Agentic AI has been the thing this year, buoyed by the notion that Large Language Models (LLMs) can fill in the cracks in existing automation approaches. Most vendors of existing Supply Chain Management platforms (SCM) are adding LLM-enhanced agentic capabilities to their tooling mix. Also, Decision Intelligence vendors are making considerable progress in streamlining supply chain processes with better data management and more traditional AI techniques.
Tim Mitrovich, CEO of Artisan Studios, argues that although these tools can help, there are still significant gaps to address in improving SCM workflows. Data is often fragmented, coming from many different systems and in various formats. And the data used frequently does not represent the same point in time.
As a result, stakeholders and decision-makers find it hard to trust this data and often rely on their experience or tribal knowledge to solve their own problems. They also often err in not solving the problems of their greater organization. This can occur due to misalignment of objectives or a lack of understanding of other organizational goals.
The platforms have made progress in modeling known problems and helping to automate routine decisions. But this also requires deep modeling of the problems, data cleansing, and access to more and cleaned data. After working with many enterprises to solve practical SCM challenges, Mitrovich also finds these larger platforms take time to spin up:
The platforms also took a considerable amount of time to implement and deliver value, with some nearing twenty-four months for companies to start seeing the payback. Many enterprises weren’t able to do complete data cleansing in this time, falling short of their goals. Finally, many of the platforms have limited collaboration with humans within the organization and with suppliers. This can lead systems to mis-model supply chains with a lack of context or a lack of validation with suppliers.
The tactical use of agents as intermediaries promises to give organizations a faster time to market with newer, less expensive SCM solutions without changing platforms. Employees and suppliers can continue to engage with their models and collaborate on optimizations to deliver value to the business and its customers.
New agentic approaches
But why should enterprises rush into agents if legacy automation tooling could support many of the same objectives? RPA bots and process intelligence development tools are all making great strides in expanding to support more automation and integration use cases. Mitrovich says:
RPA bots can work well when everything is known, and there aren’t issues or opportunities to optimize. However, RPA bots often break and stop when they encounter unknown situations, and this can slow your supply chain.
It's possible to expand RPA bots to handle more edge cases. But this requires more investment in identifying new edge cases and deciding how they should be handled. These updates can add complexity and increase maintenance costs. Mitrovich believes that newer AI agents can adapt to more situations based on instructions, roles and tools by using reasoning via LLMs when handling tasks.
For example, in the case of a mismatched purchase order, they might examine a missing line item, identify possible alternative suppliers, determine which supplier(s) could meet budget and time constraints and place an alternative order. They could also alert downstream processes and send notifications to the end customers. Mitrovich says:
Unlike RPA bots, you don’t need to identify all the failure types and paths up front, simplifying your solutions and keeping your costs in check.
Emerging use cases
Mitrovich has been seeing a variety of new agentic solutions that demonstrate some of this flexibility by using LLMs. For instance, shipping lines and manufacturers that rely on overseas suppliers are augmenting their supply chains with agentic LLMs to identify shipping delays based on unstructured data, soft signals, emails, weather reports, labor disputes, and other data sources.
Without agentic LLMs, this would have required companies to train AI models on each type of data or have people study, report, and review consolidated data. The amount of data they would review would be limited, based on time. Using LLMs, Mitrovich estimates that companies can identify potential disruptions three to seven days in advance, allowing them to begin re-planning production or finding delivery alternatives.
Another example is in contract analysis and compliance, where companies use LLMs to review entire contracts and highlight non-standard or out-of-compliance terms and conditions. LLMs are not limited to reviewing terms or fields of data in these documents, unlike prior natural language processing (NLP) techniques.
Companies are also using LLMs to analyze unstructured data, such as reviews, and feeding the information back to their product planning departments to identify new features or changes to improve their products and public perception. For example, a clothing company may see several reviews indicating that sizing runs small relative to consumer perception in its LLM-analyzed review data. They can adjust this in several ways, such as providing a better sizing chart for customers or trying other experiments to improve this perception of their products.
Mitrovich has also been tracking research at universities and in the corporate world regarding the use of LLMs in SCM. This includes how to fully automate the decision process and remove the human-in-the-loop, how to prevent LLMs from hallucinating, how to identify deeper causal relationships between market data and supply chain events, and how to identify hidden steps within defined standard operating procedures (SOPs) that prevent optimization.
Mitrovich believes that the current generation of LLMs benefits from training data from many different industries and processes. This allows them to consider things an SCM professional might not be aware of. This will enable them to propose interesting solutions that are not present in the daily thinking of those focused on the problem. This thinking might not be novel to the world, but it could be novel to the problem at hand. But he cautions:
Implementors need to take care in balancing where they allow the agent to identify this type of solution and allowing the agent to make a decision based on novel ideas. If you find the middle ground, you will find agents that can come up with solutions that can be approved via human in the loop processes that might not have been surfaced via traditional or prior methods. Too much on the edges, and you find yourself in chaos or limited by traditional thought processes.
New guardrails required
Early adopters are seeing some success in automating multiple independent SCM processes in a trustworthy way, including manufacturing, automotive, and hospitality. But it's also essential to think through new guardrails for decision authority, data provenance and traceability, and feedback loop and stability controls. Mitrovich explains:
Careful planning of guardrails, human-in-the-loop decision authority, and monitoring and feedback will help prevent the rise of many trust and safety issues. However, just like the humans that made them, the AI agents are not foolproof. Companies that have agents with no feedback and no measurement and those who don’t consider human-in-the-loop agency, will risk exposing these issues to their customers. Internally, companies need to consider the impact to their employees and look for ways to encourage them to add value while the agents perform their tasks.
Decision authority management concerns the point at which teams can trust agents to make decisions rather than requiring a human decision maker in the loop. In some cases, one bad decision can snowball into many, making a small issue large. Companies need to assess the risk/reward of decisions, decide on the level of decision authority they feel comfortable with, and set guidelines to ensure this is followed.
Data provenance tracking and traceability are required to ensure teams can drill down into how decisions were made and on what data. Mitrovich explains:
Poor decisions will happen and, just like from bugs in traditional software, being able to trace why that decision happened is important to help improve your solution. To do this properly, there needs to be a focus on providing structured data about the decisions and not just providing logs of the decisions.
Feedback loops and stability control help provide input to the agents regarding decision-making and interaction with enterprise systems to improve the agents' operations. For example, teams sometimes discover an unexpected entry point into the ordering agent. When this allows the agent to take the right action, case reinforcement learning could encourage that correct behavior. Teams may also want to limit these behaviors, perhaps to once a day, or even block them in the future if they are unsure of potential problems they may cause.
Start small
Mitrovich finds the best way to start an agentic deployment is to pick something small, simple, and measurable. A good one is the type of problem that employees might spend time on but only adds little value. Also consider looking beyond a single-process problem, even if the second agent is doing nothing more than suggesting how to address any problems found by the first agent's work.
Another essential step is becoming familiar with data quality and cleanliness, a problem many organizations face. This will inform you of how to prepare internal systems for future steps. The process of building and interacting with agentic workflows can help teams set up any infrastructure or identify services required to support future expansions.
It's also helpful to work out the implications of different types and approaches for guardrails as part of this process. On the tools and technical side, this can be informed by the problems you are trying to avoid. Mitrovich walks through a compliance example:
If you need to ensure compliance, you’ll probably want a retrieval-augmented generation (RAG) solution, which will need a vector database to store your inputs. From there, you can require citations to ensure the correct data has been used to formulate the decision, and then a policy engine to ensure your business rules, including regulatory requirements, are followed. You’ll also want to make sure that inputs to and outputs from your agent match your schema. Finally, you need to think about how to store the inputs and outputs from your agents so that you can support them with continuous improvement.
It's not necessary to implement all of this at once. A better approach is to build a proof of concept, then pick the one or two most important technical items to try to build into it. They may be the foundation you need most, or the items your team needs the most experience with setting up and working with. This will help you develop skills in that specific area when you move past POC and into production.
In terms of processes, it's important to consider where you want the human-in-the-loop in your agentic workflows. This could be as a final decision-maker or as an observer to the process. It's also important to capture the recommendations and decisions the agents make, and the results, to track the accuracy and confidence in the system. This process can also help feed information about positive and negative decisions back into the agents to support a feedback mechanism, enabling the agentic system to learn and improve.
Planning for problems
One of the biggest and first issues with these projects relates to data governance and data readiness. Mitrovich explains:
Many companies think their data is ready because they use platform X and their data is cleaned appropriately and well understood. In some cases, internal policies have failed or been ignored, and data in downstream systems becomes harder to match to the system of record. Often, we find that data is not clean, having duplicates or contradictions.
It's important to identify the root cause of these issues by working backwards from end goals. This can help shore up processes and tools to improve data quality. Performing an audit of data governance practices and ensuring that data provenance and data lineage are tracked can help raise these issues earlier rather than later.
Mitrovich predicts that LLM-powered agentic AI might allow some companies to completely reinvent their supply chains over the next year. This will present new opportunities for supply chain analysts to think beyond projections and models to identify ways for companies to disrupt their industries.
This could be supported by industry initiatives on SCM-specific LLMs, skill systems and SOPs being built into agent infrastructure. Vendors may also start innovating on top of these and packaging them into SaaS services for the smaller enterprises. He says:
Agentic AI is going to fill needs that the largest players in SCM started to fulfill, but fell short of. The large SCM solutions aren’t going away, but the complexity of working in them will shrink as we see more and more agents built and deployed.
My take
With all the hype around agentic AI, it can be difficult to determine whether it's old wine in new bottles or a genuine opportunity to try something completely different. Mitrovich offers an interesting perspective on how these Agentic LLMs might help connect the dots across existing tools and platforms more efficiently in the supply chain.
A measured approach to exploring these new tools, with an eye toward better decision auditing, data governance, and feedback loops, seems like it could apply to both existing human processes and the newer ones.
