Exclusive Insights from

Gonçalo Conceição, Director - myAZ Services, AstraZeneca

_{EXCLUSIVE EXPERT INSIGHTS}

Agentic AI Isn't the Challenge: The Real Challenge is Delivering AI-Enabled Service Management

Gonçalo Conceição, Director - myAZ Services, AstraZeneca, shares how they built an AI-powered employee support model serving 120,000 colleagues globally, and why data quality, knowledge governance, process ownership and change management proved far more critical than the AI itself.

When you first explored agentic AI, what outcomes were you expecting to achieve, and how do those expectations compare with reality?

When we first explored agentic AI, our goal was to solve a clear business problem: simplify a fragmented support environment. Colleagues were navigating more than 80 channels, and, in 2023, they raised nearly 1.8 million support tickets, most requiring human intervention. We wanted myAZ to become a single-entry point where employees could ask for what they needed in natural language, access trusted information, complete routine tasks, and only be escalated when necessary.

That expectation has broadly been validated. myAZ is now available to over 120,000 colleagues across 78 markets, supporting around 28,000 monthly active users and more than 68,000 monthly conversations, with fewer than 8% requiring escalation to a ticket or live agent. We have also seen measurable reductions in HR ticket volumes and strong self-service outcomes.

What reality has taught us is that success depends on more than the AI itself. The biggest drivers of impact have been knowledge quality and governance, thoughtful UX, and change management with a strong focus on communication and user education. The opportunity is real, but success depends less on how “autonomous” the AI is, and more on how well it is integrated, governed, and adopted.

Where have you seen agentic AI create the greatest business impact so far, and what surprised you most about the results?

The greatest business impact we have seen from agentic AI so far has been in turning myAZ into a single, scalable front door for employee support across

HR, IT and Finance. It has helped shift us away from a fragmented support landscape of more than 80 channels towards a more intuitive model in which colleagues can ask for what they need in natural language and either receive an answer, complete a task, or be routed to the right support path.

The clearest results have come through self-service and ticket deflection, particularly in HR. Pilot markets adopting a digital-first model saw up to a 24% reduction in HR case volumes for a 3-month period. In Finance, for example, we observe a monthly average of 6,000 self-served cases, equivalent to nearly 11 FTE savings.

More broadly, myAZ now supports over 120,000 colleagues across 78 markets, with around 28,000 monthly active users, more than 68,000 monthly conversations, and fewer than 8% requiring escalation. What surprised us most was that the strongest impact came not from AI alone, but from the combination of AI, quality knowledge, channel strategy, and user experience.

That really reinforced for us that the value of agentic AI comes not just from what it automates, but from how well it is integrated into the Employee Experience.

Can you share an example where poor data quality, fragmented process ownership, or inconsistent governance limited the effectiveness of an AI initiative?

One of the clearest lessons from myAZ has been that the effectiveness of AI is directly shaped by the quality and consistency of the environment around it. In our case, one of the early constraints was not the AI capability itself, but the fragmentation of knowledge sources, support channels, and process ownership across functions.

When information is spread across multiple systems, maintained to different standards, or owned inconsistently, the assistant can only ever be as reliable as the content and processes behind it.

A practical example was the variation in knowledge quality and structure across different domains. Because myAZ depends on trusted enterprise content to answer questions and guide actions, inconsistencies in article quality, metadata, maintenance discipline, or process clarity could reduce answer relevance and create gaps in the experience.

In some cases, this meant the assistant could not confidently resolve a request end to end and instead had to fall back to routing, clarification, or escalation.

This reinforced an important principle for us: AI does not remove the need for strong governance; it exposes where governance is weak. In response, we increased our focus on knowledge remediation, global knowledge standards, clearer ownership models, and stronger cross-functional governance.

That work has been just as important as the AI itself in improving reliability, user trust, and business impact.

What processes or functions are proving most suitable for agentic AI today, and what characteristics make a process ready for autonomous execution?

The processes most suitable for agentic AI today are typically those that are frequent, repeatable, rules-based, and connected to reliable enterprise systems.

In the case of myAZ, these include a range of employee support tasks across HR, IT and Finance, such as answering policy questions, checking request

status, booking time off, viewing payslips, checking expenses, requesting access, or guiding users to the right support path.

What makes a process ready for autonomous execution is a combination of clarity, standardisation, and control. The process needs clear rules, defined ownership, good underlying knowledge, and structured integration with the relevant systems. It also needs guardrails so that the AI can act confidently when conditions are straightforward and escalate when they are not.

In practice, the most suitable processes are not always the most sophisticated ones, but those with the right foundations: consistent data, strong governance, low ambiguity, and a well-defined user journey.

As agentic AI takes on more decision-making and execution responsibilities, how are you approaching governance, accountability, and human oversight?

As agentic AI takes on more responsibility, we are making sure that governance, accountability, and human oversight scale alongside capability. With myAZ, we are clear about where the assistant can answer, where it can act, and where it must escalate or follow an approval route. We only expand autonomy where the process, knowledge, and controls are mature enough to support it safely.

Accountability remains with the relevant business and process owners, supported through cross-functional governance across GBS, HR, IT and Finance. Human oversight is built into the model through scoped use cases, permissions, approvals, escalation paths, and feedback loops.

A key enabler is myAZ Compass, our internal insights and analytics tool,

which provides a single trusted view of how myAZ is performing across functions. It helps us monitor user behaviour, system performance, and business impact, and identify where governance, improvement, or intervention is needed.

Our overall approach is human-led, AI-enabled, with trust and clear ownership at the centre.

Where are agentic AI initiatives being held back today, and how much of that is a data, process, or ownership challenge rather than a technology limitation?

Today, agentic AI is often held back less by the technology itself and more by the surrounding environment in which it operates. The main constraints are usually data quality, process standardisation, knowledge maturity, governance, and ownership clarity. Where information is fragmented, or accountability is unclear, AI can still support users, but it becomes harder to automate confidently and at scale.

That has been a key lesson from myAZ. The assistant is only as effective as the processes, knowledge, and governance behind it. If a process is inconsistent across functions or markets, or if the underlying content is not well maintained, the AI will be limited regardless of how capable the model is.

So, while technology continues to improve rapidly, the bigger challenge is often organisational readiness. In many cases, data, process, and ownership are the real factors determining how far agentic AI can go.

What foundational changes have had the biggest impact in enabling agentic AI to deliver measurable value?

The biggest foundational changes have been around the environment

supporting the AI rather than the AI alone. For myAZ, the most important enablers have been stronger knowledge management, clearer process ownership, channel simplification, and focused change management. These have been critical in turning agentic AI into something reliable, trusted, and scalable.

Improving knowledge quality and structure has had a direct impact on answer relevance and self-service. Simplifying support channels has made adoption easier, and through myAZ Compass, we have created a single trusted view of performance, user behaviour, and business impact across HR, IT and Finance, which helps us continuously improve and prioritise value.

The overall lesson is that measurable impact comes when technology and operating foundations evolve together.

If you could give one piece of advice to a GBS leader who is currently evaluating agentic AI, what would it be, and what is the one mistake you would urge them to avoid?

My advice would be to begin with a specific business problem rather than with the technology itself. Agentic AI creates the most value where there is clear friction, high volume, and a process that can genuinely be simplified for users.

The biggest mistake to avoid is assuming AI can fix weak foundations. If data, knowledge, process ownership, or governance are not in place, agentic AI will highlight those gaps rather than overcome them – AI amplifies, for the good and the bad.

So, the key is to focus first on the problem, the operating model, and the foundations for trust, then use AI to scale what works.

Tell Us Your Thoughts…

What stage of Agentic AI Implementation is your shared services/GBS currently at?

Evaluating
Planning
Piloting
In Production
Scaling