Partner content: Agentic AI and enterprise transformation

‘Rethink the work before writing the code’

A Q&A with Dmitry Tikhomirov, Global Head of Microsoft Business Group, EPAM Systems, Inc.

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

Generative AI has already proved useful for drafting, summarizing, analysing and creating content. But in most cases, the work still starts with a human prompt and ends when the model gives a response.

Agentic AI changes that pattern. An agent is given a goal, not a script. It can work out the steps, choose the tools it needs, bring in other specialised agents where useful and continue until the task is complete. In an enterprise setting, that has to happen within clear guardrails, including permissions, policies, data controls and human oversight where needed.

That shift matters because it changes how organisations think about work. Instead of using AI only to help people complete individual tasks, businesses can begin to redesign processes so that agents handle more of the coordination, information gathering and decision support across systems.

The result is a move toward more adaptive ways of operating. Work can happen faster, decisions can be made with better context and teams can spend less time moving information from one place to another. That is where agentic AI starts to create a different kind of enterprise value.

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

The value is showing up in the challenging parts of the enterprise. Underwriting, customer onboarding, risk analysis, software engineering, field operations, store operations and service centers. These are the places where people spend too much time pulling information from different systems, checking documents, chasing approvals and translating between tools that were not necessarily designed to work together.

That is where agents start to earn their keep. A strong agentic workflow removes friction from the process. It helps people get to the right answer quicker, with fewer handoffs and less manual interpretation along the way.

Retail is a good example. Albert Heijn, part of Ahold Delhaize, has used Azure OpenAI as part of an employee-facing assistant within its store app. Teams can use it for questions around restocking, onboarding, product lookup and inventory information. At that scale, across hundreds of stores and tens of thousands of employees, even small improvements in everyday work can start to make a big difference.

The measurement, though, has to be more than “people saved time.” Our AI Value Capture research shows what good measurement looks like. In one software engineering example, agentic workflows reduced documentation time by 80% and code generation time by 45%, which translated into a 30% increase in net sprint velocity. In a contact center example, conversational AI drove 40% call deflection, improved efficiency on remaining calls by 15% and lifted CSAT by nine points.

So, success should be measured at the level of the process. Time saved is useful, especially early on, but the stronger measures are process cost, cycle time, decision quality, revenue impact, customer experience and the amount of work that can move through the system without adding headcount.

The practical lesson is to start where friction is visible and value can be measured. Establish the baseline before the pilot starts. Then track whether the agent improves the economics of the workflow, not just the productivity of an individual task. That is how organisations move from interesting AI activity to measurable business impact.

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

This is where many businesses lose momentum. Despite there being plenty of experimentation, there’s still a real gap between promising pilots and production-grade AI.

Our own research points to the same pattern. In EPAM’s 2025 AI report, only 26% of disruptors and advanced companies delivered AI use cases to market, compared to 9% of competent companies and just 3% of beginners. Additionally, EPAM’s AI Value Capture research adds another important signal: 43% of companies struggle to secure AI budgets because they cannot prove ROI, and 60% of successful pilots fail to scale across the enterprise.

The reasons are usually familiar. The pilot is interesting, but it is not tied tightly enough to a business priority. The team proves that the technology works but does not clearly establish the financial case. The workflow depends on data, permissions or systems that were never prepared for autonomous execution. As a result, governance arrives late and change management is delayed even further.

Organisation
The first step is to build the operating model around the ambition. That means clear executive sponsorship, accountable owners, AI-literate teams and a funding path that can move from validation to scaling. EPAM’s AI report found that only 1% of companies say they have a fully effective AI governance model, which tells us that many organisations are still building the management capability needed to scale AI safely.

Process
The work itself has to be redesigned. Agentic AI creates value when it can operate across a real business process, with access to the right data, tools and decisions. A pilot should start with measurable baselines and prove value against operating metrics such as cost reduction, revenue impact, cycle time, quality or customer experience. Without that link to the workflow, even a technically successful pilot will struggle to justify the next round of investment.

Technology
The stack has to be modular, governed and observable from the beginning. Agents need secure access to enterprise data, clear permissions, audit trails, monitoring and lifecycle management. When an agent is expected to plan, decide, call tools and act across systems, the enterprise has to know what it can do, what it has done and when a human should step in.

Organisations that navigate this well treat the pilot as the first stage of production, rather than a simple experiment. They define value before they build, prepare the process and data around the agent and put governance in place early enough to support speed.

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

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

The best way to think about Microsoft’s role is as a platform for building, running and governing agents across the enterprise.

Model choice is the starting point. Different workflows will need different models, depending on reasoning depth, cost, latency, security and domain fit. Azure AI Foundry gives enterprises a way to evaluate and deploy those models in a controlled environment, without locking every use case into a single technical path.

The next layer is agent development and execution. Azure AI Foundry Agent Service and Microsoft Agent Framework give technology teams a more structured way to build, test and manage agents in production. That matters because enterprise agents have to operate across real systems, follow permissions, call tools safely and leave a trace of what they did.

There is also the application layer, where adoption starts to become practical. Microsoft 365 Copilot, Copilot Studio and role-specific agents bring agentic capabilities into the places where people already work. For many organisations, that will make the difference between isolated experimentation and everyday use.

The final layer is governance. Agents need identity, access controls, data boundaries, monitoring, auditability and lifecycle management. Microsoft’s broader stack, including Entra, Purview, Defender and Agent 365, gives enterprises a familiar control plane for managing these new digital workers with the same seriousness they apply to people, applications and services.

For enterprises, the priority should be platform discipline from the beginning. They need to decide how agents will access data, which actions they can take, how approvals will work, how performance will be measured and where human oversight is required.

The companies that benefit most will be those that connect agentic AI to real business processes, with governance and scalability designed in from the start. That is where Microsoft is positioning itself: as the enterprise layer for agentic AI, from model selection through to production operations.

What actionable advice would you give to organisations getting started?

Rethink the work before writing the code. Too many teams start by asking where they can add AI. The better starting point is the work itself. What is slow today? What costs too much? Where are employees chasing information, checking documents, waiting for approvals or fixing the same problems again and again?

Once the problem is clear, choose a use case with a real owner and a clear measure of success. Underwriting triage, onboarding, store operations, service support and operations intelligence are good places to start because the work happens often, the pain is visible and the business value can be measured. A pilot should answer a simple question: did the process become faster, cheaper, more accurate or easier for people to use?

Make the pilot as close to the real world as possible. It should touch the data, systems and users that the full rollout will depend on. A clean demo can create excitement, but the useful learning comes when AI meets permissions, approvals, data gaps, unusual cases and the way people actually work.

Build agents with specific jobs. A broad general agent sounds attractive, but it is harder to control and harder to measure. A claims intake agent, a store colleague assistant or a policy lookup agent can have clear boundaries, trusted sources and defined points where a person reviews the work. That makes the system easier to improve over time.

Measurement has to be practical. Before the pilot starts, capture how the work happens today: how long it takes, what it costs, where mistakes happen and what customers or employees experience. Then compare that with what changes after AI is introduced. Just as importantly, count the cost of keeping the solution running, including the tools, support, training and improvements it will need over time.

Trust has to be designed in early as well. Leaders should decide what data the agent can use, what actions it can take, who approves sensitive decisions and how the organisation will review what happened. Clear rules give teams the confidence to move faster because people know where the boundaries are.

The first agent will rarely be the last, so build in a way that leaves room to grow. The setup should allow new tools, models and workflows to be added without starting again each time. Teams will also need training, managers will need new ways to supervise AI-supported work, and the organisation will need a culture of continuous improvement after launch.

EPAM’s work with Albert Heijn, the Netherlands’ largest grocer, is a good example of this approach. The production assistant built on Azure AI Foundry was designed around real store work: product lookup, stock questions, restocking, onboarding and customer support. The lesson is straightforward. Start with work that matters, give the agent a clear role, measure what changes and build the controls needed to scale.

Dmitry Tikhomirov is Global Head of Microsoft Business Group, EPAM Systems, Inc.

Find out more about EPAM and our partnership with Microsoft.

This article is sponsored by EPAM.