Partner Content: How organisations can get the most value from agentic AI

How Microsoft platforms and tools are accelerating adoption for our clients

What is agentic AI and how is it reshaping the business landscape?

Dmitry:
Generative AI was, in effect, a very capable typewriter. It could draft an email, summarise a contract or produce a passable image, but each act began and ended with a human prompt. Agentic AI changes that pattern. An agent is given a goal, not a script; it plans the steps, chooses the tools, invokes or even creates dynamic agents where helpful, and keeps going until the work is done - all autonomously and within guardrails set by the enterprise.

What we are seeing now is a shift in how organisations operate. Agentic AI allows businesses to move from human-led, AI-assisted processes to autonomous, adaptive systems, creating dynamic, data-driven environments that can act in real time.

This is a fundamentally new model of enterprise value.

Where are organisations getting real value today, and how should success be measured?

The pattern is consistent. Agents earn their keep in high-friction, multi-system workflows: underwriting, customer onboarding, risk analysis, engineering, field operations and many others. These are processes in which people spend their days gathering information, interpreting documents and bridging gaps between systems. Multi-agent architectures - where specialised agents hand work to one another - collapse that choreography.

But the bigger shift is in how value is measured.

It is no longer just about time saved. Agentic AI creates value by removing entire layers of cost and complexity, and by enabling faster, higher-quality decisions.

For example, Ahold Delhaize is using Azure OpenAI Service as part of its broader digital transformation to bring agentic AI to a very large retail footprint. Given its scale - 700 owned stores across the Netherlands, over 80,000 people who work on the shop floor, interacting with the millions of customers that come through its doors each day - even modest productivity gains can translate into significant enterprise value.

In summary, organisations should benchmark outcomes: revenue lift, cost reduction, cycle-time compression and decision quality.

Why do so many agentic pilots never reach production, and how can that be navigated?

This is where many organisations get stuck. Despite huge enterprise AI investments, 95% of pilots fail to reach production (MIT).

The biggest issue is misalignment. Pilots are often disconnected from real business priorities or do not account for governance, security and data-access requirements.

Another challenge is trying to layer AI onto existing processes. That usually delivers incremental gains, not transformation.

To move forward, firms need to focus on three pillars:

• Organisation - Build the operating model, literacy and accountability before the code.
• Process - Redesign the workflow with agents at the centre, not plastered on top.
• Technology- Choose a modular, governed, observable stack that can evolve.

The shift from pilot to production is as much about mindset and structure as it is about technology.

Where does Microsoft fit in the agentic AI stack, and what should enterprises prioritise?

Microsoft has, quietly and rapidly, assembled one of the more complete enterprise-ready agentic stacks. The model catalogue is broad and includes models from OpenAI, Anthropic, Meta, DeepSeek and other providers, alongside access to a wide range of open models. At the centre of Microsoft’s proposition is an increasingly coherent and secure agent stack:

• Work IQ, Fabric IQ and Foundry IQ form the backbone. Together, they are meant to ground agents in the reality of the enterprise: how work gets done, where the data sit and how AI systems are built, governed and improved. Work IQ brings intelligence about workflows, collaboration patterns and business activity; Fabric IQ anchors agents in governed, enterprise-scale data; and Foundry IQ provides a managed, permission-aware knowledge layer that connects structured and unstructured enterprise data so agents can access and act on trusted enterprise knowledge.
• Microsoft Foundry Agent Service is the runtime layer: the place where agents are built, deployed and orchestrated in production. Microsoft is positioning it as the enterprise execution environment for multi-agent workflows, with growing support for interoperable patterns such as Model Context Protocol (MCP).
• Microsoft Agent Framework is the developer substrate beneath that runtime. Positioned as an open-source framework for Python and .NET, it brings together Microsoft’s earlier agent-development approaches into a single orchestration layer for building, testing and managing agents with more discipline than the first generation of AI experimentation allowed.
• Copilot Studio is the low-code factory floor. It gives business teams and developers a way to create and connect agents without rebuilding everything from scratch, and Microsoft has made it a central part of its argument that enterprise agents should not remain the preserve of specialist engineering teams.
• Microsoft 365 Copilot and its expanding family of agents, such as Researcher, Cowork and many others, sit at the application layer, where the theory meets the employee. Increasingly, Microsoft is embedding specialised agents and agentic capabilities into the daily software of work-across collaboration, productivity and role-specific tasks-so that agents are not a separate destination but part of the flow of business.
• Microsoft Agent 365 is emerging as a management and governance layer, with control-plane capabilities for discovering, registering and governing agents across the enterprise estate.
• Microsoft Entra, Purview and Defender extend identity, governance and security to agents as if they were a new class of digital worker. That is a subtle but important shift. Microsoft is not merely treating agents as software artefacts, but as participants in the enterprise environment who require authentication, authorisation, monitoring and policy enforcement.
• And finally, GitHub Copilot has become a “dark factory” for code, moving from requirements to production-grade features in hours, not weeks. For enterprises, the appeal is no longer code assistance but autonomous software engineering.

If you are building with Microsoft, the priority should be the AI platform, governance and scalability from day one-ensuring systems are secure, traceable and designed to evolve.

What actionable advice would you give to organisations getting started?

Rethink the work before writing the code. The dominant failure mode is inserting AI into workflows designed for people and filing cabinets. That delivers incremental gains when the opportunity is transformative.

Six practical principles:

1. Anchor to business objectives, not technology. Agentic AI is a means; revenue, risk and cycle time are the ends.
2. Start with high-impact, low-complexity use cases - underwriting triage, onboarding and operations intelligence-and prove the operating model before scaling.
3. Rethink ROI. Value comes from eliminating cost structures, not shaving minutes. A 12x acceleration is not a 12% improvement.
4. Build for specialised agents, not generic ones. Narrow agents with clear roles, well-governed tools and human-in-the-loop checkpoints outperform “one big model that tries”.
5. Embed governance from the first line of code. Identity, data and lifecycle are not afterthoughts; they are the foundation of trust.
6. Design for the future. Whoever builds rigid, bespoke architectures today will be refactoring them by year-end. Modularity is not elegance; it is insurance.

EPAM’s work with Albert Heijn, the Netherlands’ largest grocer, shows the point: a production AI assistant on Azure AI Foundry helped store staff with restocking, onboarding and product lookup, earning EPAM the 2025 Microsoft Innovate with Azure AI Platform Partner of the Year award. The lesson is simple: when business process design, data, governance and engineering are in place, AI scales quickly and delivers returns.

This article is sponsored by EPAM.