model that’s being built to evaluate a candidate’s ability to do well for a certain role. Because of poor design, the model might end up correlating a factor like location to predict if the candidate is a fit or not. This might work out in biased evaluation sets, since location can help to identify historically under-represented sets of people in that role. This model, while apparently doing well on testing sets, is not really doing what it’s supposed to do, i.e. evaluate a candidate’s ability for a role.” said Abhinav Tushar, Head of Artificial Intelligence, Vernacular.ai
Solution
“The first thing to start with is the modelling goal itself. Are the goals of the model carrying inherent bias? Can the objective functions be made fair, by design? Many times the problem and evaluation criteria itself has a bias that needs to be resolved before anything else,” Tushar explained.
Datasets used for training, especially evaluating the model should be balanced, so that they don’t reflect real world biases. This is a common source of bias as datasets collected from the wild tend to be discriminative against arbitrary groups, the knowledge of which shouldn’t be important for the model in making predictions.
“Unknown biases can still sneak into the system, thus strong Quality Assurance (QA) processes, specifically keeping bias in mind, are important. Creating many slices of test sets covering potentially problematic situations is an important step here. These QA processes should be extended post integration too, since data drift and feedback loops after deployment can still bring biases in the system,” Tushar added.
Also Read: How enterprises dealt with AI biases in 2020
