Key takeaways

• Start with measurable business value, then choose the technology that best fits each part of the workflow
• Build reusable agents and controls on an enterprise platform instead of creating isolated bots
• Combine machine learning, agentic AI, robotic process automation and human judgment in modular workflows
• Reengineer the process before automating it to create more value than point solutions alone

Organizations exploring agentic AI face a challenge that extends well beyond building an impressive pilot. They must identify business problems worth solving, determine where agents belong in a broader workflow and establish the governance needed to operate those agents at scale.

That was the focus of a ZS-hosted webinar featuring Pritam Debnath, senior director and global head of AI technology at Sysco, and Nitesh Ratre, associate principal of technology at ZS. They discussed Sysco’s approach to selecting agentic AI use cases, building reusable enterprise capabilities and combining AI with automation and human judgment.

They also examined a demand forecasting workflow that illustrates why successful AI transformation often requires several technologies working together rather than an agent handling every step.

Watch the webinar to get the full insights from the conversation.

Nitesh Ratre (NR): What pushed Sysco to move beyond AI experiments and toward enterprise-scale agents?

Pritam Debnath (PD): Merchandising, inventory management and the broader supply chain have faced significant volatility. Businesses are trying to maintain margins, improve profitability and reduce inventory and supply chain costs while continuing to meet service expectations.

At Sysco, we don’t want to simply place AI solutions on top of existing processes. We’re asking how we can use AI to transform and reimagine those processes.

The period of isolated copilots, one-off bots and proofs of concept is giving way to a new question: How quickly can we scale the right capabilities across the enterprise?

Moving from a proof of concept to enterprise scale is a substantial journey. It requires governance, security and auditability. It also requires repeatable delivery patterns and end-to-end agent life cycle management. We developed the Sysco Agentic Ecosystem, or SAGE, with those requirements in mind. The goal isn’t to create a great pilot. It’s to create a platform that can support scalable AI applications across the business.

NR: How do you decide which agentic AI use cases are worth pursuing?

PD: We try to avoid starting with a hammer and then looking for nails. Agentic AI is evolving quickly, and it’s easy to begin with the technology and ask where it might fit. We begin with the business problem instead.

The first step is to discover, identify and prioritize use cases based on business value. That value might come from improving profitability, reducing cost per unit or improving working capital. The use case needs to connect to a meaningful business outcome.

Once we understand the problem, we ask what solution will improve the workflow most effectively. We keep the business process and the people performing the work at the center of the design.

Sometimes the workflow needs an agent that can reason, plan and execute across several steps. In other cases, a deterministic task may be better suited to robotic process automation. We don’t want to force agentic AI into every phase of a process. We want to choose the right technology for each phase and create the best possible end-to-end workflow.

NR: What does the life cycle look like once you’ve selected a use case?

PD: It begins with discovery and design. We define the business objective, understand the workflow and determine which parts should change. From there, we integrate the necessary tools and assemble the agent or broader solution.

Once the solution moves toward production, evaluation becomes critical. That includes quality checks, testing and approvals. Teams need to understand whether the agent is producing the expected output and behaving consistently.

Deployment isn’t the end of the process. We continue to monitor performance and effectiveness. Feedback from the workflow helps us refine the solution over time.

Some of the concepts are similar to the software development life cycle [SDLC] and machine learning operations [MLOps], but AI can also help automate parts of those practices. The main principle remains the same throughout the life cycle—keep business value and the business process at the center. If an agentic workflow is the best answer, use one. If another approach is more effective, use that instead.

NR: How does SAGE support that approach?

PD: Think of an agentic workflow as a structure made from building blocks. SAGE provides the reusable blocks that individual teams and applications can assemble into workflows that solve specific business problems. As a horizontal enterprise platform, SAGE enables vertical agentic applications and workflows by offering shared agents, components and other artifacts that teams can reuse across functions rather than rebuilding the same capabilities for every use case.

That product and platform approach improves development speed, supportability and ease of implementation. It also gives teams access to components that already include enterprise guardrails.

NR: Can you walk through how that model applies to demand forecasting and forecast adjustments?

PD: Demand forecasting is a common challenge, but it becomes especially complex in a large supply chain. Sysco’s demand planners manage thousands of forecast exceptions each week across distribution centers, markets and regions.

Those exceptions may come from changes in demand or problems with the input data. Historically, managing them required considerable manual work. Planners also relied on their own knowledge and checks, which could create inconsistencies. As the number of exceptions increased, so did the risk of out-of-stocks and other supply chain disruptions.

When we approached this, we didn’t start by asking how to build an agent for the process. Instead, we broke down the main business problem into components.

The first component was the forecasting model itself. The existing tool used a more classic form of exponential smoothing. We developed an ensemble machine learning model using more sophisticated forecasting algorithms. In the implementation discussed during the webinar, that model produced better forecast accuracy than the base forecast for nearly two-thirds of the stock-keeping units.

The next question was what to do with the model’s recommendations. An agent built with SAGE evaluated the output against a confidence threshold. It could send high-confidence recommendations directly to the systems that execute forecast updates.

When a recommendation required human judgment, the agent sent it to a task and workflow management system for review. The overall solution combined the forecasting model, system integrations, robotic process automation and a human in the loop.

Together, those components improved forecast accuracy, reduced out-of-stocks associated with forecasting errors and reduced the manual effort required to manage exceptions.

The example is important because no single technology handles the entire problem. Agentic AI, machine learning, automation and human decision-making all play different roles in the same workflow.

NR: The workflow also includes a forecast insights agent. What does that add for planners?

PD: The insights help planners understand what may be causing a spike or exception before they make a decision.

Without that support, a planner may need to visit several enterprise resource planning systems, gather data from different places and perform additional analysis to make sense of the exception.

The forecast insights capability brings relevant information into the workflow. It gives the planner context for the recommendation and helps make the human review more focused. The goal isn’t to remove the planner from the process. It’s to give the planner better information and reduce the effort required to reach a decision.

NR: You’ve emphasized AI-first process transformation. Does that mean point automation is no longer worthwhile?

PD: It isn’t an either-or choice. Point automation can deliver real business value, and it may be the right place to begin. Organizations often follow a crawl, walk, run progression as their capabilities mature.

If you add automation to an existing process, you can improve that process. That’s useful, and teams should pursue those opportunities when the benefit is clear. They do deliver business benefits, but that’s the difference between evolutionary versus revolutionary.

But the larger question is whether the existing process is already designed in the best possible way. That’s why we start with process engineering.

Suppose a legacy process has five steps. A team may be able to automate two of them and create a valid benefit. But as the organization matures, it should also ask whether all five steps are necessary. Could the process be redesigned with three steps and then supported by AI?

When you redesign the process before applying the technology, the potential benefit can be much larger. Point automation improves the work as it exists. Process transformation creates an opportunity to change how the work happens.

NR (to the moderator): Amanda, do we have any questions from the audience?

Amanda: We do have some. Where does the SAGE platform end and an individual agent implementation begin?

PD: SAGE provides the horizontal, reusable components. An individual business application selects the components it needs and assembles them into an agentic workflow.

Returning to the building-block comparison, SAGE supplies the blocks. A vertical application combines those blocks into a structure that solves its particular business problem.

The platform may provide reusable agents, governance capabilities or common technical services. The application team brings those pieces together with the business logic, tools and workflow required for the use case.

That separation allows teams to develop solutions for different functions while sharing common enterprise capabilities.

Amanda: Here's another. How do quality, compliance and cost management change as agents scale?

NR: Traceability is essential. When an application uses several agents, an orchestrator may decide which agent should handle a request and how the workflow should proceed.

The system needs to record those decisions and the steps that follow. Teams can then audit whether the workflow selected the right agent, performed the right actions and produced a useful response.

User feedback is another important part of the process. When a response isn’t useful, the team needs a way to understand what happened, adjust the solution and improve future performance. Without that feedback and auditability, it’s difficult to establish the level of quality needed for scale.

Human review also remains part of the design. As teams gather more data about performance, they can refine the agents and determine where more automation is appropriate.

PD: Cost management belongs in that same discussion. As software pricing shifts from license- or seat-based models toward token-based consumption, financial operations, or FinOps, becomes an important part of scaling agents.

A platform can help select an appropriate model based on the task instead of using the most expensive model for every request. The organization also needs to monitor how agents communicate with one another and how much consumption a workflow generates.

Quality controls, governance and cost management can’t be added after an agent has already scaled. They need to be part of the platform and production design from the beginning.

NR: Thank you, Amanda. One last question, Pritam: What should leaders carry forward as they move from pilots to enterprise deployment?

PD: Keep business value and the business process at the center. Don’t begin with the assumption that every problem needs agentic AI.

Build for production rather than treating governance, monitoring and auditability as later additions. Use reusable platform capabilities where they can accelerate development and give teams consistent controls.

At the same time, don’t dismiss smaller automation opportunities. They can create value and help an organization build maturity. But continue to ask whether you’re improving an existing process or redesigning the process itself.

The goal isn’t to deploy the greatest possible number of agents. It’s to build better workflows that improve business performance and help people make better decisions.

Watch the webinar to get the full insights from the conversation.