Webinar Recap: Decision Intelligence at Speed — The Agentic AI Advantage

Summary

In our Future.Now webinar, “Decision Intelligence at Speed: The Agentic AI Advantage,” co-hosted with Deloitte, we explored what agentic AI changes about the way enterprises decide and act. Twenty years ago, supply chains rewarded cost, efficiency, and a plan carried out well. That world has moved on. Late suppliers, transport delays, and equipment failures now tend to arrive several at once, so the advantage belongs to those who adapt when reality departs from the plan. AI has advanced quickly alongside that shift, and its fullest value emerges when it is tied directly to the decisions a business makes each day.

This third and final episode of our three-part series turned to agentic AI as the accelerator: not AI that predicts outcomes or writes content, but AI that reasons, orchestrates, collaborates, and helps drive decisions across the value chain. We looked at what sets it apart from the technologies that came before, how to match the level of agency to the decision at hand, and what it takes to scale with governance in place from the start. A live look at Aera, along with client examples, showed how the pieces come together in practice.

Key Takeaways

• Advantage now comes from adapting, not from planning perfectly. For years, a strong plan and steady execution were enough to run a supply chain well. The harder work now begins after the plan exists. Suppliers run late, equipment fails, and shipments slip, often at the same time. Decision intelligence is the connective tissue between that volatility and fast-moving AI capability, linking data, context, and execution in a continuous loop so teams know which action to take.

• The right level of agency matters more than maximum autonomy. AI covers a wide spectrum. It runs from rule-based automation through machine learning and generative models to agentic systems that plan, use tools, and adapt in pursuit of a goal. Agency develops along a maturity curve, and full autonomy is rarely the right target. The appropriate level depends on business impact, confidence, reversibility, regulation, and the cost of getting a decision wrong. People remain essential throughout, setting objectives, defining guardrails, and watching outcomes.

• Value amplifies when use cases connect end to end. One use case that gathers AI capability around a single decision makes a sensible starting point: scope stays manageable and value can be measured. The larger gains arrive when those use cases are strung together, so the output of one becomes the input to the next. Agentic AI serves as the orchestrator. It selects capabilities, sequences them, and holds context across systems without replacing the applications already in place. Results then feed back into earlier decisions, and the whole chain learns together.

• Discipline and reusable foundations are what carry a capability to scale. One life sciences program followed a platform-first principle: govern the foundation once, then reuse it many times. A structured intake weighed each candidate for value, feasibility, data readiness, risk, and reuse potential. A common blueprint embedded security, governance, auditability, and explainability before any solution was created. More than 14 use cases were prioritized on that basis, and downstream integration effort fell by close to 30 percent. Scale comes from that kind of discipline, not from more disconnected pilots.

• The operating model matures in stages, and roles evolve alongside it. In the AI assisted stage, work stays human owned while AI surfaces priority opportunities. In the AI augmented stage, the platform recommends next steps and supports handoffs, while people direct the workflow and remain accountable for the decisions that matter. In the autonomous stage, AI takes on more routine execution, and people monitor performance, handle exceptions, and shape policy. Effort shifts toward judgment and improvement, so roles, skills, and controls need to be redesigned alongside the technology rather than after it lands.

• Most organizations are converging on a hybrid development model. A platform-led approach relies on out-of-the-box capability, which simplifies the landscape, standardizes controls, and speeds time to value where processes are common. A fully custom build maximizes control and differentiation, though it calls for deep engineering capacity, heavy governance, and long-term upkeep. The hybrid model sits between the two, pairing a reusable platform foundation with company-specific agents for proprietary data and differentiated reasoning. More than 60 percent of organizations are now moving that way. The choice should follow time to value, differentiation, integration complexity, total cost, and organizational maturity.

• Structured and situational decisions run on the same foundation. A structured decision starts with a trigger, such as a stockout signal or a supplier delay, and the platform weighs options across prediction, optimization, heuristics, machine learning, and agents. A situational decision starts with a person asking a question in plain language, yet draws on the same engines and the same decision memory. Both are grounded in the Decision Data Model, write actions back to source systems, and are logged so the system learns over time. In the demonstration, a morning briefing showed 67 recommendations generated overnight and 45 executed autonomously.

• Trust is engineered into the architecture. Agentic Ambient Intelligence discovers before it acts. It scans the environment, reasons through the situation, and surfaces adjacent risks along the way. A live registry indexes every object at creation, so catalog scale and context window stay decoupled and accuracy holds as the environment grows. Orchestration is a structural consequence of the core engine rather than a probabilistic call, so routing stays deterministic and the model reasons within a structure that decides what runs. Governed views honor the same access controls a business user has, and every decision can be simulated, traced, and observed in real time.

Speakers

	Archana Ravi, Director, Growth and Solution Engineering, Aera Technology Archana brings 17 years of experience in supply chain strategy, operations, and product management across industry and consulting. Archana has partnered with leading food and beverage, retail, CPG, chemicals, and pharmaceutical companies to transform and optimize complex supply chains. She works closely with Aera customers to help them realize the full value of decision intelligence, define their North Star vision, and turn big aspirations into measurable results.
	Kevin Overdulve, Partner, Supply Chain & Network Operations Practice Leader, Deloitte Kevin leads the supply chain practice at Deloitte Belgium and drives the firm’s global decision intelligence agenda. He also serves as the Global Aera Technology Alliance lead, helping organizations transform supply chain decision-making through AI.
	Hitesh Parmar, Associate Director, Deloitte Hitesh advises clients at Deloitte UK, working at the intersection of digital transformation and AI-driven organizational change. He specializes in helping enterprises bridge the gap between rapid AI adoption and accountability, focusing on the transition from tool-based AI to agentic systems.

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 copilot or an AI tool on top of our data. How is this different?

A: A BI layer or a copilot produces an answer, and that answer lands in someone’s inbox and waits. Aera produces an action and then executes it. The difference lies in where it operates: at the decision layer rather than the language layer, sitting on top of the data platforms and LLM investments you already hold rather than replacing them. What you get is a complete loop, running from the trigger through the options, the recommendation, the action, and the feedback that follows. The useful question, then, is not whether your current AI gives good answers, but where those answers go once they arrive.

Q: If an agent reasons its way to the wrong action, what stops it?

A: Several safeguards, and they work together. Governance is enforced at the invocation boundary, in code, on every call, with permissions applied at the engine and object level at the moment a decision is made. The LLM never makes an access control decision. Every autonomous step carries two validation checkpoints, and a failed check sends structured feedback to the agent so it can correct that step rather than carry on. Business semantics and policy controls sit alongside the deterministic engines to keep decisions within the boundaries you have set. The architecture assumes the model will sometimes be wrong and contains the error rather than trusting the model not to make one. An action can never exceed the authority of the person who authorized it.

Decision governance also lets you place human checkpoints wherever the risk warrants them, and transparency does a great deal of work here. Seeing the logic, the reasoning, and the rationale behind a decision, with an explicit execute step in the early stages, gives teams a safeguard as they move toward fuller automation. Every recommendation is recorded along with whether it was executed and what resulted, which creates a learning history you can use to refine the logic over time. Decisions go wrong in every organization today, though they are rarely visible when they do. Bringing them into the open, where they can be governed and learned from, is a clear step forward.

Q: Does our data go to an external model, and are we locked into whichever model gets picked?

A: The LLM receives structured, governed output from data views, scoped to the query and to the permissions of the person asking. It sees only what that user is authorized to see. Aera is LLM neutral by design, with specific models assigned to specific tasks, and both per-tenant routing and bring-your-own-LLM are live today. On compliance, the platform is SOC 2 Type II and ISO 27001 certified, holds ISO 42001, is EU AI Act ready, and aligns to the NIST AI Risk Management Framework.

In regulated industries such as life sciences, organizations often prefer to keep everything inside their own infrastructure, and that approach works well. Others are comfortable with hosted models. Either way, nothing ties you to a single one. New model versions tend to be iterations of those before them, so moving to a newer version is straightforward, with cost the main consideration. What differentiates you as an organization is how you match use cases to the right model and the right application. The orchestration layer and the routing logic are where that advantage sits.

Q: How much time and effort does it take to set up and configure Aera?

A: Less than most people expect. Organizations typically get going within the first month. The Decision Data Model can be set up in around two weeks, and a first skill usually follows within one to two months. From there, the first milestone of measurable value tends to arrive by the three-month mark.

Q: What does this do to the time it takes to create new decision logic, and to the people who do that work?

A: Aera Intelligence goes live in days, supported decisions follow in one to two weeks, and full automation lands in eight to twelve weeks. Prompt-to-skill creation, where you describe the decision problem and Aera drafts the specification and the logic, ships this summer. The work itself moves up a level. Less time goes into writing logic, and more goes into defining the boundaries that logic runs inside and reviewing what comes out of it.

Q: What does this cost to run at our volume, and how do we keep it predictable?

A: A discovery-first architecture means only the relevant subset of your environment reaches the reasoning step, so token cost stays bounded per interaction. Continuous automation runs on deterministic engines, which is why the 50 million decisions executed in 2025 carried no LLM cost that scaled with volume. Continuous automation and on-demand reasoning also share a single platform, so there is no second set of infrastructure to run, monitor, and reconcile. For the LLM spend itself there are three commercial paths: Aera API calls, your own LLM contract, or Aera embedded models. You are never forced to carry model costs you already hold under contract.

Q: Do you need a platform-first approach to reach decision intelligence at speed? What if the business wants a unified capability, but money has already gone into several individual capabilities that are not connected?

A: There is no need to rip and replace what you have. Fragmented capabilities tend to struggle with decision intelligence because they lack a shared data foundation for decisions, and that foundation is precisely what the Decision Data Model™ provides. It harmonizes data from the systems you already own, so earlier investments continue to count toward the unified capability you are working toward.

Q: How do you retain past decisions in the system?

A: Every decision is captured with its full context: the data behind it, the recommendation, the action taken, and the outcome. Together these create a permanent, auditable decision history. That record is also what Aera learns from, so acceptance patterns and outcomes continuously improve the recommendations that follow.