Summary

In this article, TechLeader delivers a practical framework for enterprise AI agent strategy, based on direct implementation insights from John Thompson, former Global Head of AI at EY. Drawing from one of the world’s largest private GenAI environments, Thompson outlines seven core design principles that move agent systems from pilot to production. Topics include orchestration architecture, autoscaling logic, embedded governance, polymorphic workflows, human-agent collaboration, and tooling resilience. For CTOs, Chief AI Officers, and enterprise architects, this is a field-tested blueprint for deploying intelligent agents that deliver measurable outcomes at scale.

‍

‍

‍

The past year has seen a wave of agent demos, prototypes, and vendor frameworks. But for enterprises that need results and where decisions impact compliance, revenue, or global execution, most of those tools fall short.

EY has taken a different approach. The firm's former Global Head of AI, John Thompson reveals how the firm runs one of the world’s largest private generative AI environments where over 300,000 employees interact with agent systems designed for decision-making at scale.

In Thompson’s view, the real work of enterprise AI agent strategy begins after the pilot. “We’ve been building simple agents for over a year now,” he said. “We’re moving into intelligent and polymorphic agents because that’s what it takes to support business complexity.”

TechLeader unpacks exclusive insights shared by Thompson to offer a tactical framework for CTOs, Chief AI Officers, and enterprise architects ready to go beyond surface-level tooling.

The 7 Design Principles to Scale and Sustain your Enterprise AI Agent Strategy

Thompson’s insights can be broken down into 7 design principles, which are key to ensure that AI agents can be scaled across enterprises.

EY’s early deployments of AI agents focused on narrow tasks with clear boundaries. As those systems matured, the design requirements became more complex. Agents needed to coordinate across domains, operate under policy, and scale under load.  

These seven principles emerged from that progression. Each one reflects how EY moved from working examples to systems that are now part of day-to-day operations.

‍

‍

Design Principle 1: Filter for Quality Before You Train

One of EY’s first major agent deployments focused on proposal generation. The team used thousands of documents to train the system, which included winning and losing bids, polished and outdated formats.

The result was uneven. “It didn’t work very well,” Thompson said. “We put all sorts of proposals in there. Good ones, bad ones. And we realized we needed to go back.”

The fix came from curation. EY ran a human-in-the-loop evaluation cycle, filtered for only the best material, and retrained the model. The performance shifted. The agent which is now known as Deal and Delivery Assist, replaced a process that used to require multiple contributors and several weeks. Today, it returns a first draft in under five minutes.

Design Principle 2: Orchestration Beats Overgeneralization

Enterprise requests are rarely simple. They may combine tax rules, regulatory obligations, geographic constraints, and cost structure. Thompson’s team addresses this not by expanding the size of individual models but by layering them.

“You can have an agent that’s actually acting like an orchestrator for a variety of other agents behind it,” he explained. The orchestrator receives a prompt, segments it, and sends each part to a model optimized for that specific domain.

This architecture reduces system fragility. General-purpose agents risk underperforming across every domain. An orchestrated structure improves precision, traceability, and reuse.

Design Principle 3: Scaling Needs Rules, Not Just Runtime Logic

When deadlines shorten, agents may need to complete work faster than originally planned. EY’s systems include autoscaling logic that allows agents to clone themselves and run tasks in parallel.

“If I project it’s going to take me 24 hours to do this process, and I need to get it done in six, I’ll clone myself five times,” Thompson said. Each clone handles a segment of the work and the result is returned faster.

But speed comes at a cost. “Your hosting costs and your compute are going to be 10, 15, even 20 times larger,” he warned that without budget controls, autoscaling becomes a liability.

Design Principle 4: Governance Must Be Embedded at Launch

EY’s agent systems are governed before they go live. They are trained using verified materials, deployed with policy enforcement and operate within hard boundaries.

“An agent needs to be trained like a person,” Thompson said. “You need to give it all the training material you would give someone in that role. And you need to govern it with all the policies that make sense.”

This includes ethics rules, data access controls, geographic restrictions, and defined outcomes. If an agent can spend money or issue a recommendation, its permissions and audit trail must be clear.

Design Principle 5: Let Agents Extend rather than Invent

Some of EY’s most advanced systems operate polymorphically. These agents start with partial processes and finish them by identifying gaps, pulling in reusable modules, or creating new steps based on the objective.

“You started with the three human steps that were built,” Thompson explained. “Then the agent adds ten more, pulls from libraries, and runs the whole thing.”

This allows teams to deploy faster. They don’t need to build end-to-end workflows from scratch. The agent expands the chain based on what the result requires.

Importantly, it does not do so blindly. Each step is logged, sourced, and evaluated against the original goal.

Design Principle 6: Agents Must Know Who to Ask

Many workflows include a moment when external input is required. This could involve an expert opinion, a compliance check, or feedback from another department.

Thompson’s view is that agents should initiate those interactions. “They’re going to have to have a catalog of people,” he said. “With their experience, their expertise, and what it might cost to ask them a question.”

The logic here isn’t complicated. If the agent needs input, it should reach out through a structured system, instead of guessing or halting. For internal requests, the expectation is response. For external ones, pricing may apply.

Design Principle 7: Expect Your Stack to Shift

EY does not rely on a single vendor framework. Thompson names partnerships with Microsoft, Google, Adobe, and IBM, but makes clear that modularity is essential.

“You may pick a small company,” he said. “They may get acquired, go out of business, or change their model.”

The systems EY builds are designed to keep working when that happens. Orchestration layers separate logic from the stack. Internal APIs, model wrappers, and agent registries allow flexibility without disruption.

Deployment Timeline: What Thompson Sees Coming for AI Agents by 2026

“We’re going to see polymorphic agents, agents building agents, by the end of 2026,” Thompson said. Based on internal roadmap velocity, he projects the following timeline for agent evolution at enterprise scale.

‍

Deployment Timeline by TechLeader

Enterprise Playbook Summary

For teams designing enterprise AI agents beyond the pilot, Thompson’s rules offer a clear execution model:

  • Use high-signal, curated training data. Quantity will not save you.
  • Route tasks to specialized models. Don’t expect a single system to carry the load.
  • Monitor autoscaling. Fast agents can create invisible infrastructure debt.
  • Write governance into the system before launch. Treat agents as accountable actors.
  • Allow workflow extension, but lock to objective-based evaluation.
  • Include a human registry. Expert input should be traceable and priced.
  • Plan for framework churn. System logic must be independent of vendor survival.

These aren’t optional optimizations. They are operating constraints shaped by scale.

Conclusion

Thompson describes an implementation reality. His teams are building systems that handle regulated work, execute decisions under pressure, and adapt without slipping out of control.

For enterprise leaders, agent strategy is a matter of what holds up when policy, performance, and permission collide.

And EY's work shows what that looks like when it’s live.

At TechLeader, we understand the stakes enterprise leaders face when navigating AI strategy. That’s why we’ve built a focused set of resources tailored to decision-makers, engineers, and innovators working in generative and enterprise AI.

  • TechLeader Voices: Interviews with leading experts who break down everything from the complexities of being a tech leader to navigating the ever-evolving AI landscape.

Check out the latest issue of TechLeader Voices to see how cybersecurity leaders are approaching agentic and explainable AI to build systems that are resilient, auditable, and ready for scale.

TechLeader Voices is built for those leading the shift. Subscribe to our free newsletter to stay sharp on the moves that matter in enterprise tech.