Webinar Recap: Agentic AI in Action — Redefining Enterprise Decision-Making with Aera

• Agentic AI extends LLMs into live decision processes.
  Traditional use of LLMs has centered on generating text or insights, often detached from operational workflows. With agentic AI, that capability is embedded directly into decision-making systems, allowing agents to reason in real time while connected to enterprise logic. This enables organizations to shift from static recommendations to dynamic, continuous decision-making that drives measurable outcomes.
• Autonomy can scale without loss of control.
  A frequent concern with AI-driven decisions is the balance between autonomy and oversight. The webinar emphasized that Aera ensures transparency and governance across every action taken by agents. Guardrails such as explainability, traceability, and auditability mean that decisions can be automated at scale without eroding the trust of business users or compliance leaders.
• Unstructured data strengthens decision flows.
  Most enterprise systems focus heavily on structured data — transactions, forecasts, and plans. But much of the critical business context sits in unstructured formats such as emails, contracts, and documents. Agentic AI now incorporates this unstructured data directly into decision processes, widening the scope of information that informs decisions and improving the ability to anticipate and respond to risks.
• Agent teams orchestrate complex decisions.
  The live demonstration showcased how multiple AI agents can be composed into a team that works collaboratively. In the supply chain risk management example, one agent detected a disruption, another simulated response scenarios, and others evaluated feasibility and executed actions. This orchestration mirrors how cross-functional teams of people work together, but with far greater speed and consistency.
• No-code tools accelerate adoption.
  Deploying intelligent agents doesn’t require advanced programming skills. With Aera’s no-code tools, enterprises can quickly compose, configure, and embed agent teams into existing decision flows. This lowers the barrier to adoption, empowering business leaders and functional experts to create AI-driven decision processes without relying solely on technical teams.

Access the full webinar recording here.

Q: Can you explain what is more beneficial — RPA or agentic AI — in terms of business value from an organizational perspective?
A: RPA is primarily focused on reducing FTEs by automating tasks people used to perform manually. The system monitors human activity and then mimics those actions — such as keyboard inputs — to reduce call center headcount or other task-based labor. Its business value lies almost entirely in task automation and cost reduction.
Decision intelligence, enabled by agentic AI, goes further. The first value driver is the ability to make new decisions that people cannot make today simply because they do not have the time or capacity. The second is the ability to make better decisions by bringing more context into the process, removing bias, and leveraging traditional AI and machine learning models. This improves the accuracy, granularity, and frequency of decisions, increasing the value of each one. The third is reducing the cost per decision by automating and digitizing the process, similar to RPA. The difference is that agentic AI operates at the level of intelligence: it is adaptive, contextual, and capable of reasoning, while RPA was never designed to be intelligent.

Q: How will the steps of understanding and recommending differ for implementation partners like Deloitte when developing skills in Aera using agentic AI?
A: In the design phase, the key is to decide which parts of a process you want to make more agile and adaptive, and which parts should remain under very strict, deterministic control. This trade-off is essentially between non-deterministic and deterministic reasoning.
With non-deterministic reasoning, powered by LLM agents, you can address a much wider range of situations — especially those that are too complex or variable to codify in rules. At the same time, some decision steps must remain tightly controlled, and for those, traditional deterministic approaches such as decision trees remain appropriate.
Customers are already experimenting with this blend. For example, some have taken an existing deterministic skill, one that uses machine learning but still relies on deterministic logic, and augmented it with agents. These agents apply learnings from past decisions, managing modifications or overrides automatically. At the tail end of the decision process, an agent can decide to change a recommendation when user behavior in the past has shown a different choice to be more effective.
This is critical to skill design. The point is not simply to “plug LLMs everywhere,” but to identify where they bring unique value — such as reasoning with unstructured data, orchestrating processes without codifying every scenario, and dealing with high variability. Agents also add the ability to learn from historical user decisions more easily, making skills adaptive and continuously improving.

Q: Does the Aera AI agent need to be both autonomous and intelligent, or can it simply automate tasks?
A: Aera is a platform, which means there is flexibility to support many use cases. The most impactful use cases are centered on decision-making, but automation of tasks is also possible. As shown in Laurent’s demo, tasks can be automated within the broader context of intelligent decision-making.

Q: What technical sophistication is required to interact successfully with an agent?
A: Aera is designed for business users, not technical experts. Users do not need advanced training to work with agents. Interaction happens through natural language or guided interfaces, making the platform accessible to business professionals across functions.

Q: If OpenAI GPT is used for LLMs, how is data confidentiality maintained?
A: OpenAI is only one option. Aera is not an LLM provider but a platform that can connect to whichever LLM endpoint the customer chooses. Enterprises can configure secure, approved endpoints for any LLM they want to use. In Laurent’s demo, OpenAI was used, but the endpoint could just as easily have been an internal or alternative model.
All exchanges with an LLM happen via secure APIs. There is no data storage or copy retained on the LLM side — it is strictly send-and-receive. This is similar to how enterprises have long managed secure connections with other systems. The result is 100% locked-down confidentiality, regardless of whether OpenAI or another provider is used.

Q: Is the email parser an out-of-the-box module in Aera? Can it parse attachments and summarize?
A: Yes. The email parser is an out-of-the-box agent module, and it is capable of parsing attachments as well as summarizing them.

Q: How do you get from the outcome of an agent team to a recommendation in a Skill?
A: The outcome of an agent team is stored as structured output, typically in JSON format. That output updates the target subject area, and from there the rest of the process is handled just as it would be for any other Skill, resulting in recommendations being populated for the user.

Q: Does Aera provide predefined components for supply chain and marketing, or is it a blank canvas?
A: Aera provides predefined components called Aera Skills, which are composable capabilities for digitizing decisions across different domains. These cover functions such as supply chain (orders, inventory, etc.), finance, marketing, and more. Customers do not need to start from scratch unless they want to build something highly customized.

Q: How does Aera interact with planning applications such as Blue Yonder and o9 in supply chains?
A: Aera is built to orchestrate and enhance existing planning ecosystems rather than replace them. Integration happens via APIs with major planning platforms such as Blue Yonder, o9, and Kinaxis. Many clients use Aera alongside APS systems to bring decision intelligence into their current planning processes.

Q: How can agents improve forecast accuracy using two different data systems?
A: Agents improve forecast accuracy in two ways: by learning from user behavior and by working with multiple data sources.
First, they analyze how business users interact with forecasts, especially where overrides are made. Forecasts generated by machine learning models usually fall into groups: some product sets show low accuracy, others are very smooth, and some are consistently high in accuracy. Overrides happen frequently, but organizations often do not track whether these interventions improve or worsen accuracy. Agents can record and analyze this behavior, identifying patterns such as chronic over-forecasting or under-forecasting. Over time, the agent can recommend where user intervention adds value (e.g., new product launches or volatile categories) and where it tends to create errors. This reduces wasted effort while improving the overall forecast.
Second, agents can support forecast accuracy improvements that involve multiple data systems. In this case, demand sensing techniques come into play. This typically involves traditional machine learning or deep learning models that incorporate external signals like point-of-sale data. The goal is to improve short-term accuracy, granularity, and frequency. While not inherently “agentic,” these ML-driven approaches complement the agent’s role by strengthening the base forecast, which the agent can then monitor, adjust, and enhance using its own learnings from user behavior.

Q: Are there MLOps-style hooks to monitor agent performance and drift?
A: Yes. The agentic framework includes performance testing, logging, explainability, and monitoring tools to track agents. These are built into the development experience.
The platform supports no-code, low-code, and pro-code development. Users can interact through notebooks such as Jupyter, through the visual interface, or programmatically via SDK. AutoML capabilities are also included, enabling experimentation, correlation analysis, and drift detection at both the data and model levels. This ensures ongoing oversight of agent performance and reliability.

Q: Have you deployed learning agents to improve forecast accuracy?
A: Improvements to forecast models themselves are usually handled through data science and machine learning techniques. Agents, however, play an important role by identifying and explaining the drivers behind deviations in forecast accuracy. By capturing these learnings, agents help improve the structure of forecasting models and ensure that deviations are better understood, ultimately contributing to higher accuracy over time.

Q: It seems that Aera uses a data crawler to bring information from different systems. How is the knowledge graph for this data built? Does it need to be provided externally, created by the user in Aera, or does Aera build it automatically?
A: Aera provides prebuilt crawlers with mappings to ERPs such as SAP, which can be configured and customized as needed. At the core is the Decision Data Model — the master repository of enterprise information represented in a unified data model. It is continuously fed by real-time data streams and includes thousands of predefined datasets across domains such as master data, production, and purchasing. Organizations can use these out of the box or tailor them to their own business context. In addition, the Data Workbench allows users to access internal and external data, refine and harmonize it, and then publish it into the Decision Data Model for use in decision-making.