I spent part of this weekend trying to understand what Sysco is actually doing with artificial intelligence. I was less interested in the headlines, the press releases, or even how much money the company hopes AI will save. I wanted to understand what Sysco is building underneath all of it.

That led me to a presentation by Pasindu Liyanage of Sysco LABS titled Engineering Excellence in the AI Frontier. One slide in that presentation changed how I looked at Sysco’s AI strategy. The slide was titled “From Chaos to Coherence: Orchestration & Agentic Building Blocks.”

That title may tell us more about the future of enterprise AI than another announcement about a new chatbot ever could. Sysco does not appear to be thinking about AI simply as another application. It appears to be building something much more important: an intelligence layer for the enterprise.

The Problem With Building AI One Application at a Time

Most companies are beginning their AI journey in a predictable way. Marketing experiments with one AI tool, finance experiments with another, IT builds a chatbot, and operations finds an automation platform. Somewhere else, someone connects a large language model to company documents while another team begins experimenting with an AI agent.

Individually, many of these projects can create value. The problem begins when a company tries to move from a handful of experiments to dozens or even hundreds of production AI systems. Every AI application begins needing many of the same underlying capabilities.

It needs access to models. It may need access to company knowledge, authentication, documents, databases, APIs, or business systems. Once AI begins taking actions instead of merely answering questions, the company also needs ways to control, monitor, test, and govern what those systems are doing.

If every department builds these capabilities independently, AI can create a new generation of technology silos. The company may solve individual problems while creating an increasingly fragmented technical environment underneath them.

That is very close to the problem Pasindu presented. His architecture slide lists challenges including fragmented development efforts, vendor lock-in, lack of interoperability, lack of centralized governance, and difficulty integrating AI into the broader ecosystem.

Those are not really model problems. They are enterprise architecture problems, and that distinction matters.

From Chaos to Coherence

Sysco’s answer appears to be the centralization of the common AI foundation. Pasindu’s diagram places a Centralized Enterprise AI Platform at the center of the architecture. Above it are reusable GenAI agents, GenAI applications, configuration tools, and what the slide calls a Northbound Application Layer. Below it sit model providers and knowledge stores.

That changes the way a company can build AI. Instead of every application constructing its own intelligence stack, applications can use shared capabilities. Instead of every development team deciding how to connect to models, data, knowledge, and tools, the enterprise can provide a common foundation.

This also means that individual applications do not have to carry the entire burden of intelligence themselves. The intelligence begins to separate from the application.

I think that may be the bigger idea.

For decades, software applications have contained both the business process and much of the logic required to operate that process. AI creates the possibility of moving some of that intelligence into a reusable layer that can work across many applications.

The Model Does Not Have to Be the Architecture

Another interesting detail in Pasindu’s diagram is the number of technologies represented underneath the platform. The slide includes technologies and services associated with Gemini, AWS Bedrock, Azure OpenAI, Llama, Vertex AI Search, Amazon Kendra, and Chroma, among others.

That suggests another important architectural principle: the enterprise does not necessarily need to organize itself around one AI model.

Models will change. Prices will change. Capabilities will change. Some models will be better at reasoning, others may be better at speed, cost, coding, document analysis, or specialized tasks.

A company that directly connects every business application to one model provider can eventually create another form of lock-in. A common intelligence layer provides a different possibility. The business application can ask for intelligence while the platform determines how that intelligence should be delivered.

That is a much more durable architecture.

The model becomes a component of the system rather than the system itself.

Then Come the Agents

This is where Sysco’s architecture becomes especially interesting. Pasindu’s diagram explicitly includes Reusable Gen AI Agents as part of the centralized platform architecture.

That means the reusable building blocks are not limited to model access. The enterprise can begin creating specialized intelligent capabilities that can be used across multiple applications and business processes.

One agent might work with documents. Another might interact with structured data. Another could call an API, evaluate information, or perform a specialized business task. Those agents can then become components of larger workflows.

This is very different from thinking about an AI agent as one all-knowing digital employee that somehow understands every part of the company.

The emerging architecture looks more like an organization itself. Specialists perform specialized work. Orchestration determines which specialists are needed. Applications provide business context. Company systems provide information. Governance determines what each component is allowed to see and do.

That is a far more realistic way to think about enterprise agents.

Applications Remain Specialized. Intelligence Becomes Ubiquitous.

This is the part of Sysco’s architecture that I think matters most for the rest of us.

For decades, businesses have purchased specialized applications. ERP handles financial transactions and operations. CRM manages customers and opportunities. Warehouse systems manage inventory. HR platforms manage employees. Document systems manage policies, procedures, contracts, and knowledge.

I do not think those applications are going away, and I do not think they need to.

What may change is the intelligence surrounding them.

Instead of trying to put the entire company into one giant AI application, a common intelligence layer can potentially reach across specialized systems when it has the authority to do so. The applications continue doing what they were designed to do, but intelligent systems can retrieve information and perform authorized actions across those boundaries.

That leads to a principle I keep coming back to:

Applications remain specialized. Intelligence becomes ubiquitous.

The intelligence does not have to live inside every application. It needs governed access to the applications, data, and knowledge required to accomplish the task.

That creates a very different kind of enterprise.

What Does This Mean for a Mid-Market Company?

Sysco operates at enormous scale. A mid-market manufacturer is not going to recreate Sysco’s engineering organization, and it should not try.

But the architectural principle may scale down much better than the infrastructure itself.

Imagine a company running Microsoft Dynamics 365 Business Central. Business Central remains the operational system of record. Customers, vendors, inventory, orders, and financial transactions stay where they belong.

SharePoint might hold governed company knowledge such as policies, procedures, specifications, contracts, and other important documents. Other systems might continue handling CRM, production, logistics, or specialized business processes.

Then an intelligence layer begins operating across those systems.

One agent might identify revenue opportunities. Another could investigate a customer credit problem. Another might prepare a distributor price-change submission, monitor the health of a business process, reconcile channel data, or gather the information required for a management decision.

The applications have not disappeared. The ERP has not disappeared. The company’s knowledge systems have not disappeared.

What changed is that intelligence can now operate across those boundaries.

That may be the architecture that allows an established company to become intelligent without replacing every application it already owns.

The Hard Part May Not Be Building the Agent

This also changes the question executives should be asking.

Much of today’s AI conversation focuses on how quickly someone can build an agent. That is becoming less interesting because building individual agents is getting easier.

The harder challenge may be building a company that can safely operate hundreds of intelligent systems.

Who gives an agent authority? Which data can it access? Which actions is it allowed to take? When does it need human approval, and when can it act by itself?

How is its performance evaluated? How does the company know when it fails? Who is responsible for maintaining it? How do leaders prevent ten departments from building ten different versions of the same capability?

Those are not simply AI questions. They are questions about governance, operating models, architecture, responsibility, and management.

That may be why centralized platforms such as the one Pasindu showed matter so much. The platform is not merely a convenient place to connect models. It can become the place where the enterprise creates consistency around intelligence.

Sysco May Be Showing Us the Next Stage

There is an important caution here. One presentation does not tell us everything about Sysco’s technology environment, and I would not assume that one diagram represents every production system inside the company.

But the presentation does give us a useful window into how people inside Sysco LABS are thinking about enterprise AI architecture.

And the direction is worth studying.

Sysco does not appear to be asking only, “Where can we add AI?”

The architecture suggests a much larger question:

What foundation do we need if intelligence is going to become part of the entire company?

I think many established businesses will eventually have to answer that question.

The Internet did not simply give companies websites. Over time, it changed the architecture of the enterprise. It changed how companies communicated, sold products, connected systems, interacted with customers, and moved information.

AI may create a similar shift.

We may be moving from companies that use AI applications to companies with an intelligence layer running across the organization.

Sysco is building at a scale most companies will never need. But the architectural lesson may be much more universal.

The question for the rest of us is not whether we can build Sysco’s platform.

It is:

What is the smallest version of this architecture that an established company can build today?

I think that is where things are about to get very interesting.

The Autonomous Economy Is Coming – John Wheeler

MCP Isn’t the Revolution. It’s the Infrastructure. – John Wheeler

Sources:
(7) Post | LinkedIn

Engineering Excellence in the AI Frontier | Pasindu Liyanage