Webinar Recap: Beyond the Control Tower — Agentic AI and the Autonomous Supply Chain

Summary

In our Future.Now webinar, “Beyond the Control Tower: Agentic AI and the Autonomous Supply Chain,” we explored what happens after a supply chain problem becomes visible. A decade of investment has given most organizations real-time insight into their operations, and that investment has paid off. Control towers, dashboards, and exception alerting surface problems as they emerge, and often before they land. The difficulty lies in the distance between seeing a problem and resolving it, a distance measured in data pulls, cross-functional meetings, and decisions that arrive after the window to act has closed.

We looked at how agentic decision intelligence closes that distance by adding a layer of recommendation, execution, and impact tracking on top of the systems already in place. A live demonstration on real supply chain data followed a single excess inventory risk from detection through simulation, approval, and execution in the source system. We also examined how autonomous decisions stay governed as trust in them grows.

Key Takeaways

• Visibility is largely solved. Decision speed is not. A decade of supply chain investment has delivered real-time insight, and it does what it was meant to do. What remains open is the interval between seeing a problem and acting on it, an interval filled with data collection, cross-functional meetings, and escalation. Excess inventory, stockouts, longer lead times, and rising logistics costs can each be traced back to a decision that arrived late.
• Granularity and scale determine whether decisions actually get made. Many organizations decide at the SKU or SKU-location level when the situation calls for batch-level detail, and expiry risk hides in that gap until it surfaces at quarter end as a write-off. Scale compounds the problem, since a single region with a dedicated team may perform well while that same diligence never reaches the rest of the network. One organization was carrying roughly 15 million euros in short-dated, blocked, or excess stock. Another was seeing 30,000 shortages a week, a volume no team can work through by hand. Aera operates at the granularity where those losses can still be prevented, and applies the same rigor across every SKU and every region.
• Assembling the data is the largest time cost. Audience polling placed data collection ahead of root cause diagnosis as the biggest drain on time, with more than half of respondents selecting it. Much of what matters sits outside structured systems, in supplier contracts, policy documents, and email threads, and it gets compiled manually if it gets compiled at all. Aera ingests both structured and unstructured sources into its Decision Data Model, which shortens the path from question to root cause considerably.
• The missing layer is recommendation, execution, and impact tracking. Real-time visibility sits at the bottom of the stack, and root cause analysis, scenario simulation, and risk response sit above it. Both are widely deployed today, though usually in separate systems owned by separate teams. The layer above those two is where agentic decision intelligence operates, recommending decisions, executing them, and measuring what they returned. Impact tracking carries more weight than it might appear to: without a record of what a decision was worth, a delegation threshold cannot be governed, improved, or defended.
• One foundation supports structured and situational decisions alike. Highly structured decisions with a known playbook can run with the human out of the loop, which is where the largest automation opportunity sits. Decisions that call for validation keep the human on the loop, with each recommendation surfaced for acceptance or rejection. Situational questions, such as which SKUs have fallen below safety stock during a disruption, keep the human in the loop through Aera Intelligence, a conversational interface that draws on the same data and the same actions. All three rest on decision memory, continuous learning, and stewardship.
• Governance is what makes autonomy practical. Thresholds are set by decision type and by value, guardrails are enforced in code on every call and written into policy, and every decision carries an audit trail. Nothing acts beyond the authority of the person who authorized it. Aera’s confidence scores, built from the outcomes of prior decisions, let organizations automate above a chosen level and route everything else for review. A control room view within the platform tracks how recommendations are performing, what value they have returned, and which root causes keep recurring.
• Value arrives in weeks rather than years. Decision support through Aera Intelligence can go live in one to two weeks, once systems are connected and questions can be asked against governed data. Structured skills and agents follow in two to six weeks, generating recommendations that teams accept or reject. Full automation of high-confidence, high-volume decisions typically lands in eight to twelve weeks. Hershey deployed three loss-prevention skills on that path and reported 2.7 million dollars in savings within 11 weeks, along with three times as many shortages identified as its manual process had caught.

Speakers

	Suraj Ramalingam, Senior Solution Engineer, Aera Technology Suraj brings extensive experience in global supply chain management, having held senior roles at Salesforce and Procter & Gamble. At Aera, Suraj focuses on enabling customer success through technology-driven transformation, helping organizations operationalize AI-driven decision-making at scale.
	Aditya Verman, Sr. Manager, Product Marketing, Aera Technology Aditya works at the intersection of AI and enterprise operations. His focus: turning complex, AI-driven capabilities into clear business value. He brings a B2B software background and a close eye on how autonomous decision-making is reshaping the enterprise.

Full Recording

Access the full webinar recording here.

Q&A

During the presentation, attendees submitted questions. Below are several of the key questions and answers.

Q: We already have a control tower. Does this replace it?

A: No. Your control tower keeps doing detection, and it remains good at exactly that: showing you what is happening across the network. Aera handles what comes after the alert, which is the intelligence and the automation of the decision itself. Hershey runs Aera alongside Kinaxis, S/4HANA, and JDA, and nothing was ripped out to make room for it. Think of it as an intelligence layer sitting on top of what you already own.

Q: How is this different from putting an LLM on top of our data?

A: The language model participates in the reasoning rather than conducting it. Decisions run against your constraints, your objectives, and your policies, and the model never makes the access control decision. Reasoning is one part of a larger system that also includes deterministic engines, policy enforcement, and execution back into source systems.

Q: If deterministic logic is cheaper and more reliable, why use agents at all?

A: Deterministic logic is the preferred path, precisely because it is cheaper, more reliable, and easier to scale. The limit is how far it can carry a process before a person has to step in. In a five-step process, deterministic engines might have taken you through step three, at which point a user was pulled in to handle steps four and five. Language models extend that reach, so the human is engaged later and less often. The aim is to minimize the back-and-forth before a final decision, and the model is used only where the deterministic engine cannot go.

Q: How do you stop the system from making a decision you would not have approved?

A: Three mechanisms work together. You set the thresholds by decision type and by value, so not everything has to be fully automated. Guardrails are enforced in code on every call and written into a policy document, so each decision passes through them. And every decision carries an audit trail you can review and learn from. Nothing acts beyond the authority of the person who authorized it, and when a decision needs human approval, Aera flags it. The platform is also self-service, so customers configure and adjust these rules themselves.

Q: Within Aera Intelligence, can you ask the agent through the chat interface to carry out one or more of the listed actions, and it routes them for execution? What happens if a use case has not been documented? Will the agent fail?

A: You did hear that correctly. Aera Intelligence executes actions that customers have already given it access to. Where an action is identified but not yet defined, Aera engages the human to either grant that capability or execute it outside the platform. The agent goes as far as it can and brings a person in to fill any gaps. It will work out how to solve the problem, but it only executes what you have given it permission to execute.

Q: You mentioned that Aera can take in procedures and unstructured documents. Does that mean it can connect to SharePoint and document management tools? How do you ensure it does not pull information from sources it should not reach?

A: Yes, and this can be done through agents as well. Aera reads only the sources your administrators explicitly configure and permission, so there is no open-ended crawling. Access is defined at the connector level, with role-based access control governing who and what can reach each source.

Q: What does data quality management provide? Can it correct or enrich data?

A: Both. Master Data Management Skills run checks for accuracy, consistency, and completeness, monitor quality through a control tower view, and generate correction recommendations, with the option to write corrected values back into the source ERP such as SAP. Aera also enriches master data from observed reality, for example by comparing actual lead times against the values recorded in master data and recommending updates. The data improves alongside the decisions built on it.

Q: What does the data foundation require from us?

A: Less than you might think. The Decision Data Model™ goes live in about four weeks, supported by ready crawlers and connectors, and it reads from your existing systems. You do not need to consolidate your data estate first, which is the step that usually costs a year.

Q: What systems and data infrastructure typically need to be in place before an organization is ready for decision intelligence? If Aera is making and actioning decisions at the SKU level, how much granular visibility is required?

A: Aera typically hydrates its Decision Data Model with data from ERP, APS, data lakes, and other transactional systems that hold data relevant to the decision. The usual approach is to start with the main systems and add others based on the use cases you want to put into Aera, rather than building out a complete picture in advance.

Q: Can the platform decide to scrap stock, or propose that slow-moving inventory be sold while it still holds value?

A: Yes. The Ageing Inventory Management Skill identifies slow-moving, obsolete, and excess stock, including a pre-alert for items likely to become obsolete. It then recommends actions such as transfers to locations with demand, order modifications, and discounted sales where the stock still holds value.

Q: Where should we start?

A: Start with your most pressing problem, using one skill, one function, and your own data. The best candidates are decisions that repeat at volume with a clear cost and service trade-off, and that you are simply not able to get to all of today. Stockout prevention and excess inventory are the two most common entry points. From there, that first deployment becomes the basis for a roadmap toward more complex decisions.