“How do you best operationalize AI?“

By Harley Davis

One of the biggest challenges many businesses face is how to effectively use the good work done by their data scientists, AI experts and business operations experts to improve business outcomes.

For example, data scientists typically work with whatever data they can find to understand client behavior, or how the business runs, to implement predictive analytics and machine learning (ML) models. However, these models are used sporadically, offline, to generate propensity or risk scores that are only updated when the models are retrained. They don’t respond to situations as they occur.

The potential for integrating AI models into operations has never been greater. Organizations across industries are moving quickly to bring their analytics work out of the shadows and into daily operations. To do this successfully requires good tools used in the right ways with a focus on data and understanding the nature of your models.

A new generation of tools for operationalizing AI

One of these good tools is a new generation of digital decisioning platforms that can insert the models right into the transactional decision making – whether it’s processing a claim, at the point of purchase, or during client interactions on a website.

Originally based on human-readable business rules, decision management systems have evolved to manage more complex business requirements by combining the specificity and clarity of business rules with the ability to apply AI-based ML models at the point of decision making. All of this is wrapped up in a decision modeling framework that lets businesspeople understand how decisions are made and how the elements come together to more efficiently run the core transactional and operational decision making of a business.

The potential benefits are significant, but how do you get there? There are three keys to help ensure success.

Key 1: Reimagine your process

Start from the decision you’re trying to make and the data you have available, and work to bridge the two. What does this mean in practice?

Create a decision model. With modern digital decisioning platforms, you can model how you make decisions. For instance, a large health insurance company may start with the following decision: “Should I approve this claim, reject it, or send it to a human adjuster?”

Break down the decision into its component parts. If the claim does not meet the rules of the policy, reject it. If the claim is very likely to be fraudulent, reject it. If the claim might be fraudulent but we aren’t sure, send it to a human. Otherwise, accept it.

Continue down this path for each of the component parts: What are rules for the policy of the claim? How do those rules apply to this claim?

Identify the data needed for the decision. To answer the questions above, you need data about the claim itself, thresholds and other information in the policy.

If the claim is likely fraudulent, the process will be a bit different. Here you need to apply an AI or predictive analytics model. This model takes various parameters around the claim and the client as input and produces a numerical confidence score indicating the degree of reliability of the model’s judgment.

Key 2: Assemble a decisions team

Once you’ve identified the process, bring together the key stakeholders:

Businesspeople who understand the decision. The businessperson can help model the sub-decisions to determine the business rules and thresholds for the AI model and how they’re plugged into the decision model.

IT people who build the decision model in a digital decisioning platform. They choose and insert the model into the transaction flow, identifying and sourcing the data elements needed for the decision. The IT person also develops a governance process and test plan to ensure that the decision service will work correctly.

Data scientists who construct the AI models used in the decision. They need to provide the context for how the model can be used and, perhaps, modify it to take into account the execution context and available data at decision time. They’re also responsible for other issues that impact the decision model, such as getting feedback from the decision for future iterations and managing issues such as KPI drift and bias detection.

Key 3: Map the data to the decision model

Finally, map all of the parts of the decision model to the available data. Some data may have to be mapped from one form to another – to match the format expected by the analytics model. And you may have to adjust the decisions or change the transaction stream to get the needed data.

This whole process is outlined in the diagram below:

decision process model

By moving to AI-based models that embody deep understanding of the business and leverage big data, you can make more accurate and customer-centric decisions. If you can store and use the result of these decisions, you can also provide valuable pre-curated data to help retrain and refine the models dynamically, making the process scalable.

Watch this demo video (04:59) to see how you can make business decisions with greater flexibility and accuracy by infusing business rules with machine learning.

At IBM Think 2020 in San Francisco, 4 – 7 May, attend sessions that can enable you to apply AI to help improve decisions at every level of your organization.

Sign up for the IBM Automation Insider

What automation opportunities are out there? How do you get results?
Every other month our experts share five pieces of strategic content to help you drive growth through automation.