Q&A: Agentic AI in retail and supply chain planning

Jun 16, 2025 • 8 min

We sat down with RELEX experts Max Mononen, Product Director, and Rich Kurhajetz, Senior Strategist, to explore what agentic AI is, how it works in practice, and how businesses can successfully implement it.

What is agentic AI?

Max: AI agents are small autonomous entities that act, adapt, and collaborate with superhuman efficiency. Within a system, they can operate individually or as a network of decision-makers. You give them an objective, and they continually diagnose, optimize, and perform tasks to solve complex problems and achieve that objective. They can be triggered by data systems and, in the future, by other agents.

AI agents are small autonomous entities that act, adapt, and collaborate with superhuman efficiency.

I do want to call out an important nuance. The agent’s job is to achieve a goal. When I say an agent “acts,” I mean that it decides what task will support that goal and then calls on a tool to perform it. There are various tools at its disposal – algorithms, codes, and mathematical calculations – and those tools perform the action. The agent then analyzes the outcome, determines whether or not it has succeeded, and calls on tools in a constant loop until the goal has been achieved.

This is why agents aren’t much use without a robust platform of quality data and proven machine learning and optimization capabilities. Agents are only as effective as the tools they can use.

How are AI agents different from chatbots?

Max: The biggest difference between agents and gen AI chatbots is that agents take action and chatbots provide information.

For instance, at RELEX, we have Rebot, our friendly gen AI assistant that behaves like a copilot. In its current iteration, it accesses the RELEX repository of industry- and solution-specific knowledge, but it does not access data in a customer’s RELEX environment. So, you can ask Rebot to explain an industry term or solution capability, or outline different ways to measure demand forecast accuracy, and it will give you those answers. But it doesn’t look at your data. It acts as a 24/7 consultant, helping companies accelerate onboarding, ease daily operations, and improve users’ understanding of the solution so they feel more confident in their decision-making.

The biggest difference between agents and gen AI chatbots is that agents take action and chatbots provide information.

Agents are the natural evolution of conversational chatbots. But unlike gen AI assistants, they live in your environment and therefore can take action based on your data – transparently, safely, and in collaboration with humans.

Rich: From a user perspective, interactions with agents and with gen AI assistants can look very similar. There’s usually an input box where you type your questions and prompts. But the biggest difference is what’s under the hood. With a generative AI chatbot, LLM, or GPT, you ask a question, and it answers — but it doesn’t take action. Agentic AI can perceive, decide, and act based on the parameters and rules you provide.

How does agentic AI function in practice, and what are some key use cases?

Max: Put simply, anything that a user can do within the RELEX system, an agent could do. And because the agent doesn’t have the limitations humans have in terms of time, it can handle far more tasks. But an agent usually has just one action it’s instructed to complete. This restriction is part of the guardrails that make sure it doesn’t change anything it isn’t supposed to.

As a concrete example, let’s look at the business rules engine (BRE) behind the RELEX core logic. The BRE allows users to set up rules and workflows that are triggered automatically, depending on input, data, and thresholds. Within an agentic system, you could use an agent to run such a workflow.

An agent usually has just one action it’s instructed to complete. This restriction is part of its guardrails.

For example, you could use a prompt like this: “Help me create a dashboard showing my lost sales value for these different types of items in these different locations. Summarize the actions that I should take to prevent those lost sales.” Instead of building those dashboards yourself in the user interface, you can pre-configure them and prompt them when you need them.

Rich: Automation in general is a major use case. Think about retail and supply chain planning systems. There is so much math, so much data, so much complex decision-making. With agents, there’s an incredible amount of latitude for automation. We’re not talking about automating everything in your end-to-end platform. Humans should be in the loop. But we are talking about tasking and training agents in very specific ways to support planning decisions.

With agents, there’s an incredible amount of latitude for automation.

For instance, right now, the number of planning exceptions that need a review has to remain manageable for a human. But that can leave you with the question: “Did we miss something really important?” The power of AI lies in the ability to parse through all that data, performing high-level analysis and prioritization, taking action on smaller issues, and surfacing bigger issues that require human review.

This breadth and depth of data analysis opens up a whole new field of advanced analytics that previously might not have been something a user was trained for. Now, with specific agents, an entry-level or mid-level user could unlock some very advanced analytics through something as simple as prompt engineering.

How will agentic AI impact the future of work and user roles?

Rich: I think what people are wondering is: “Am I going to lose my job, or am I going to be doing different work? Is this role going to be eliminated, or are we going to start creating more data scientist roles?”

Fundamentally, the work is still there. You’re doing higher-level work and cutting out the mundane tasks, but you still need those analyst roles to review exception handling and the remediation of simpler tasks.

However, agents will change how users interact with their planning systems and add value. Who does the best with AI? Well, whoever asks the best questions. It takes an educated user.

Max: It’s human-computer interaction for a reason. Even if agents take care of a lot of tasks, you are the one controlling them. As the super user, you set the rules, goals, and targets that follow your company’s strategy and priorities.

Who does the best with AI? Well, whoever asks the best questions.

The biggest difference is that with agents, you now have thousands of junior analysts available. You can guide them to take specific actions, changing parameters based on your objectives. As agents handle repetitive tasks, you can focus on the tasks that add the most value and help take your business forward.

Also, team compositions will change because even now, teams that use AI tools and copilots are outperforming teams that don’t. This means that businesses will be able to scale and achieve more with the same number of people they have now.

What guardrails should be in place when implementing agentic AI?

Rich: Just slapping AI into your current system without a good plan puts you in danger of doing a lot more bad work at scale. You can do quite a bit of damage very quickly because you can now make a massive amount of changes.

So, what guardrails should you have in place? Well, think about people, processes, and technology. You wouldn’t give an entry-level analyst access to making major changes to your monthly or annual plan. Those kinds of role-specific permissions should apply to agents as well.

Max: Right, think about the permissions you’re giving your agents. You can break those permissions down into three basic categories.

First, decide what tools to give those agents. As I mentioned, agents can only use the tools at their disposal. Limit them to the tools they’ll need to accomplish the goal you’ve set for them.

Think about people, processes, and technology. Those kinds of role-specific permissions should apply to agents as well.

Second, determine rules for the agents themselves. How autonomously can they act? Let’s take that example from earlier where we talked about prompting an agent to help with lost sales. To start, we ask for an analysis to pinpoint which locations and products are experiencing the most lost sales. Then you might start asking for suggestions. What should I do about this lost sales problem in this region’s stores? The agent makes suggestions and asks if you want it to take action. It asks your permission until its suggestions are proven trustworthy, or until you’ve set up rules or other agents that can automatically take the appropriate measures.

Third, set clear rules of engagement. This comes down to governance. Define what area each agent operates within, what workflows it handles, and the types of behavior it should follow when you prompt it. And have hard guardrails in the background. You don’t want it to be like that chess agent that started cheating when playing against the biggest chess AI system because its only instruction was to win – it didn’t say anything about playing by the rules.

Where should companies start their journey toward implementing agentic AI?

Max Mononen: The first advice I’d give is just to get started. If you’re using RELEX, you already have access to Rebot and many different AI capabilities. When it comes to agentic AI, start dabbling and playing with different systems. That’s the easiest way to learn. But be careful with your data. Contact your legal department to ensure you get proper tools and permissions that won’t jeopardize information security or data privacy. Don’t upload company data to external systems, but try out available tools.

Just get started. That’s the easiest way to learn. But be careful with your data.

Second, determine what business outcomes you are looking to achieve and then identify the ones that are a fit for AI. Don’t try to do everything with agentic AI. You’ll just end up creating agents with no value. Even though you can do more now, you still need to prioritize. Look at the time usage of your team and consider where you are losing money, where your biggest gains would be, and what you don’t have time to do right now.

With agents, you can start looking at high-impact topics you didn’t have time for before.

Many people immediately think of reducing reporting time, and yes, reporting tasks can be a nuisance. But at the same time, that’s a very low value add task. With agents, you can start looking at high-impact topics you didn’t have time for before, reviewing the long tail of problems and determining where you can apply these agents to help you visualize data and take action.

Third, if you haven’t already, start implementing a platform that will give agents what they need to complete tasks – the machine learning, optimization, and data management capabilities companies are already using to accelerate and improve their supply chain and retail planning.

Wondering what tools AI agents could use? Check out the guide to AI at RELEX