A model for successful implementation of Artificial Intelligence Models
AI Growth Journey:
Let me start this story with some growth statistics. Artificial intelligence (AI) is changing the competitive environment in industries around the world. The potential comes from automating decision-making process through replicating human-like reasoning. The number of AI startups since 2000 has magnified to 14 times as per Forbes. As per statistica, the funding of AI startups has grown significantly.
AI will automate 16% of American jobs as per Forrester. At the same time, the proportion of US patent applications that contained AI rose from 9% to about 16% between 2002 and 2018.
As per Gartner “In 2019, nearly 37% of enterprises implemented AI — depicting an increase of around 270% in the past four years
Challenges in AI Implementation:
One of the greatest challenges is the absence of a clear implementation strategy. Success requires a strategic approach and implementation of AI. These include identifying areas for improvement, establishing goals with clearly defined benefits, and ensuring a continuous feedback loop of the improvement process.
A successful model for AI Implementation:
Some of the factors that will have a direct impact on successful implementation are listed here. We choose the causal factors to develop a potential model for AI implementation. These factors had been identified based on survey of successful and not so successful implementations.
Proposed model for Successful AI implementation (Source: Dwivedi et. al. 2022)
1. Data quality
Data Quality plays a pivotal role in determining the success of an AI implementation. As per Gartner report in 2021, every year, poor data quality costs organizations an average $12.9 million. Poor data quality leads to poor decision making both in the short run as well as the long run. Following are some of the aspects to improve data quality:
· Identifying and imputing missing values in the records.
· Identifying and removing duplicate records.
· Identifying and treating anomalies in data fields.
· Identifying and treating inconsistent records.
2. Industry specific data to build AI applications
In order to develop efficient AI-models, a sufficient quantity of good quality data is required. Often companies lack availability of sufficient data to start with AI at first place. A key challenge here is to have enough actionable data that can used to enable key AI insights.
3. Lack of highly trained and skilled professionals
Advances in AI are slowed across industry by a global shortage of workers with the efficient skillset in AI viz. Deep learning, NLP and robotic process automation. The talent available is limited and as per a survey conducted by O’Reilly in 2021 more than 50% of the organizations realize a skill shortage in ML modeling and data science professionals. Though many companies are emphasizing on closing the skill gap by organizing an in-house or external training program, but with new advancements in AI it is difficult to meet the ever-increasing knowledge base.
4. Computing resources to test more demanding methods
Training machine learning models can be a computationally intensive task. Choosing the right infrastructure to train and operate the machine learning algorithm will have a significant impact on the performance and execution of the model. There are two primary processors used as part of the most AI process: central processing units (CPUs) and graphics processing units (GPUs). CPUs are used to train most traditional machine learning models. GPUs are suitable to train deep learning models and visual image-based tasks. These processors handle multi-thread calculations in parallel and are much more expensive as compared to CPUs. There are three main operation model used by organization viz. On-premise, cloud service model and a hybrid model combining both on-premise and cloud model.
5. Trust within organization
For many organizations analytics and AI are key tools for taking business decisions. Conventionally business decisions are taken based on expert knowledge, experience and business surveys and still many organizations lack a high level of trust in the way their organization uses data and analytics. Organizations need to be communicated and cost benefit analysis of using AI based decision making needs to be cultivated
6. Data Algorithm and decision bias
With the growing use of AI in decision making a challenge arises to make an unbiased decision-making algorithm. Some biases can be introduced while using AI algorithms. Some of the biases can arise because of using biased data. Before using the base data, it needs to be assured that data is free from all kinds of biases and is relevant for the use case it is used for. Other times human biases also impact the outcome of a model where human overlays are used alongside model which tend to add further bias to the outcome.
7. Absence of collaboration model
Developing an AI based decision engine would involve outputs from business and data science team. This can include analysts, data scientists, data engineers along with some business users with a collaborative effort being deployed by all stakeholders. With an increased use of remote working structure organizations need to have an effective collaboration environment supporting data ops and model’s ops capabilities.
8. Data privacy and GDPR
General Data Protection Regulation (GDPR) and The California Consumer Privacy Act (CCPA) both of which came into effect in 2018 has stressed on the issue of privacy and opened up a tight compliance guideline. AI needs to ensure that there is a balancing act on data protection, privacy and ethics.
9. Integration with operational engines
Effective AI doesn’t end at using AI technique, it is a process which also involves operationalizing with the existing/recommended platform. Optimization of the integration platform involves ingesting AI into it to improve execution and make AI driven decisions performance by standardizing the decision process from data integration to solution output. Another notable advantage is AI is to train the training data set to automate the statistical modelling process without any manual intervention
10. The cost factors
Almost all business organizations in the world are embracing AI. The benefits delivered by implementing effective AI in the business process are colossal. Effective AI also involves cost benefit analysis. There are numerous factors leading to the pricing of AI:
· The current level of AI maturity
· The level of AI maturity the company needs to enter
· Architecture of AI to be used
· The performance of AI algorithms
· The complexity of AI algorithms
· The amount and availability of data to be consumed by AI engine
While it should be admitted that this is an ongoing journey and many more such drivers would be added to the feature store as we move on.
Hope you like the story !!!