Q&A

From AI Pilots to Platforms: Designing Enterprise-Grade Agentic AI

• By David Ramel
• 09/01/2026

Agentic AI is changing what enterprises are being asked to architect. Instead of building applications that simply send a prompt to a model and return a response, organizations are increasingly exploring systems that can make decisions about how to accomplish a goal, invoke tools, interact with business systems and coordinate multistep work.

Microsoft's Azure Architecture Center describes an AI agent as a system that can use a language model to determine what actions to take, which tools to call and how to sequence steps to complete a task. With multi-agent architectures, specialized agents can coordinate to tackle more complex work. But Microsoft also cautions that additional agents bring more coordination overhead, latency and potential failure modes, making architectural discipline increasingly important as systems grow.

That production concern is also visible in platforms and frameworks emerging around enterprise agents. Microsoft Foundry Agent Service, for example, includes capabilities around identity, role-based access control, observability, versioning, publishing and shared tooling, while the open source Microsoft Agent Framework is aimed at building, orchestrating and operating production-grade agents and multi-agent workflows. Together, those capabilities highlight a broader shift: once agents move beyond isolated experiments, concerns such as governance, reusable services, shared context, security and operational visibility become architectural issues rather than afterthoughts.

That's the challenge Samantha St-Louis will tackle in the upcoming From AI Pilots to Platforms: Designing Enterprise-Grade Agentic AI session at Live! 360 Tech Con 2026 in Orlando.

Samantha, AI Solutions Lead at Global Knowledge, is a Microsoft MVP, Microsoft Certified Trainer (MCT) and founder of CloudFirst AI. Her work focuses on helping organizations move beyond experimentation to create secure, scalable and production-ready AI solutions using Microsoft technologies, with expertise spanning Azure AI, Copilot, AI agents and enterprise AI adoption.

Her introductory-level session, scheduled for Nov. 18 from 9:30 a.m. to 10:45 a.m., focuses less on demonstrating a particular AI tool than on the architectural and organizational decisions needed to turn disconnected experiments into an enterprise platform. Among the issues on the agenda are coordinating multiple agents, preserving shared context, reducing duplicate infrastructure and keeping technical decisions aligned with long-term business outcomes.

For Samantha, one of the central problems is that organizations can mistake a successful proof of concept for evidence that an AI system is ready for production. A pilot may prove that an agent can perform a task, but an enterprise deployment also has to answer questions about security, ownership, integration, governance, evaluation, cost and lifecycle management.

Her argument for platform thinking doesn't mean every problem needs a sprawling multi-agent design. Rather, it means enterprises should decide what capabilities belong in individual solutions and what should be standardized underneath them. Shared identity patterns, data access, observability, orchestration, governance and common tools can keep teams from independently rebuilding the same capabilities while the number of agents continues to grow.

We caught up with Samantha to discuss why AI pilots so often fail to become enterprise platforms, the warning signs that indicate an organization is creating AI sprawl, and how architects can connect agentic systems to measurable business results.

PureAI: What Inspired You to Present on This Topic?
Samantha: There is a growing and expensive gap between AI experimentation and enterprise impact.

Organizations are running pilot after pilot, but a successful pilot is not the same thing as a production-grade AI solution. The architecture required to prove a concept quickly can be different from the architecture required to deliver measurable efficiency, capability or revenue impact at scale in a production-ready solution.

That is why I am increasingly skeptical of the traditional POC mindset in AI. Proving that an agent can work is often easy. Designing a secure, governed, integrated, reusable solution that can operate across the enterprise and generate real ROI is the harder and far more important problem.

This session is about closing that gap: moving from isolated pilots to intentional enterprise platforms designed for measurable business value.

What Is the Clearest Sign That an Organization Has Outgrown Individual AI Pilots and Needs a Platform Approach?
The clearest sign is when duplication and underutilization start to appear at the same time.

If multiple teams are independently building similar connectors, retrieval layers, governance processes or integrations, the organization is already creating unnecessary complexity. But an even stronger signal is when the number of agents keeps increasing while adoption, utilization and measurable business value remain low.

At that point, the issue is no longer a lack of experimentation. It is a lack of platform strategy.

A platform approach helps organizations determine what should be reusable, what should be standardized, which agents are actually creating value, and where investment should be consolidated rather than continuously adding more disconnected solutions.

Where Do Organizations Most Often Lose Control as Agentic AI Adoption Begins to Scale?
Usually at the boundaries between agents, data, tools and ownership.

It is relatively easy to govern one agent. It becomes much harder when many agents can access different systems, invoke tools, generate actions, hand work to other agents and operate with varying levels of autonomy.

Organizations also lose visibility surprisingly quickly. They may know an agent exists without knowing exactly what data it can access, which tools it can invoke, what other agents depend on it, how much it costs to operate or how its performance is being evaluated.

Scaling agentic AI therefore requires more than traditional AI governance. You need operational visibility into the entire agent ecosystem.

The best time to develop your agentic operational system is before you deploy the first agent. The second best time, is right after you've read this line!

What Should an Enterprise Standardize First to Prevent AI Sprawl?
Enterprises should standardize the operating model as early as the technical architecture.

That includes defining front-door versus back-door agents, identity and access patterns, shared data architecture and residency, and the common services agents will use. But organizations also need to establish lifecycle management: who can create an agent, who owns it, who monitors it, how it is evaluated and when it should be retired.

I often compare agents to employees. Before you hire someone, you define their role, their access, who manages them and how their performance will be measured. Agents should be treated with the same discipline.

Without those standards, organizations quickly accumulate agents with overlapping capabilities, unclear ownership, inconsistent access to data and no structured way to manage them over time.

How Can Multiple Agents Share Context Without Creating Unnecessary Duplication?
The key is separating shared enterprise context from agent-specific context.

Organizations often make the mistake of giving every agent its own copy of the same knowledge, retrieval logic, integrations and memory. That becomes expensive to maintain and creates consistency problems almost immediately.

Instead, agents should be able to reference governed shared services and common knowledge sources while maintaining only the context necessary for their specific task or interaction.

I think of it as creating a shared context and capability layer rather than giving every agent its own isolated universe.

That approach reduces duplication, improves consistency and makes changes much easier to manage.

What Architecture Decision Has the Greatest Influence on Whether an AI Initiative Can Scale?
The architecture decision with the greatest influence on whether an AI initiative can scale is whether you design the solution as an isolated application or as part of a reusable enterprise architecture.

A surprising amount of long-term scalability is determined very early. If every agent is tightly coupled to its own model, data source, integration logic, security model and orchestration layer, scaling becomes increasingly expensive.

A stronger architecture separates those concerns. Models can change. Agents can change. Business processes can change. But shared capabilities like identity, data access, observability, tools, orchestration patterns, governance and evaluation should not have to be rebuilt every time they do.

Equally important is how close the architecture is to the business user. Builders who design solutions without first deeply understanding the business logic, workflows, decision points and user needs often create technically impressive systems that nobody wants to use. Those solutions are fundamentally designed for technologists rather than for the people doing the work.

The strongest architectures therefore do two things well: they create reusable technical foundations, and they are designed with the business user, not simply for them. That combination is what allows an AI solution to scale both technically and operationally.

How Should Organizations Connect Agentic AI Architecture Decisions to Measurable Business Outcomes?
Start with the outcome before designing the agent.

Organizations sometimes begin with the technology -- "We want an agent" -- and then search for something useful for it to do. I prefer the reverse approach.

Identify the business process, determine where value is currently being lost and define how success will be measured. That might mean reducing processing time, increasing revenue, lowering operating cost, improving conversion, reducing errors, accelerating employee productivity or improving customer response times.

Then architecture decisions can be evaluated against those outcomes.

For example, greater autonomy may increase speed but introduce additional risk. A more sophisticated multi-agent architecture may improve performance but also increase latency and operating cost.

Good architecture is not the most technically impressive architecture. It is the architecture that delivers the required business outcome with the appropriate level of cost, reliability, governance and risk.

What Is One Warning Sign That a Company Is Moving Too Quickly from Experimentation Toward Production?
One of the clearest warning signs is when an organization is preparing to move an agent into production without a formal architecture and design document.

That documentation should define the agent's end-to-end architecture, business function, data flows, integrations, permissions, security controls, governance model, lifecycle, ownership, evaluation framework, feedback loops, maintenance requirements and failure-handling processes.

If those elements have not been documented and reviewed, the organization may have a working prototype, but it does not yet have a production-ready system. In fact, 99% of the time, once you push it to production the agent fails. Like throwing noodles at the wall and hoping it sticks!

Production should mean that the organization understands not only how the agent works today, but also how it will be governed, monitored, improved, maintained and eventually retired.

How Can Attendees Learn More About This Topic, and Prepare for Your Session?
I would encourage attendees to come to the session with one real AI initiative from their organization in mind.

Think about what would happen if that use case became extremely successful. What if one agent became ten? What if one department became twenty? What capabilities would suddenly need to be shared? Where would governance become difficult? What would you wish you had standardized earlier?

During From AI Pilots to Platforms: Designing Enterprise-Grade Agentic AI, I'll walk through the architectural patterns and strategic decisions organizations need to make as they move from experimentation toward enterprise-scale agentic AI.

My goal is for attendees to leave with a much clearer picture of what belongs in an individual AI solution, what belongs in the enterprise platform underneath it, and how to design today's AI initiatives so they do not become tomorrow's technical debt.

Note: Those wishing to attend the session can save money by registering early, according to the event's pricing page. "Save $400!" said the organizer, whose pricing page lists Sept. 25 as the Super Early Bird registration deadline for Live! 360 Tech Con 2026, which is presented by the parent company of Pure AI.