Motivated by multiple drivers, enterprises across nearly all industries are increasingly embracing artificial intelligence (AI) and machine learning (ML) to enhance efficiency, profitability, and customer experience while improving evidence-based decision making. Ever-increasing volumes of available data, both structured and unstructured, combined with ongoing innovations in the software and infrastructure space capable of handling large data volumes efficiently, is facilitating this adoption.
Implementation of AI technology and ML solutions can require significant investment. Based on our experience spanning multiple industries, we have identified key considerations which can help any implementation of AI/ML be much more efficient, leading to a successful adoption (as compared to AI technology “sitting on the shelf”) and enhanced return on investment.
Business challenge identification: The first step toward a successful implementation of any AI or ML solution is to identify business challenges the organization is trying to tackle via AI/ML and gain buy-in from all key stakeholders. Being specific about the desired outcome and prioritizing use cases driven by business imperatives and quantifiable success criteria of an AI/ML implementation is helpful in creating the roadmap of how to get there.
Data availability: Enough historical data, relevant for the business challenge being tackled, must be available to build the AI/ML model. Organizations can run into situations where such data may not yet be available. In that case, the organization should develop and execute a plan to start collecting relevant data and focus on other business challenges that can be supported by available data science. They can also explore the possibility of leveraging third-party data.
Data preparation and feature engineering: This is one of the most important steps in the development of an effective AI model. In this step — in addition to the usual data cleansing, data integration, use of AI tools such as Natural Language Processing to incorporate structured data, judicious and creative feature engineering, creating the training and test data, etc. — it is also important to consult with the business stakeholders and the legal team to ensure that the data/features being used in the model comply with any relevant regulatory frameworks and laws (e.g., Fair Lending). It is also important to incorporate “existing wisdom” in this step. For example, if the objective is to build a fraud detection model, prevalent fraud patterns already known to the organization’s investigation unit should be incorporated. In addition to enhancing the effectiveness of the model, this builds confidence for the end-users of the solution, thus facilitating adoption of the model.
Selection of an appropriate modeling approach: For any given business challenge, it is common to find that multiple AI and ML algorithms are applicable. Often, the simpler algorithm or model with fewer parameters may be a better choice (assuming the performance of different models is similar). A particularly important step in this process is to consider model explainability — is the selected model able to provide human-understandable, plain-English explanations and reasoning behind its decisions? In certain regulated industries, reasons behind decisions made by an analyst or algorithm are a requirement. Many AI/ML algorithms are, by nature, “black-box” in that the contributing factors for the model outcome are not clear. Model explainability packages, such as LIME or SHAP, can provide human-understandable explanations in such situations.
Strategy for operationalization: Having clarity around how the predictions and insights from AI/ML fit into daily operations is clearly needed for a successful implementation. How does the organization plan to use the model scores/insights? Where does the AI/ML model “sit” within the operational workflow? How will the model insights/score be consumed in the process? Is it going to completely replace some of the current manual processes, or will it be used to assist the analysts in their decision-making? Will the solution be implemented in the cloud or on-premise? How will the data flow into and out of the AI/ML solution when implemented? Is there a funded plan for procuring the necessary hardware and software? Having a well-defined roadmap that addresses such questions will go a long way in making sure that the solution gets operationalized and does not sit on the shelf.
Phased implementation approach: The human factor is one of the hurdles faced in any AI/ML implementation effort. People are often uncomfortable with sudden and dramatic changes to their existing processes. A phased implementation approach can help mitigate such concerns. We often suggest a pilot phase, in which the AI/ML solution runs in parallel with the existing process — so that relevant teams have an opportunity to compare the outcomes of the two and become comfortable with the new process.
Training, skilling, and enablement: Of course, it is important to build teams with expertise in various areas of the AI/ML space. Ensure that the relevant skills and resources to support the operation of the AI/ML solution are available. Any skills gaps should be bridged by either training the existing resources or bringing in new resources with appropriate skills.
Thinking through each of these recommendations and having a clear strategy from the beginning to address them will greatly enhance the chances of success and return on investment for any AI/ML implementation.
Connect with the authors:
Managing Director – Emerging Technologies Global Lead, Protiviti
Senior Director – Machine Learning and AI Lead, Protiviti
Director – Machine Learning and AI, Protiviti
Artificial Intelligence, Machine Learning