AI Infrastructure

How to Add AI Supply Chain Planning to Legacy Systems | Mike Landry, ketteQ | TFiR

0

Supply chains change faster than the plans that run them, and a plan can go stale almost the moment it is published. Most large enterprises still run planning systems built 20 or 30 years ago, and those architectures cannot use modern AI or cloud computing. Bolting an LLM on as a sidecar does not fix it, because the AI only sees a copy of the data.

In this interview on TFiR, Mike Landry, CEO at ketteQ, walks through how to add agents on top of existing planning systems, govern what they do, and decide when humans step in.

Guest: Mike Landry, CEO at ketteQ
Show: The Agentic Enterprise
LinkedIn: Mike Landry
Watch the discussion on YouTube

Here is what every supply chain planning leader needs to know.

Technical Deep Dive

Q: How do you add AI to legacy supply chain planning systems without replacing them?

Mike Landry, CEO of ketteQ, says older planning systems such as Kinaxis were built on architectures that are 20 or 30 years old and cannot use modern AI or cloud computing. ketteQ connects an intelligence layer called PolymatiQ and an orchestration layer called Quintus on top of them, where the data is already organized for planning and the ERP integration is built. Landry says connection takes minutes and the speed to value is quick because the existing systems and data stay in place.

“The speed to value is very quick by sitting on top of the existing infrastructure the data and the systems that are already there and implemented years ago, but just not meant for today’s fast-paced world.”Mike Landry, CEO, ketteQ

Q: Why do supply chain plans go stale so quickly?

Landry points to disruptions on the supply side while customer expectations keep rising, including demand from marketplaces like data centers. Demand and supply imbalances now show up faster, so businesses have to catch changes and uncertainties quickly to keep the promises they make to customers. He calls this an ongoing challenge for businesses and for supply chains in particular.

“Now the world is just operating faster and faster.”Mike Landry, CEO, ketteQ

Q: What does an AI agent do differently in production when a plan goes stale?

Landry says Quintus is free-range AI rather than scripted like a sidecar, so it handles novel questions with novel answers and learns from each interaction. ketteQ deployed it to 93% of its customers three or four months before the interview, and prompt volume is going up. He adds that Quintus is being connected to Copilot and Claude, so Copilot or SAP’s Joule can draw on supply chain planning expertise.

“We’re free-range AI, we’re not scripted like the sidecar.”Mike Landry, CEO, ketteQ

Q: What role do AI agents and LLMs play in supply chain planning decisions?

In PolymatiQ, thousands of agents run heuristic experiments in parallel instead of one heuristic plan at a time, to find the best supply plan for make, move, buy, inventory or demand. Landry compares this to having 1,000 planners no one could hire. On the user side, the LLM sits inside Quintus with a harness, so memory, skills, governance and security live in the customer’s data rather than travelling with the LLM.

“We can literally swap the LLMs out in minutes and we can use a collection of them for what they’re good at doing.”Mike Landry, CEO, ketteQ

Q: How should AI agents integrate with ERP and planning data?

Landry says data integration is different in the agentic world. With MCP, or by going to the ERP’s API, ketteQ trained Quintus on the data model of systems like Kinaxis, which uses a hierarchical structure rather than a relational one, so it can read and write where the data sits within a minute. A billion-dollar pet food company called PPF uses the solvers to write answers back into Kinaxis as if they were generated there, deployed in weeks without changing the platform.

“With that training and an MCP access, within a minute, we access that data structure right where it sits to be able to read and write.”Mike Landry, CEO, ketteQ

Q: How do you keep agent-written Python code safe in production?

Landry says some of the Python code Quintus writes might go into production, and skills can be turned into production missions. He notes that companies doing vibe coding only got so far because of roadmaps, versioning, support and DevOps. ketteQ opened its development environment so consultants and customers can inspect and extend the Python code, with revisions managed through GitHub and version control.

“We essentially have opened our development environment so consultants and customers can go in and look at that Python code and extend it.”Mike Landry, CEO, ketteQ

Q: How do you govern what each AI agent is allowed to do?

Landry says governance starts at the foundation: ketteQ’s software runs on AWS, and AWS Bedrock houses the LLMs and makes them interchangeable. Beyond that, security is defined in the customer’s data based on role and on what the agent is doing. Agents appear in a mission control setting where Quintus leads the crew, and each agent can be extended or changed over time.

“All that collection of agents you see in a mission control setting where Quintus is the captain of a crew.”Mike Landry, CEO, ketteQ

Q: How does a human override an AI agent, and how do you know when to?

Landry softens the idea that humans are always in the loop, because the world is changing so fast. At one $27 billion company, Quintus drafts supplier expedite emails with the reasons, and a human approves each one, so the person no longer has to find, prioritize or execute the decisions manually. As trust builds, the human can tell the agent to send emails or update delivery dates itself for specific suppliers, parts and situations, and ask only to be notified.

“At some point then I’ll say, okay, for these suppliers, for these parts, for these situations, Send the email yourself. I don’t need to see it.”Mike Landry, CEO, ketteQ

Q: When do agents make supply chain decisions and when do humans get involved?

Landry says humans stay involved, but it will look different by company and use case, and so will the pace of adoption. Today a planner may wake up to 75 emails from contract manufacturers and freight forwarders in Asia, so something that should take 5 minutes takes 3 days. Automating some of that back and forth, including unstructured email, can improve the company’s clock speed, and users can tell Quintus to respond a certain way next time, which gives it authority to act on their behalf.

“The latency between planning and execution is essentially added up to a slower clock speed for a business.”Mike Landry, CEO, ketteQ

Q: How do you audit what an AI agent decided and why?

Landry says an audit trail of who is doing what and when is table stakes for a planning system. Quintus also sees people make decisions, the impact of those decisions, and the automation of them, so the business can find insights in that audit. The aim is to improve the quality of decisions and workflow, whether AI or a person made them.

“Being able to find insights from that audit so that the business can improve on its decision-making and its workflow and the value and the quality of those decisions, whether it’s made by AI or it’s made by a person.”Mike Landry, CEO, ketteQ

Q: Which actions can an AI agent take without human approval?

Landry says “always in the loop” was a bold statement and that he does not know how much automation will be adopted in the next few years. Some use cases may not need a human in the loop, which is the latency he wants to collapse. Where people still decide, they will have more insight and tools to judge quality, and early adopters will pull ahead because learning and speed compound.

“We have to manage the risk, but early adopters are going to gain an advantage over the laggards unlike any other kind of technology adoption.”Mike Landry, CEO, ketteQ

Q: What still stops supply chain agents from running decisions autonomously?

Landry says it is not mainly an engineering problem. Companies cannot change overnight because of culture, processes and workflow, and some still use fax machines. He believes the technology is ahead of company readiness, and that leadership, readiness and the ability to change will be the gating factor on the path from planning intelligence to orchestration, automation and intelligent execution.

“I don’t think the engineering is going to be the gating factor with companies moving forward into the agentic world.”Mike Landry, CEO, ketteQ

Q: What is ketteQ’s approach to supply chain planning and orchestration?

Landry says ketteQ started in 2019 with a mission to provide supply chain planning and orchestration solutions that adapt to a changing world and the increased speed of business, from demand to supply and from customers to suppliers. It built a cloud-native, AI-ready architecture for companies that make and provide products across a number of industries.

“We’ve been able to build an architecture that was cloud-native and AI-ready…”Mike Landry, CEO, ketteQ

Resources & Documentation

  • Quintus, the ketteQ orchestration layer and user experience for supply chain planning

***

👇 Click to Read Full Raw Transcript

Swapnil Bhartiya: You may not be aware, but supply chains change faster than the plans that run them. Tariff shifts, demand swings, and the plan goes stale almost the moment it is published. Most large enterprises still run planning systems built 20 or 30 years ago, and the architectures are too old to leverage modern AI or cloud computing. Now, bolting on LLM on a sidecar doesn’t fix that because the AI only sees a copy of the data. It needs better integration. And when agents write and run code in production, hallucination, governance, and safety stop being theory and they become engineering problems. And to talk about all of these problems and the solutions today, we have with us Mike Landry, CEO of ketteQ. Mike, it’s great to have you on the show.

Mike Landry: Thanks. It’s great to be here. Appreciate you having me.

Swapnil Bhartiya: Let’s start with the foundational question. Tell us about ketteQ and what do you folks do?

Mike Landry: ketteQ was started in 2019 with a mission to provide supply chain planning and orchestration solutions. that could adapt to the changing world and increased speed of businesses from demand to supply, customers to suppliers. And so we’ve been able to build an architecture that was cloud-native and AI-ready and really provide some incredible capabilities for companies that make and provide products across a number of industries today.

Swapnil Bhartiya: ketteQ is 7 years old in a space that feels like it should be 70 years old. As we all know that supply chains have changed, evolved over time, the dependency, the flow, geopolitics, trade wars, even setting all of that aside, the economy is evolving, supply chain is evolving. Yeah,

Mike Landry: and you listed a few great examples, but there’s a number of disruptions on the supply side. Meanwhile, on the customer side, The expectations go up. The speed of changes and customer demands and marketplaces like data centers. And you can start to think about some of these places that, and those disruptions and demand-supply imbalance or disconnects we see show up in our daily lives, you know, through COVID and even before. But now the world is just operating faster and faster. And so to be able to catch those changes or disruptions or uncertainties and manage them, So that you’re running an efficient business, that you’re making promises to customers and delivering on those promises in an effective way is an ongoing challenge for businesses and supply chains in particular.

Swapnil Bhartiya: Many of your customers already run planning system, SAP IBP, Kinaxis, O9s, and others. When they ask why they need ketteQ on top of those, what is your justification?

Mike Landry: We talked about our company is 7 years old, but the folks that are in the company that are leading it have been around for a long time. So we understand the business challenges, the math problem, and have built something that’s contemporary on technology that, quite frankly, the older systems, like you called out a few of them, the Kinaxis in ’09, they were built on designs and architectures that are 20, 30 years old. And it just can’t leverage the AI technology that’s available. It can’t leverage the cloud computing that gives you a chance to provide Optimal answers at speed and scale. So what we’ve built does just that: optimal plans for demand and supply at speed and scale. And that’s important if you’re trying to drive real-time decision-making and manage the disruptions in a in a fast and thoughtful way. So what we’re able to do now is connect that layer of intelligence and orchestration on top of the other systems, which is a big advantage because the data is already organized for the purpose of supply chain planning. The ERP integration is built. The And what we get a chance to do is by connecting in minutes to where the data sits, we can essentially upgrade the experience and improve the outcomes by using the intelligence layer we call PolymatiQ and the orchestration layer we call Quintus to change essentially how they see and manage the business from suppliers to executives, from customers to suppliers. And there’s a number of use cases that we can point to that has just a tremendous amount of value. But the speed to value is very quick by sitting on top of the existing infrastructure the data and the systems that are already there and implemented years ago, but just not meant for today’s fast-paced world.

Swapnil Bhartiya: We are living in an AI moment. AI has been around for decades, but LLMs, generative AI, agentic AI, and MCP has changed the game. What role is AI playing in your own solutions?

Mike Landry: Yeah, on the decisioning layer, what we have in PolymatiQ, it’s basically agents, thousands of them running tries in parallel instead of one heuristic plan at a time, is trying thousands of heuristic experiments or decisions and bringing that back to make the best supply plan for make, move, buy, or inventory or demand. So in that point, agents are as if you had 1,000 planners running those tools. Now, no one would be able to hire that, but that’s the output that we’re able to generate with our PolymatiQ. On the user experience side, we have Quintus. And what we did, and you pointed this out already, is the LLMs are so powerful. In older systems with older technology, you’re having to put the LLMs as sidecars and they move the data there. And there’s only so many things you can do with an LLM and just a copy of their data. But for us, what we have is the LLM is put inside our product and we have a harness for that LLM. So the memory, the skills, the governance, the security is actually in our customer’s data. It doesn’t carry with the LLM. So we can literally swap the LLMs out in minutes and we can use a collection of them for what they’re good at doing. So Quintus provides a path to those LLMs and And it essentially has at its disposal to be able to reason through any question, task, or automated mission. It has the native Python engine, the PolymatiQ solvers, the ability to write SQL code, ’cause our data is in a relational data, and all our technology is an open technology stack. And all of that just gives what we call Free-Range AI. So we can answer any question, run any task, or provide any automated mission without any— and it goes from supply chain to data to user adoption to, you know, anything about the business, Quintus can kind of help you. gain insights and make decisions.

Swapnil Bhartiya: Let’s talk about how things actually happen in the real world. LLMs and agents hallucinate. They make mistakes. They pick side quests and they actually lose track of what they were supposed to be doing. So a planning system publishes a plan, conditions change, and the plan starts to go stale. What does an agent like Quintus do differently in production once that happens?

Mike Landry: Yeah. And like you’re pointing this out, it’s the ability— because we’re free-range AI, we’re not scripted like the sidecar. And so in that case, it’s novel questions with novel answers. And the world changes. And so how does it track and adjust to those changes is part of the big wow factor of the LLMs. And so in those interactions, it learns. Right. And we see our customers, we deployed this to 93% of our customers 3 or 4 months ago, and we see the prompts going up. And every time it does prompts for our customers, Quintus is learning and becoming the world’s best supply chain expert. We’re connecting Quintus to Copilot and to Claude. So all of a sudden now Copilot or Joule from SAP gets a PhD in supply chain planning because these agents are talking to each other. And Quintus has the ability to generate so much insights and intelligence for the purpose of the context of supply chain, for making and executing decisions. So the world of humans and machines will begin to build trust together and an understanding, and other agents will come alongside as well. But we have such a big head start because we see the prompts, and every time those prompts happen within our customer base, and it’s going up exponentially, they’re creating skills and missions and building understanding and trust with Quintus, and it’s able to do more and more with the complete confidence and governance that businesses need to take advantage of this agentic world.

Swapnil Bhartiya: Let’s talk about the security side. Quintus writes and runs Python code on the fly to answer questions. How do you keep that safe in production?

Mike Landry: The Python code, because we’re a native Python engine, some of that might be in production. And we, or we could turn those skills into production missions. And so there’s really a balance. You know, we look at the world when this started at the beginning of this year, all the vibe coding that companies wanted to do. And we saw companies that could only get so far. Because you have the reality of roadmaps and versioning and support and DevOps. So what we’ve done is we essentially have opened our development environment so consultants and customers can go in and look at that Python code and extend it. There’s millions of Python developers and thousands of Python libraries. That gives us an advantage to adopt the ones that are going to be helpful for a given business or a use case that they’re trying to address. Being able to do that in a thoughtful way that manages those revisions with GitHub and version control and all of that is part of the extendability of our platform. So we have customers that are handling use cases like, you know, customer escalation or supply expedite that are outside planning but leverage the planning intelligence and automation and orchestration of our technology.

Swapnil Bhartiya: What does that integration look like? Are customers gluing the pieces together, or is it a single pane of glass where they work through the systems easily?

Mike Landry: You know, because of the advent of, you know, data integration is very different in the agentic world too. And so with MCP connecting to an API that already exists with these other vendors, or when we’re implementing our own platform and we’re going to the API of the ERP systems, we essentially have trained Quintus on the data model. of these systems, of like something like Kinaxis has a hierarchical data structure. It’s not relational. It’s very foreign and unique. Our team deployed it 80 times back at our, before we started ketteQ. So we have a really strong opportunity to train Quintus on what that data structure is like and how to interpret that hierarchical data model for the purpose of a supply chain context and reasoning. And so with that training and an MCP access, within a minute, we basically access that data structure right where it sits to be able to read and write. And we have customers like a pet food company called PPF. It’s a billion-dollar pet food company. They’ve been using the solvers that we have to essentially write answers right back into Kinaxis as if it was generated there. And that’s just a big advantage because they were able to— we were able to deploy that in weeks without having to change out the platform. That’s a much bigger decision. And a longer process. So when we can add intelligence and orchestration right on and deliver it in minutes, deploy it in days, we just have a lot of opportunity to prove value before a customer even makes a— has a cost or a commitment.

Swapnil Bhartiya: Where are most of these decisions actually made? When the customer goes live, what does that actually look like? And what is the timeframe there? Is it that agents are making the calls? And when do humans get involved, or are they totally out of this loop?

Mike Landry: Yeah, no, I think humans will always be in the loop, as they say. But it’s going to look different from company to company and their pace to adopt agentic automation is going to look different by company to company as well, or use case to use case. So, you know, today the world moves at the speed of humans. And so there’s latency. We have customers that will get emails from their contract manufacturers and freight forwarders in Asia. And some of our planners will wake up, there’ll be 75 emails. And then by the time they get through it and respond in terms of what should be made and what should go on the next container, the person who’s receiving that email may be done for the day. And so then they get up the next morning and process that return. So something that should take 5 minutes takes 3 days. And the latency between planning and execution is essentially added up to a slower clock speed for a business. And if we can automate some of that back and forth, some of the system updates, even when it comes to unstructured data of spreadsheet— of emails, we have an opportunity to really improve the clock speed of the company to remove that latency, particularly for global companies. So that’s really what we’re in the business of doing. But it comes down to just like you point out, who’s making decisions and when can the machine take some tasks or even take a decision when it’s appropriate. An interesting thing is the way that systems are changing. Before, that would be a rule or parameter that someone would have to go to another screen and change where it would be an auto-approval. Well, now with Quintus and the world of agents, as learning happens, you can just basically tell Quintus, hey, next time when a customer, when a supplier tells you this, you can go ahead and respond that way. And so in that conversation, really, with the agents, it can learn and take on authorization from the user to go and start to do things on their behalf. So it’s a— those are the kinds of things that will happen at the user level and the company level as technology is implemented to work with humans to improve the overall clock speed of the company.

Swapnil Bhartiya: Quintus works through a network of agents, a demand agent, a supply agent, and inventory agent. How do you govern What each of those agents are allowed to do and not allowed to do, because that can be a real challenge.

Mike Landry: Yeah, for sure. And so there’s governance and security that happens at the foundational level. So all of our software runs on AWS and there’s AWS Bedrock that provides the LLM housing and interchangeability. And so there’s security and governance at the foundational level there. And then in terms of what the agents do, And what they’re allowed to do. There’s security that’s defined in our customers’ data from the governance and security based on the role and what the agent’s doing. And so all that collection of agents you see in a mission control setting where Quintus is the captain of a crew. And so you can really kind of organize that way and extend or change any of those agents over time. And the interesting thing is we see our customers using Quintus in ways that we didn’t anticipate. So of course, the supply chain questions, what are the supply orders that impact important customer orders and this quarter’s revenue? There’s a lot of reasoning that has to happen with that kind of question to gain the insights that someone’s looking for, because this ties finance and sales and suppliers all together to see how they interchange and how they relate. Well, now we see customers asking things like, where are the inconsistencies in my data? How did you find that? Where do the ramifications of that and how do I fix it? So it’s very much a different question, but it’s one that always happens when you deploy planning is we don’t have the data. And so the agents can help with that too. Another challenge with deploying these decision support systems is always user adoption. Well, guess what? We see customers asking Quintus, how are the users adopting it? Who’s making overrides? Who’s making approvals? When are those decisions good ones? And so Quintus is helping guide the deployment of our technology as well as the technology itself, which is fascinating to start to see. You mentioned

Swapnil Bhartiya: humans will always be in the loop, but when an agent like Quintus is about to How does a human step in to override it if needed? And how would you even know if an override is needed?

Mike Landry: Yeah. And I say always in a loop is, that’s kind of a strong statement because the world is changing so fast. I mean, we see it this year. So I guess I would, might soften that a little bit. But when we see companies deploying this, like one company is a $27 billion company, they’re not going to have it write emails to suppliers. It’s going to write, here are the scenarios suppliers that have a need to expedite. Here’s why. And here are the emails I’m ready to send. So they’ll send those emails, but it’ll be on a human deciding yes. So the human won’t have to find, prioritize the decisions, or execute the decisions manually. So it’s going to save the people that do those expedites today a lot of time. And it’s going to build trust. So at some point then I’ll say, okay, for these suppliers, for these parts, for these situations, Send the email yourself. I don’t need to see it. Same thing coming back. Suppliers send emails back. It’s still a manual communication for today. But when they come in, it triages. Okay, here’s 40 emails. And all you have to do is update the estimated delivery date. Do you want me to do that in the system? Yes. Okay, good. Next time, update the system yourself. I don’t need to see it. Just let me know that it did, that you did it. And so you can do it right away. Here are 10 emails that is asking for a new date, and these dates are okay. You’re still going to make the customer order. Do you want me to update those dates? Yes. These dates is not going to make the customer order. Do you want me to expedite? Here’s an expediting fee. Do you want me to do that? So as it triages, it will be able to categorize what is It thinks it should do, gets approval, and then just lets the human be still in a loop to add context and judgment. And all of that’s— we came from consulting. And this is going to, I think, going to open up a renaissance or a tremendous opportunity for consultants in that process that I’m describing, because it is a process of transformation. Yeah.

Swapnil Bhartiya: We talked about humans overriding decisions. Let’s talk about the next step. After an agent has made a decision, how do you audit what was decided and why? Because that is also critical.

Mike Landry: Yeah. And so I think that’s part of the wonder of this system is that, as I mentioned earlier, Quintus is involved in seeing people make decisions, the impact of those decisions, and as well as the automation of those decisions. And so to be able to provide an audit trail That’s like table stakes for a planning system, who’s doing what and when, or even in the ERP and transaction system. But being able to find insights from that audit so that the business can improve on its decision-making and its workflow and the value and the quality of those decisions, whether it’s made by AI or it’s made by a person.

Swapnil Bhartiya: Earlier you used the term humans are always in the loop, but let’s think about a world where things are more automated. Let’s assume Quintus runs a mission override and agents like procurement agents process purchase orders. Which action can an agent take on its own without a human approving them first?

Mike Landry: Yeah, so, well, yeah, like I was saying earlier, I said a bold statement, always in a loop. That’s I don’t know how much automation is going to be adopted in the next few years. And there’s probably some use cases where they’re not in the loop and they don’t need to be in the loop. And that’s the latency that we’re talking about collapsing. But in terms of when people are making those decisions, it will just be made with a lot more insight and tools to be able to make those decisions faster and better. And then, like I said earlier, to judge the quality of those decisions. And then everyone’s learning together, the agents and the people. It’s a new way to look at the business. And so there’ll be, you know, those that are able to adopt and gain an advantage in speed and decision-making over their competitors faster, that learning is just going to compound. And so I think this is a place where we have to manage the risk, but early adopters are going to gain an advantage over the laggards, unlike any other kind of technology adoption, because the learning and the speed compounds for those

Swapnil Bhartiya: Things are moving fast and a lot is still unsolved. What engineering problems still need to be solved before supply chain agents can safely run more decisions on their own autonomously?

Mike Landry: Yeah. Hmm. I mean, that’s a good question. I think there, you know, companies are still don’t, can’t change overnight. There’s culture, there’s processes and workflow. I mean, companies are still using fax machines in places. And so the speed of adoption and transformation is one that is not necessarily an engineering problem, but is one from a leadership, I believe, is going to be the biggest advantage or challenge that a company might have. We already see, I believe, what we’re doing right now is Blazing a trail for a company and enabling that trail to go from planning intelligence to planning orchestration to planning automation to then intelligent execution. And so the technology’s really there. And it’s quite frankly ahead of companies ready to adopt that technology, in my belief. So I don’t think the engineering is going to be the gating factor with companies moving forward into the agentic world. It’s going to be their leadership and readiness and ability to change. and transform, that’s going to be the gating factor more than engineering, I believe.

Swapnil Bhartiya: Mike, thank you so much for taking time with us and chat with us and try to explain this challenge that enterprises are facing when it comes to the whole supply chain. Thanks for your time. And I would love to have you back on the show. Thank you.

Mike Landry: Thank you. I really enjoyed the conversation and it was fun and interesting. And I appreciate the way that you kind of led the conversation on a natural course. Appreciate the opportunity and the chance to have this dialogue with you.

What Data Does Continuous Threat Exposure Management Need? | Romanus Prabhu Raymond, ManageEngine | TFiR

Previous article