Lowe’s is fighting to prevent AI agent overload

Modern Retail spoke with Chandhu Nair, svp of data, AI and innovation at Lowe’s, about how the company makes decisions on whether to build and invest in new AI features. Nair has been working on Lowe’s AI initiatives for several years and is particularly interested in the issue of AI sprawl — what he defines as when AI agents are narrowly focused, poorly coordinated, difficult to maintain and built in silos across an organization. He also spoke about the rules and guardrails the retailer puts in place to ensure quality responses and security. This interview has been edited for clarity and length.

How are you addressing what you call AI “sprawl” and building governance around AI?

“The evolution that you’re seeing in the industry is from these conversational chat assistants using generative AI to many more agents. You have a quick way to build an agent that can take multiple tool sets and perform a very specific task. An example of an agent that we have is an invoice reconciliation agent. That can go back and look at different documents on a procurement order and invoice, and if there are discrepancies, it can understand and correct it.

It helps automate a lot of the mundane processes that no one wants to do. Now, the challenge with technology like that is that it’s easy to get carried away and build a whole set of agents, and that’s the sprawl that can happen with it. What we had to do was create, essentially, one taxonomy of where we apply these agents. Within Lowe’s, we’ve created this AI transformation office, which has a governance process that looks at which areas have the right use cases to apply these agents.

When you deploy these agents, there is a “human-in-the-loop” framework to all these. We usually start with a human in the loop. For certain low-risk tasks, [humans are] always observing and only intervene when there is something that needs to be done. Certain tasks carry no risk. Everything that we do goes through our Lowe’s-built AI foundry, which is a Lowe’s-based AI platform. It has observability, it has explainability to it, so I can track how these agents are performing, etc.”

What do you mean by “sprawl”? What can go wrong when you have too many AI agents?

“The inherent technology — which has now obviously gotten a lot of attention with agentic AI — the idea is you can use the LLM to specifically go after a particular task. You give it to them in simple English, and it can use different tools. It can use an API, it can use an Excel sheet to solve that particular task you gave it. That’s the whole idea of agentic AI. The sprawl is because it’s so easy to do, because I can now prompt it and say, ‘Hey, use tools 1, 2, 3,’ and do it.

The challenge is to make it work in a very consistent way. If you and I were to do that same task in Excel, we would have that context of what needs to be done. Inherently, the models have to learn the context of how that needs to work. Humans also have certain control parameters that we have built into our memory and context in terms of what we would do or we would not do. The agents have to be trained to make sure it’s consistent and it works within those boundaries and frameworks.

It’s fairly easy to build an agent. You could have multiple engineers build out agents, and suddenly you would have a list of agents without a purpose and not working consistently every time. It may work 20% of the time or 80% of the time; the challenge is to make sure it works 100% of the time.”

What kind of rules do you have around building new AI tools?

“We certainly have guardrails at every layer of the technology stack, and then, we also have guardrails on the process itself.

I’ll start with the process side. We use four parameters. One is, ‘What is the true ROI for the business case?’ It has to be meaningful for the company. The second is the capital investment that is needed. Those are traditional; you apply that to practically any technology investments you do. The third dimension is the risk dimension. Does this carry any sort of brand risk? Is there a privacy risk? Is there a security risk? Is there a sprawl risk?

Probably the most important part, the fourth parameter, is that someone has to reimagine the business process with this technology. If that person, that change management is not there, the technology is not going to be useful. It’s not going to use itself, so it’s also important that the right change leader is attached.

From a technology perspective, we look at all different levels. We set guardrails at the model level to make sure that it is responding to the Lowe’s context. For example, we don’t want our Mylow companion in the stores to be answering questions on politics or things that are not relevant for the Lowe’s context. We wanted to be very specific to home improvement and what the store associate has to do.

And then, at the application level, these setups have several layers of security. There is a whole set of security layers that we’ve built so you cannot inject artificial prompts and find there’s a way to get to our systems.”

Are there any issues in AI implementation that you’ve seen from other retailers that you try to avoid?

“I don’t know I want to name any specific failures or other companies. But look, I think some of this is just the fundamentals of building good technology. The base of doing it is a little different with this technology, but the fundamentals of doing good technology have not really shifted.

If you think about when the internet came in, or when social media came in, guardrails were put in place even during that time. I think the basic principles that we are adopting are similar. We have to either bring in different methods of doing it, just because the technology is new, or add in more layers.

Initially, when the models came in, there was lots of news around models hallucinating and giving inconsistent responses. We take those learnings, from things that happen in the marketplace, and ensure that we have the right process to control the hallucinations. Those types of hallucinations are even more important when I talk about agents. You can’t really have an agent act if your model is inconsistent in the way it reasons.

We, thankfully, have great partners with all the key players in the ecosystem — the OpenAIs, the Googles of the world; we work with them and try to bring the best practices from them as this is evolving. We also have to hire the right talent who has expertise in the space to make sure this is built the right way.”

How does Lowe’s think about using AI at a high level, both in commerce and corporate usage?

“We had to understand what will change when this technology becomes relevant at scale. We broke it down to three areas, and that became our AI framework.

It’s how customers shop with us. That’s the whole commerce angle of what we do with then how we sell, which is our stores, our associates and how we engage our customers and sell through that. For our pro business, we have a lot of inside, outside sales agents. So we look at how to use this technology to superpower these associates. And then, how we work: Think about all the functions, from tech and marketing and merchandising, and any of these support functions to support the stores and our customers.

We then looked at, ‘How do we apply AI into each of these buckets: shop, sell, work?” The sell side was the area that we started on — our stores. We have about 450,000 associates and 1,700 stores. Each of our stores is pretty big. They are 100,000-square-foot stores. Because our associates are trade specialists, and the product knowledge is limited to a certain set of folks, we launched an assistant for every associate in the store that can take any product question the customer has.

The other challenge, unlike a traditional grocer or any other retailers, is that people come to our stores saying, ‘How do I fix a leaky faucet?’ or, ‘I’m looking to build a raised garden bed. What do I need to do that?’ And then, ‘What tools do I need to put that together, and can you show me how to put that together?’

With this assistant, regardless of which department you are in or which product category you service at any store, you can ask a question on the device with the Mylow companion and service that customer. It will tell you not just how to build the raised garden bed, but it will also show all the products you need to buy, where they are in the store, how much is available, and whether you can get shipping on them.

Everything is available in one place. It just makes the life of the associate and the customer, that whole experience, easier. It’s been a little over a year since we rolled that out to our first store. It’s the same underlying technology that is used for the consumer commerce side, in terms of Mylow, but between the two, we get about a million questions a week. Think about every engagement that AI is helping convert.

Since we rolled that out, we have seen about 300 basis points of improvement in our net promoter score, which correlates to sales numbers. I’m not at liberty to share the exact numbers here, but you get the point of how this has evolved. Obviously, it’s not an overnight switch. We had to iterate. We had to work with the store operations. We walked through every store that we could with the tool to continue to iterate and get feedback.”