Most GBS Functions Have an AI Vision. Very Few Have Finished Building It.

By Te Sheng Chang Head of Digital, AGOS Asia

Ask a room of GBS leaders whether AI belongs in how they run their function, and almost three quarters will say yes, without hesitation. That is the number we reported in our September 2025 white paper with Roland Berger, co-authored with Laurent Doucet, Partner at Roland Berger Asia, and it tracks with what I hear in the room. Then I will ask how far along they actually are in building the system to deliver on it, and the room gets quieter. Usually somebody mentions a pilot that has been running longer than planned, or a use case still sitting in review. Most GBS functions I speak with have the vision locked in. Very few have finished building the operating system underneath it, and a lot of them have not fully started.

Vision: the easy part

74% of the GBS leaders we surveyed said they have a clear vision for AI. That is a real number, and leaders have genuinely moved past debating the direction. But a vision is the cheapest part of any transformation you will ever pay for. An offsite produces it, a slide deck carries it, and a townhall applauds it. What it will not do, on its own, is change what anyone is doing on Monday morning. Every other number in our study comes in lower than this one, and that gap is what the rest of this piece is about.

Ideas: where the first crack shows

In our study, only 37% of organisations have internal AI training programmes running, which is really a proxy for something bigger. It tells you whether the function has built any internal capability to turn curiosity about AI into proposals, pilots, and eventually usable output. Roughly two out of every three GBS functions have not built that layer, which means they are hoping good ideas will show up through individual initiative, a vendor demo, or a consultant's deck. Hope is not an ideas engine. My honest view, having sat with a lot of these programmes, is that most of them underestimate what AI requires from the people using it. This is not like RPA or analytics a few years back, where exposure and a few optional workshops were enough to get teams moving. AI is changing how people make decisions, not just how they run tasks, and the organisations generating real ideas are the ones who figured that out early. Treating this skill gap as familiar territory costs you about 18 months before you realise.

Rao Murugesh, who runs Global Shared Services at Roche, described the underlying issue in our research: "The biggest challenge we face is fragmentation. Our teams operate in silos, each with their own tools, processes and data systems. This makes it difficult to move forward cohesively, especially when we are trying to scale AI initiatives across the organisation." What makes Roche interesting is not that they solved fragmentation cleanly. Nobody has. What they did was build the conditions for AI ideas to surface from inside the function, through Roche Tech University, a permanent internal capability that gives their people a shared baseline in AI and data. Once that layer was in place, the ideas coming out of Roche stopped being one-off experiments proposed by individuals. They started coming from teams who could see where AI would help inside their own processes, grounded in the work rather than imported from a vendor deck. Good AI ideas come from people who understand both the technology and the business problem, and who have been given the frame to work through the two together. If 37% of organisations have built that layer, the other 63% are waiting for ideas to surface from individual initiative, which is a much slower and less reliable engine.

Implementation: where approvals kill momentum

40% of GBS leaders told us AI is meaningfully freeing employees for more strategic work, which is marginally better than the training number but still a minority. Puneet Gupta and Lalitha Kumar at Air Liquide described the constraint directly: "We have strong ideas and good early results, but the time between identifying a use case and getting it into production is painfully long. Most of that delay is not technical. It is governance, approvals, and risk reviews that were not designed with AI in mind." 

The bottleneck is not talent, and it is not the technology. Most companies are still running AI through approval processes built for traditional IT work, where scope gets fixed upfront and every release must line up against criteria that were written before machine learning was on the agenda. AI does not behave like that. A use case can be built in days, change direction once real users get their hands on it, and look different by the time it shows up in risk review. The processes meant to govern it are moving at quarterly-release speed, and the mismatch produces backlogs inside large GBS functions. Pilots stack up, and very few of them graduate to production, because nobody sat down and properly designed the path between the two.

Measurement: the part almost nobody has built

This is the number I would sit with longer than the others. Only 20% of organisations in our study have a framework for measuring and rewarding AI contributions, which leaves 80% running programmes they cannot formally measure and cannot credibly defend when the CFO asks what the return has been. Here is where I think most programmes get it wrong. Measurement gets treated as a finishing step, something you add once the tools are live and the teams are trained. That sequence is backwards. If you cannot show whether an AI tool is reducing errors, saving time, or improving decision quality, you have nothing solid to scale from. The measurement framework, covering everything from personal productivity tools through to value-added analytics, needs to be designed in from the start. It is the scaffolding the rest of the programme hangs off.

Russell Parry at AstraZeneca is one of the few exceptions I have seen close up. AstraZeneca has run a modular AI training programme since 2023, and because measurement was built into the design from day one, he can actually point to outcomes: "We have seen measurable gains in both productivity and retention since rolling out our modular training approach. People want to work where they are being invested in, and they want to work on things that feel like the future." Without that discipline, everything upstream floats. You cannot say with confidence what is working, and sooner or later the investment case starts to wobble because there is nothing underneath it to point to.

Where the real gap is

Vision 74%. Training 37%. Implementation 40%. Measurement 20%. The further you move from the strategy slide into day-to-day operations, the smaller the number gets, and that is the honest shape of AI in GBS right now. The ambition is not the issue. The support system underneath the ambition gets thinner the deeper you look into the organisation, and most leaders I talk to have not fully reckoned with that yet. The GBS functions that pull ahead in the next 18 months will not be the ones with the sharpest vision, because most leaders have got that part sorted already. They will be the ones willing to do the less visible work, which means building the training layer, so ideas come from a system and not by accident, fixing the approval path so good ideas do not die in review, and putting proper measurement in place before the programme drifts. The vision is there in most organisations I speak to. The system behind it still needs to be built.

AUTHOR BIO Chang Te Sheng is Head of Digital at AGOS Asia, a GBS and digital transformation advisory firm based in Kuala Lumpur. He is co-author of the 2025 Roland Berger x AGOS White Paper on GBS and AI transformation, produced with Laurent Doucet, Partner at Roland Berger Asia. He will be speaking at AGOS GBS Summit 2026 on 10 September at Sheraton Petaling Jaya, the ninth edition of the annual GBS strategy event in Malaysia.

SOURCE NOTE All statistics and direct quotations from external leaders are drawn from the 2025 Roland Berger x AGOS White Paper on GBS and AI Transformation, co-authored by Chang Te Sheng, Head of Digital at AGOS Asia, and Laurent Doucet, Partner at Roland Berger Asia. The full paper is available on request.

Request here