Inside EY’s Enterprise Playbook of Engineering AI at Scale

AI is dominating conversations across every industry, reshaping how enterprises innovate, operate, and compete in this new era of intelligence. But beyond the hype, what does it really take to build scalable, secure, and enterprise-ready AI systems?

To unpack these challenges and opportunities, we brought together several of EY’s leading distinguished technologists and global engineering leaders for an exclusive multi-speaker conversation on the future of enterprise AI. Their insights span across Retrieval-Augmented Generation (RAG) systems, AI governance, cybersecurity, autonomous agents, digital engineering, intelligent commerce, and the evolving role of humans judgement in an AI-first world.

Together, they offer valuable perspectives on how organizations can responsibly scale AI while remaining agile in a rapidly evolving technological landscape.

Industrializing AI through modern engineering, resilient architecture, and AI-native delivery

Pat Sullivan, EY Global Digital Engineering Lead and Distinguished Technologist Program Leader

As digital engineering becomes central to enterprise transformation, how is the leadership role evolving in an AI-first world? Especially with GenAI reshaping how software is designed and built?

Sullivan: I’d say the role is expanding from “chief technologist” to “chief value engineer.” Digital engineering is no longer just how we build; it’s how we compete. In an AI‑first world, especially with GenAI, the mandate is to design operating models, platforms, and guardrails that let every team safely harness AI to deliver outcomes faster, with higher quality and lower total cost.

Here’s what is changing rapidly, and my advice as to how leaders must drive change to deliver the value:

• Change the mindset from projects to products, with AI as a first-class teammate
• Move from the traditional software development life cycle (SDLC) to a product development life cycle (PDLC) & agentic development life cycle (ADLC)
• Lead with data and governance as the core engineering concerns
• Evolve talent, teaming, and culture through intentional AI upskilling 

What doesn’t change is accountability. AI raises the bar for speed, but it doesn’t absolve us of engineering rigor. Our job is to turn AI’s raw capability into trustworthy, repeatable, and economically smart outcomes that solve business challenges. That means building a platform that makes the right thing the easy thing, equipping teams with the skills and guardrails, and measuring value end to end with results directly tied to business performance.

The companies that do this well will not be stuck in AI technology, but will lead in the business of AI.

What does a modern digital engineering stack look like today? And how should organizations rethink architecture to support speed, scalability, and continuous innovation?

Sullivan: The modern digital engineering stack is cloud-native at its core – built on microservices, event-driven architectures, and API-first design principles that allow teams to move independently and scale selectively.

But the real shift isn’t just technological; it’s organizational. You have to decompose your monoliths – not just in code, but in governance. That means federated ownership of services, platform engineering teams that build internal developer portals, and a golden path that makes the right way the easy way.

AI is now a first-class citizen in this stack, sitting alongside observability, security, and CI/CD as a foundational capability – not an add-on.

Many enterprises struggle to move from AI pilots to production. What are the key engineering disciplines required to industrialize AI at scale?

Sullivan: The graveyard of AI pilots is full of great models. One that never made it to production, and the reason is almost always the same – organizations invest in the science but underinvest in the engineering.

Industrializing AI requires mature AI & MLOps practices. This includes automated retraining pipelines, model versioning, and robust feature stores that provide consistent data for both training and inference. Beyond that, you need a culture of measurement. Every model in production should have clear KPIs, fallback logic, and a human-on-the-loop (not in-the-loop) strategy for edge cases.

At scale, AI is a software engineering problem as much as it is a data science problem. And the companies winning are the ones treating it that way.

With increasing pressure to deliver faster, how can engineering teams balance velocity with resilience, security, and long-term maintainability?

Sullivan: Speed and stability are not opposites; they’re actually reinforcing when you build the right foundations.

The fastest teams I’ve seen are fast because they’ve invested in automated testing, trunk-based development, and chaos engineering that builds confidence in their systems. Security has to shift left and be embedded in the developer workflow. Rather than being bolted on at the end through a compliance gate. The bigger risk I see for enterprises is treating technical debt as a tax on the future rather than a liability on the balance sheet today. 

Leaders who fund debt remediation with the same rigor as new feature development are the ones who sustain velocity over time with the financial rigor to sustain future innovation.

Engineering AI at scale through trusted RAG architecture

Dipanjan Sengupta, EY Distinguished Technologist, EY Consulting Global Delivery Services – AI Engineering Leader

Retrieval-augmented generation is becoming foundational. According to you, what design principles separate experimental RAG systems from production-grade, enterprise-ready architectures?

Sengupta: Currently, many businesses view RAG in isolation. All are coming up with different point solutions for different use cases. This works experimentally. But in my view, it is the wrong approach for enterprises, because it makes scaling unattainable.

To make RAG production-grade and enterprise-ready, enterprises need a platform-based approach that adheres to basic security and compliance principles. This ultimately enables consistency, reliability, and developer productivity.

Production RAG systems externalize orchestration, tuning, and decision logic into configuration while automatically surfacing evidence, references, and confidence signals, ensuring consistent behavior, explainability, and easier validation across use cases.

My current research is focused on new capabilities for RAG that have so far remained elusive but would prove immensely important for enterprises. RAG does not have a very mature model for handling multimodal content, including content beyond text, such as illustrations, pictures, or graphs.

The successful addition of this functionality, along with agentic AI, will help transition RAG from simply delivering reference material to functioning as a dynamic, bona fide research assistant.

As AI systems grow more complex, how should organizations rethink architecture? Are we moving toward platform-based AI ecosystems rather than model-centric design?

Sengupta: We are absolutely moving toward platform-based AI ecosystems, and it’s a good thing! As I mentioned earlier, a platform-based approach is the most practical, safe, and efficient option.

Contrary to what vendors want us to believe, it is impossible for a single model to produce the best solution for all tasks. As an architect, I would prefer to choose the implementation of a specific task as I see fit, depending on its severity, accuracy, and cost. 

Everyone in the platform-based AI ecosystem benefits from using a single platform hardened with compliance and security measures. If not, every business unit in an enterprise is effectively an island, with its own compliance and security considerations and processes that must be maintained. It creates excessive fragmentation.

Platform ecosystems better support collaboration across multidisciplinary teams. The platform model explicitly separates concerns for data engineers, model developers, prompt engineers, and application teams through shared services and interfaces. This separation allows each role to move faster without stepping on others. Something that model‑centric designs struggle to support as organizational scale increases.

When advising our clients at EY, we recommend a platform-based approach. One grounded in developer productivity, offering consistency, reliability, and cross-team collaboration. While models will always evolve, platforms are made to absorb the change without constantly re-architecting.

Governed retrieval is critical for enterprise trust. How do you balance data accessibility with compliance, auditability, and IP protection in RAG pipelines?

Sengupta: Companies often fail to conceive RAG holistically as an end-to-end responsibility. And, as a result, they treat governance as an add-on rather than embed it into the infrastructure.

From my view, even a wrong citation is a hallucination; however, we shouldn’t put the burden on RAG. Instead, we need to stitch together the right services properly. This means that governed retrieval must be intrinsic.

Companies should adopt a platform-centric approach to unite compliance and auditability concerns. All while embracing a configuration-driven security model for RAG that integrates frameworks such as LLM Guard, MS Presidio, LangKit, and enterprise data protection solutions.

Enterprises should define clear trust boundaries using virtual enclaves. These provide logical isolation for data, policies, and controls across projects, domains, or business units, allowing organizations to enforce the right level of access and protection without forcing a single restrictive policy across all RAG use cases, balancing accessibility with control at scale.

This enables flexibility to plug in any type of security guardrails to enforce pre-ingestion masking, secure retrieval controls, vector database safeguards, and output filtering, ensuring that both the retrieval and generation steps prevent unauthorized access and data leakage. Collectively, this approach mitigates data leakage and other emerging RAG-specific vulnerabilities.

What are the most common failure points you see when organizations attempt to scale AI systems? And how can they build for reliability from day one?

Sengupta: The first failure point is using a model-centric approach. The number of constraints placed on organizations using GenAI without a platform-based approach will prove too many to deploy at scale. This is especially true now that most enterprises are diving into agentic AI deployment.

The research highlights that many failures occur because systems retrieve only localized fragments, missing higher‑order associations across documents. Incorporating community‑level knowledge enables more stable and context‑rich responses for global or ambiguous queries, reducing failure rates as scale increases.

We are also putting too much pressure on LLMs. Whereas many of our concerns around planning, orchestration, and validation should be engineered into the system.

Enterprise data is complex. And systems built for clean text will fail in real-world environments, which is why RAG is so important.

RAG will help with potential issues around scaling up newer autonomous AI agents. By acting as their reliable knowledge access layer, providing verified context, up-to-date information, and guardrail-aligned evidence, AI agents can plan, decide, and act with far greater accuracy, safety, and situational awareness.

While, of course, the challenge of integrating existing systems and services with the larger GenAI landscape is a perennial obstacle, we can best address this by having a standard framework to integrate with the GenAI platform leveraging industry-standard specifications (OAS, Async API, etc.).

Confronting AI-driven cyber threats in the age of autonomous attacks

Maez de Guzman, EY Distinguished Technologist, EY Global Cybersecurity Managed Services Emerging Markets Leader and EY Consulting Global Delivery Services Cybersecurity Leader

How is AI reshaping the cyber threat landscape, particularly in the rise of identity-based and LLM-powered attacks?

Guzman: AI is reshaping the cyber threat landscape by enabling scalable, highly sophisticated identity-based attacks. Synthetic identities now merge real and fabricated data to bypass traditional detection. While LLMs automate social engineering and multi-vector campaigns at scale.

The industrialization of identity attacks is a critical shift. LLM-driven autonomous agents are emerging, capable of probing, adapting, and exfiltrating data without human oversight. For example. This year we have seen a lot of deepfake voice synthesis that clones executive voices from seconds of audio. This creates a repeatable fraud vector that undermines foundational authentication models.

Now we are also seeing that the organizations that adapt fastest are those that stop asking, “Does this look suspicious?” Instead, they start asking, “Can we verify if this is real?”

Zero trust has been a key framework for years. How must it evolve to address AI-driven threats like synthetic identities and automated fraud?

Guzman: Zero trust must evolve beyond static policy enforcement to incorporate adaptive, AI-driven risk analytics. This means continuously validating identity and behavior in real-time, integrating signals from AI models that detect anomalies indicative of synthetic identities or automated fraud. The framework should embrace dynamic trust scoring that adapts to evolving threat patterns rather than relying solely on predefined rules.

Synthetic identities are now fabricated at an industrial scale. Automated fraud nowadays mimics legitimate user behaviour well enough to pass static risk checks. Identity proofing must now become AI-resistant by design. Session-level trust must also move from binary to continuous with an explicit anti-automation axis, forcing gating events when the risk engine detects automation patterns. Overall, the cadence and depth of trust evaluation also need to evolve.

Supply chain attacks are becoming more sophisticated. As a leader, where do you see the biggest blind spots in enterprise cybersecurity today?

Guzman: Supply chain attacks reveal critical blind spots in third-party risk management and software integrity verification. The complexity and opacity of modern supply chains enable adversaries to inject malicious code or tamper with updates. Many enterprises lack full visibility into supplier security postures and often miss enforcing end-to-end cryptographic validation of software components.

While software supply chain security has advanced through software bill of materials (SBOMs), dependency scanning, and provenance tracking, the model supply chain remains largely unaddressed. Organizations are rapidly integrating pre-trained AI models into production pipelines under market and efficiency pressures, often without rigorous scrutiny of model provenance, security, or integrity.

With the growing complexity of AI-enabled threats, how should organizations rethink cybersecurity education and workforce readiness?

Guzman: The term literacy has evolved beyond traditional reading and writing. It now centers on the ability to adapt to new technologies and grasp emerging concepts. AI literacy, in particular, should be measured by how effectively individuals understand, interact with, and leverage AI-driven tools and threats, rather than just theoretical knowledge. Organizations should focus on cultivating this adaptive mindset, ensuring our workforce can navigate and counter AI-enabled cyber risks with agility and insight.

At EY, we embrace continuous learning and are focused on supporting our people with the training they need to upskill today. 90% of EY people have undergone foundational AI training. And over 161,000 advanced AI learning courses have been completed or initiated. Beyond upskilling our people, we’re focused on embedding ethical considerations into AI to build trust and reduce risk. To support this, we’ve established EY AI principles covering accountability, security and privacy, transparency, fairness and inclusivity, and professional responsibility.

Building Smarter Enterprise AI with Human Oversight

Sridhar V Dasaratha, EY Distinguished Technologist, EY Global Client Technology AI Research Leader

As enterprises adopt GenAI, where do you see the strongest use cases for SLMs over LLMs. Particularly in regulated industries like finance?

Dasaratha: Cost matters for enterprises, and large‑scale LLMs aren’t always the most efficient choice. Many use cases involve repeating focused tasks on similar data, where smaller models can perform just as well.

For example. Solving a specific task in finance may simply require a smaller fine‑tuned model, rather than a large general model with knowledge across many domains. As a bonus, these smaller models can also run on‑premise if needed, reducing data movement and easing privacy concerns.

Financial AI agents are gaining traction. According to you, what governance and risk considerations must be addressed before they can operate autonomously?

Dasaratha: As we increase the deployment of autonomous agents, the biggest concern is reliability. Agents are often coordinating complex workflows across multiple systems – invoking tools, retrieving data, and interacting with other agents as part of a broader orchestration. In this kind of scenario, the reliability must be high for it to be something that enterprises can trust.

To enable safe autonomy in real world settings, strong controls need to be built into the system. Agents must operate within defined boundaries. They should know when to act, when to refuse, and when to escalate – with human oversight and human review built in for approval and high‑risk decisions. At the same time, there is increasing pressure to reduce human intervention as agentic AI systems take on more tasks, which makes maintaining this balance critical.

In addition, robust evaluation is needed to continuously monitor how these systems perform in real‑world conditions and detect any degradation or failure over time. These are areas that need to be significantly advanced from where they are today for autonomous systems to be successful.

Document intelligence has evolved rapidly. As an expert, how do you think GenAI is changing the way organizations extract, interpret, and act on unstructured data?

Dasaratha: Document intelligence spans the entire spectrum of natural language processing tasks related to business documents. The main differentiator now is that we’ve shifted from the need to create complex pipelines to extract information and train task-specific models to prompt-based systems, where we can describe the task more flexibly without retraining. This also simplifies how information is handled end-to-end, making it easier to convert unstructured data into structured outputs that can be used directly in downstream workflows.

With the evolution of GenAI, it is now possible for individuals to use an LLM to work through long, complex documents. And not just extract information, but interpret it in context.

For example. One can analyze a 100-page document against another document, such as a policy or guideline, to assess compliance or extract relevant insights. This applies to the full range of documents workers deal with every day and increasingly works across modalities in a unified way.

What skills and mindset shifts are most critical for the next generation of AI talent entering the workforce? 

Dasaratha: The future workforce will increasingly involve both humans and AI working together. In this kind of environment, it’s critical that people remain actively involved. The next generation of talent will need to effectively collaborate with AI – using these systems to augment how they process information, while also relying on unique human strengths such as judgment, intuition, and imagination.

It is important that humans do not become complacent or overconfident, leading us to over‑rely on AI. We cannot relax our oversight. AI systems need to be rigorously and continuously checked, with human judgment acting as a safeguard for the overall system.

As we delegate more tasks to AI, maintaining human skills and judgment becomes critical. These capabilities are essential for validating and challenging AI outputs. Without them, organizations risk a negative feedback loop where a lack of human expertise undermines effective oversight of the systems they rely on.

Just because we have AI assistants doesn’t mean we stop being the experts. Future AI talent will need to combine technical depth with strong critical thinking, continuous skepticism, and the ability to effectively guide and work with AI systems.

Designing scalable AI architectures for the future enterprise

Wade Tsai, EY Distinguished Technologist, EY Global Client Technology Architecture Leader

When building AI systems at scale, how do you align enterprise architecture with rapidly evolving AI capabilities?

Tsai: There is a commonly held misconception that architecture is very rigid, when this is not the reality at all. In fact, architecture can be very adaptable and flexible, if it was designed for that purpose. Especially when standards and protocols are ingrained into the design.

For example. HTTP gave rise to the dynamic internet architecture; control groups led to the rise of containerization/Kubernetes and the microservices architecture; and Transport Layer Security, embedded in all modern web browsers, enabled secure network designs while maintaining versatility.

So, as we look to engineering AI at scale, it is critical that enterprises design an architecture that is dynamic, standards‑driven, and governed end‑to‑end. So that it can adapt as AI capabilities accelerate.

What distinguishes successful industry-specific AI tools from generic platforms, and how should organizations approach this build vs. buy decision?

Tsai: The real business value from AI will come from capturing and operationalizing domain expertise. Generic platforms won’t be able to replicate this kind of nuance, which really is what distinguishes successful industry-specific tools.

However, the build vs. buy decision is ultimately a situational one. At EY, for example, we have clients and deadlines to meet. If the opportunity is here now and no product exists, we must build it. But if there’s a vendor solution that provides the same capability for half the cost or more performance and is not subject to any concerns, then we buy it. 

As AI becomes embedded across systems, how do you ensure interoperability and long-term scalability in enterprise environments?

Tsai: Interoperability and long-term scalability in enterprise environments will be enabled by enforcing open protocols and architecture standards that allow agents and platforms to communicate consistently, even as the ecosystem evolves. Think back to HTTP. It really transformed the way we think about the internet by becoming the common protocol that all browsers honor. I see MCP and A2A playing that same role as our internal standards for AI. And by having this same baseline, we will mitigate the impacts of architectural sprawl and vendor lock-in.

Zeroing in on enabling scalability, it will also be critical that we ensure agents stay manageable. This means we’ll also need AI governance and control mechanisms to ensure that we know what agents have access to and how they behave. 

What should today’s engineers focus on learning to stay relevant in an AI-first world?

Tsai: My prediction is that the current architecture design for AI will eventually plateau in terms of performance without a major breakthrough. Right now, AI makes assumptions about the way we reason and delivers responses, accordingly.

However, the ability to think through a problem, reason a problem and then command the tools that are powered by agents to solve the problem – that’s what’s going to drive value for individuals and enterprises in the future.

By understanding what AI does, we take the fear out of it. And by removing that element of fear, we can maximize its ability and ours.

The Rise of Agentic Workflows in Enterprise Systems

Sasmita Panda, EY Distinguished Technologist, EY Consulting Global Delivery Services – Technology Strategy and Transformation Architecture Leader

Agentic workflows are redefining enterprise operations – how do you see them transforming traditional software systems?

Panda: Enterprise software has long been built on fixed rules and deterministic logic. Agentic workflows mark a fundamental shift to intent‑driven orchestration. One where autonomous agents reason, collaborate, and adapt in real time – turning systems from passive records into active engines of action. The real breakthrough isn’t just productivity. But re‑architecting for governance, trust, and observability, so humans set intent and guardrails while agents execute intelligently at scale.

In what ways is GenAI reshaping omnichannel commerce, particularly in personalization and customer engagement?

Panda: GenAI is reshaping omnichannel commerce from fragmented, history‑based personalization to continuous, context‑aware engagement driven by a unified view of customer intent across channels. Combined with agentic AI, it transforms the 5C’s into a real‑time, intent‑driven engine – powering intelligent interactions, generative content, precision campaigns and autonomous decisions. The result is a shift from transactional commerce to relationship‑led platforms, where relevance, trust and experience define competitive advantage.

As AI changes how software is built, what new skills do developers need to remain effective and competitive?

Panda: As AI reshapes software development, developers are evolving from code producers to designers of intelligent systems. The competitive advantage now hinges on proficiency in prompt engineering and model integration, strong command of data quality and feedback loops, and the ability to design effective human‑in‑the‑loop workflows. In an AI‑native enterprise, value comes from building governed, adaptable systems. One that delivers outcomes responsibly at scale, not from writing more code.             

As a leader in tech, what changes are still needed to improve representation and advancement for women in AI and engineering?

Panda: Advancing women in AI and engineering requires redesigning the career systems, not just increasing participation. This means transparently allocating high‑impact opportunities, evolving leadership models to value systems thinking and ethical judgment and holding inclusion accountable like any business metric. In AI‑led organizations, diversity is not an initiative—it’s a design requirement for trust, relevance, and sustainable innovation.

In conclusion

The insights shared by EY’s distinguished technologists highlight that successfully engineering AI at scale will depend not only on models, but on the architectures, governance frameworks, cybersecurity strategies, and engineering cultures built around them.