AI in retail has to be semi-automated. Here’s why

Retailers need more decision automation, faster coordination of supply chains, and faster interactions with consumers, which means they will increasingly rely on AI. Automated decisioning systems will soon be making fine-grained micro-decisions on the retailer’s behalf, impacting customers, employees, partners, and suppliers. But these systems can’t run autonomously — they need human managers.

So what exactly should this human management look like?

Every system for making micro-decisions needs to be monitored. Monitoring ensures the decision-making is “good enough” while also creating the data needed to spot problems and systematically improve the decision-making over time.

Consider the following retail example: A fashion retailer that had historically applied blunt rules to determine markdowns decided to implement a new AI-powered solution. The system performed well for the first few weeks, making more frequent and more surgical decisions than human managers were able to contemplate. But at the start of the swimwear season, the system identified a slow initial sell-through that triggered all swimwear to be marked down to sell-out. As a result, the retailer lost millions of dollars of margin and was left with no swimwear.  

Why did this happen? The aggressive markdowns were triggered because the first three weeks of sales were lower than expected. A human merchandiser would not have panicked and would have realized that this was due to a couple of particularly cold weeks. But the unmanaged and unmonitored AI system simply executed on its logic.  

The example above illustrates why the best approach to deploying AI is typically semi-automation: automation that involves some level of human oversight. When optimized for each decision, semi-automation can help retailers save time, empower employees, and greatly improve profitability, while avoiding costly pitfalls.

4 approaches to semi-automation

The four models for semi-automation range from heavy to very light human involvement. 

First, human in the loop (HITL) is the most basic framework for semi-automation, where decisions are rarely made without human involvement. Such a system provides recommendations based on automated calculations, but a human ultimately makes the decision. For example, pricing software calculates the ideal price of a dress to maximize profitability, but the pricing manager must sign off on each decision.

The next model is human in the loop for exceptions (HITLFE), where humans are removed from standard decision-making, but the system engages a manager when human judgment is required. For instance, if the automated system has two vendor options for stock replenishment, the buyer is required to step in and make the final call.

Then there is human on the loop (HOTL), which means the machine is assisted by a human. The machine makes the micro-decisions, but the human reviews the decision outcomes and can adjust rules and parameters for future decisions. In a more advanced setup, the machine also recommends parameters or rule changes that are then approved by a human.

Finally, there is human out of the loop (HOOTL), which is where a human simply monitors the machine. The machine makes every decision, and the human intervenes only by setting new constraints and objectives. 

Common pitfalls of full automation

Selecting the right model to use is a design problem. As we have seen, automation is not “all or nothing,” and decisions are not created equal. The right model should be determined based on the decision’s complexity, volume, velocity, and “blast radius,” which measures the potential downside. For example, if the decision is simply to recommend a blue dress instead of a red one because blue is out of stock, it’s a low-risk decision that can be fully automated with limited oversight. However, if the worst outcome results in misordering thousands of dresses or in expensive markdowns like in the swimwear example, then human oversight and accountability is more critical. It’s also important to recognize that automated systems can and will evolve over time, enabled by new technology, the desire to make ever more fine-grained decisions, and management’s confidence in automating business operations.

Making a successful shift to semi-automation

The key to the successful deployment of any AI system is to start with a quantified business problem. With this, retailers must foster a data-driven culture where the whole team is engaged in determining how best to improve specific business decisions. This also necessitates a change in how retailers do their jobs. For example, merchandising managers in the past might have had to set prices for several dozen dresses a day based on stock, sales data, and competitor activity. But now, with personalized promotions and recommendations, the same manager might be responsible for millions of decisions a day. This requires a fundamental shift from making decisions to making decisions about decisions — i.e., managing rules and parameters rather than making specific pricing decisions. 

Semi-automation of business-critical decisions must be approached carefully, with regard to the potential heightening of the blast radius of risk. Once the decision to automate has been made, retailers must shift their attention to decision algorithms — the logic and rules that enable retailers to execute on the micro-decisions. Miscommunication between the data science team and the rest of the organization can lead to errors and missed opportunities, potentially creating a reluctance to change that can be quite difficult to reverse.

Automation is the future, but not without human intervention

Whichever model you adopt, it’s critical to put AI on the organization chart to ensure that human managers feel responsible for its output. To succeed, retailers must understand the different ways they can interact with AI and pick the right management option for each AI system. Selecting the best level of semi-automation will ensure that the retail businesses realize the full potential of AI.

Michael Ross is Senior Vice President of retail data science at EDITED. He is a non-executive director at Sainsbury’s Bank and N Brown Group plc. He also cofounded several companies, including DynamicAction, ecommera, and figleaves.com. Prior to that, he was a consultant at McKinsey and Company.