Some companies are using AI for end-to-end decision-making, but not all decisions can be made without human intervention. Here are some real-world cases.
As artificial intelligence (AI) ascends in the marketplace, the burning question remains as to how far it can be trusted when it comes to the “last mile,” the final decision that follows the analytics and recommendations that AI yields.
In medicine, AI and analytics crunch through reams of data and scientific research to come up with a series of recommendations for a difficult diagnosis, but it is the expert medical practitioner who makes the final decision.
In a loan-approval process, automated decision-making software reviews an application and third-party data to determine a lending decision, but the loan underwriter or supervisor makes the final decision.
“Not all decisions in organizations can be fully automated, and some of these will require human intervention,” said Arash Aghlara, CEO of Flexrule, which produces decision automation software. “Decision automation should allow scenarios in which fully automated decisions are not possible because of ambiguities, uncertainty, and so on regarding the decisions. Instead, these require domain experts’ inputs and intervention.
Where AI can make decisions for an organization
Yet, if a company wants to use an end-to-end automated AI decision-making process, there are operational areas where AI can work. These are areas where there is an ironclad rule set that organizations are comfortable with and that have virtually no chance for exceptions.
- In your internal budgeting and budget administration processes, you can automate the levels for line-item approval based on dollar amount. Software can prompt you when amounts exceed $50,000 or $100,000 and must be approved by the CFO or CEO.
- In sales promotion and fulfillment, based on dollar amount and a customer’s past patronage, automatic discounts on merchandise can be given.
- A piece of factory equipment can be custom-configured based on the performance and environmental requirements in a given plant without the need for a salesperson or sales support engineer to intervene.
All of these examples have one common thread: They work on highly definable logic that the business inserts and that has little or no chance of varying from one case to the next. In these ironclad cases, it is possible to implement end-to-end automation without human intervention.
Taking automation a step further
There are, of course, companies that want to push the envelope and that have made a strategic commitment to go forward with end-to-end automation despite the fact that the automation may sometimes be wrong.
An example is Bestow, a life insurance company that made the decision to go to end-to-end AI decision-making on life insurance, with no human intervention.
“In leaving the decision making to the AI and its algorithms, we believe that we have solved one of the major challenges when human underwriters make these decisions,” said Ben Hsieh, director of product development. “In human decision-making, there can be inconsistencies and bias.”
Hsieh recognizes that there are times when algorithms and data models must be refined and when human intervention would be nice, but the speed to market of the company’s automated underwriting and the satisfaction levels of customers are making that risk worth it.
So what does all of this tell companies in general? That you can make a strategic decision to go “all AI” in your decision making, as long as you understand the risks and that taking the risk is worth it.
For most companies, though, a hybrid man-machine approach for “last mile” decisions works best.
“We use machine learning to automatically decide who should make the pricing decision—the salesperson or the model,” said Yael Karlinsky-Shichor, an assistant professor of marketing at Northeastern University. “What we find is that a hybrid structure that lets the model price most of the quotes that come into the company but lets the expert salesperson take those cases that are more unique or out of the ordinary actually performs even better.”
This feature originally appeared in TechRepublic by Mary Shacklett.