There's a version of the future where every Global Capability Centre has a gleaming AI Centre of Excellence at its heart, staffed with engineers, AI ethicists, LLM specialists, and a Chief AI Officer who reports directly to the CEO. Boards are excited. Analysts are bullish. Vendors are lining up in the lobby. That future is real. It's just not imminent. And the enterprises who understand why will avoid some very expensive mistakes.
The Narrative Has Outrun the Readiness
Let's name what's happening: the AI-native GCC is the hottest story in the India GCC market right now. Every advisory firm, every IT services company, every real estate developer pitching a new SEZ and every hiring company is packaging their offering around AI. The storyline is seductive: India's talent pool, combined with AI, will transform the GCC from a cost centre into a genuine innovation engine.
The problem is that a compelling narrative is not the same as enterprise readiness. And when you strip away the conference keynotes and analyst reports, most large enterprises are simply not in a position to build and sustain a full-blown AI COE inside their GCC, at least not yet.
This isn't pessimism. It's realism. And the distinction matters enormously for how CXOs should be thinking about their India strategy right now.
Eight Inconvenient Truths About Enterprise AI COE Readiness
1. The Core Business Isn't AI-Ready Yet
Before you build an AI COE, ask yourself a more fundamental question: is the underlying enterprise AI-ready?
In most mid-to-large enterprises, the honest answer is no, not fully. Data is fragmented across ERPs, legacy systems, and business units that have never agreed on a common taxonomy. MDM initiatives are perpetually "in flight." Even basic reporting often requires manual reconciliation that no one talks about in board presentations.
An AI COE built on top of this substrate will spend the majority of its time doing data engineering, not AI. That's not a failure of the COE, it's a structural reality. And it means the headline ambition ("AI-native GCC") masks what the centre will actually do for the first two to three years.
Enterprises that are honest about this will sequence things correctly: data estate first, AI capability second. Those that aren't will announce an AI COE and quietly pivot it into a data analytics team twelve months later.
2. The Talent Math Doesn't Work the Way Everyone Assumes
India has world-class AI talent. That's true and important. But the talent required for a high-functioning enterprise AI COE is not the same as the talent required for a GCC doing IT services, finance operations, or analytics support.
A credible AI COE needs people who can operate at the intersection of domain knowledge, ML engineering, data science, enterprise architecture and enterprise business context. These are not fresh graduates you can put through a six-month training programme. They are experienced professionals who have already made their career choices, and those choices increasingly include Indian AI product companies, global tech firms, and well-funded startups that offer equity, speed, and the chance to build something genuinely new.
Competing for this talent as a non-tech enterprise GCC is genuinely difficult. Compensation is one factor. But the more stubborn factor is the work itself. The best AI engineers want to work on problems that stretch them, not maintain models built around business-as-usual workflows. Unless the enterprise can articulate a genuinely ambitious AI problem statement, the talent proposition is weak.
Add to this the churn risk. AI talent in India is highly mobile. Building a COE is easy. Building a stable, institutionalised COE with knowledge retention, mentorship structures, and a defined career ladder, that's a multi-year organisational investment that most enterprises have not fully costed.
3. Governance Is Still Being Invented
Every week brings a new framework, a new regulation, a new incident that reshapes what responsible AI deployment looks like. The EU AI Act is live. The US is developing sector-specific guidance. India's Digital Personal Data Protection Act is being operationalised. Many enterprises are simultaneously trying to comply with multiple regulatory regimes across multiple jurisdictions, and the rules keep changing beneath their feet.
Building an AI COE without a clear, enterprise-wide AI governance framework is building on sand. Yet most enterprises don't have that framework. They have a policy document, maybe a review committee, possibly an AI ethics charter that was drafted and is being continuously redefined. The actual decision rights; who approves an AI model for production use, what the escalation path is when a model behaves unexpectedly, how IP generated by AI tools is owned and protected, are often genuinely unresolved.
In this environment, an AI COE will either move slowly (because every output requires legal review) or move fast and create liability (because governance wasn't established upfront). Neither outcome is acceptable.
Getting governance right takes time and requires alignment between Legal, Compliance, Business, and Technology leadership, alignment that doesn't happen in a quarter. Enterprises that try to shortcut this step will pay for it, either in regulatory exposure or in internal credibility damage when something goes wrong.
4. Geopolitics Has Entered the Technology Stack
This is the dimension that rarely appears in AI COE business cases, but it should. The AI landscape is increasingly shaped not just by technology choices but by geopolitical ones.
The US-China technology decoupling has direct implications for enterprise AI architecture. Certain chips, certain models, and certain cloud infrastructure choices are now entangled with export controls, sanctions risk, and government scrutiny in ways that were unimaginable three years ago. An enterprise building an AI COE today must ask which models it intends to use, which infrastructure it will run them on, which vendors it will depend on, and whether those choices are durable in a world of accelerating technology nationalism.
For US-headquartered enterprises in particular, the question of whether to use open-source models with Chinese origins, or to rely on hyperscaler infrastructure that may itself be subject to shifting regulatory treatment, is not a technical question. It is a boardroom question. And most boards are not yet equipped to answer it with confidence.
Meanwhile, allied governments are increasingly treating AI capability as a strategic asset. Some jurisdictions are beginning to require that certain AI workloads run on domestically hosted infrastructure. This creates new constraints on where an AI COE can operate, which models it can deploy, and which data it can process, constraints that are not yet fully visible but are clearly directional.
Enterprises that stand up a full AI COE today may find that a significant portion of their architectural decisions need to be revisited as the geopolitical picture clarifies. That's a real cost, and a real reason to be measured.
5. Data Sovereignty Is a Harder Problem Than It Appears
Closely related to geopolitics, but distinct enough to warrant its own treatment: data sovereignty is rapidly becoming one of the most complex operational realities facing enterprise GCCs.
When an AI model is trained, fine-tuned, or even prompted on enterprise data, questions arise immediately. Where is that data being processed? Which servers are it traversing? Which model provider is retaining it, and under what terms? If customer data from a European business unit is used to train a model running on a US hyperscaler's infrastructure, is the enterprise in compliance with GDPR? If an Indian GCC is processing data originating from a healthcare business in the US, what HIPAA obligations apply, and to whom?
These questions do not have simple answers. The legal frameworks are evolving, the vendor terms-of-service are dense and frequently updated, and the enterprise's own data classification policies often haven't caught up with the new reality of AI-driven data flows.
Establishing a full AI COE before these questions are resolved, at an enterprise level, not just an IT-policy level, creates significant exposure. And resolving them requires legal, compliance, and business alignment that is genuinely time-consuming to achieve across a global enterprise. GCCs that handle data from multiple geographies face a layered sovereignty challenge that cannot be wished away by a vendor's data residency assurance.
6. The Vendor Landscape Is Chaos, Deliberately So
Consider what a CXO faces when evaluating AI infrastructure today. There are foundation model providers: OpenAI, Anthropic, Google DeepMind, Meta, Mistral, Cohere, and a dozen others, each with different capabilities, pricing models, terms of service, and update cadences. There are hyperscaler AI platforms: Azure OpenAI Service, AWS Bedrock, Google Vertex AI, each offering access to some of these models with varying degrees of customisation, compliance tooling, and enterprise support. There are specialist AI orchestration platforms, vector database vendors, RAG framework providers, AI observability tools, prompt management systems, and fine-tuning infrastructure providers.
And every single one of these categories is consolidating, pivoting, merging, or being superseded on a monthly basis.
The honest assessment is that no enterprise, however well-resourced, can evaluate this landscape comprehensively, make confident long-term bets, and build institutional knowledge around a specific stack, all at the same time as stand up a new capability inside a GCC in a different country. The cognitive and procurement overhead alone is enormous.
There is also the question of vendor lock-in. AI capabilities today are often deeply embedded into specific provider ecosystems. Switching costs are not trivial. An enterprise that builds its AI COE on a particular stack may find, eighteen months later, that a competitor or open-source alternative offers dramatically better price-performance, but the switching cost is prohibitive. This is exactly the kind of regret that CFOs are paid to prevent.
The prudent response to a chaotic vendor landscape is not paralysis, but it is also not a full-commitment COE build on a stack that may need to be rearchitected in two years.
7. The Platforms Are Still Mid-Transformation, And So Are the Services
Perhaps the most underappreciated dynamic of the current moment: the AI platforms themselves are not stable products. They are rapidly evolving services that are adding new capabilities, deprecating old ones, changing pricing structures, and critically, consuming more and more of the value chain that was previously provided by third-party specialists.
Eighteen months ago, enterprises needed specialist vendors for AI-powered document processing, conversational AI, code generation, and data synthesis. Today, those capabilities are increasingly native to the foundation model APIs themselves, or to the hyperscaler platforms built on top of them. Third-party vendors who built businesses on top of GPT-3-era limitations are scrambling to find differentiation as those limitations disappear.
This matters enormously for a GCC building an AI COE. The roles you are hiring for today, the specialists in particular tools, frameworks, or fine-tuning approaches, may be substantially different in their scope and value eighteen months from now, as the platforms absorb more of the complexity. The training curricula you build, the certifications you invest in, the vendor relationships you establish, all of these have a shorter shelf life than equivalent investments made in, say, cloud infrastructure five years ago.
This is not an argument for waiting indefinitely. But it is an argument for building with a lighter architectural footprint than the full COE model implies, maintaining optionality, avoiding over-specialisation, and preserving the ability to pivot as the platform landscape stabilises.
8. The Organisational Change Problem Is Underestimated
Here is the dimension that almost never appears in analyst reports: the biggest barrier to enterprise AI value is not technology. It's behaviour change.
Deploying an AI tool into a business process requires the humans in that process to trust it, use it, and adapt their workflows around it. That requires change management, training, communication, and critically, leadership role modelling from senior leaders who are themselves genuinely comfortable with AI-augmented work.
Most large enterprises are not there yet. Middle management has a complicated relationship with AI, they understand, at some level, that AI may restructure their teams, and their incentive is often to slow-walk adoption rather than champion it. Without deliberate effort to address this, which is a leadership transformation challenge, not a technology challenge, even a well-built AI COE will find its outputs sitting unused on a shelf.
Building the COE without simultaneously investing in the cultural and behavioural change required to absorb its outputs is the single most common mistake enterprises make. It produces a technically impressive function that generates limited business impact, and eventually gets quietly reframed or restructured.
But the GCC Should Not Wait — There Is Real Work to Do Now
Recognising that the moment for a full AI COE has not yet arrived is not an invitation to passivity. Quite the opposite. The GCCs that will be ready when the moment does arrive are the ones doing deliberate, foundational work right now. The question is not whether to act, it's what to act on.
-
Deepen enterprise context, relentlessly. Of all the things a GCC can do to prepare for an AI-enabled future, this may be the most underrated. AI that works in a lab but fails in production almost always fails for the same reason: the people building it did not understand the business deeply enough. GCC teams that have spent years embedded in enterprise workflows, that know why a particular process was designed the way it was, that understand the edge cases finance won't document and the exceptions operations handles by instinct, these teams are sitting on an asset that no model, no platform, and no vendor can replicate. The work right now is to institutionalise that knowledge: structured business context programmes that rotate GCC talent across functions, deliberate documentation of process logic and decision rationale, and senior GCC leaders who invest time with their home-market counterparts not to take orders but to develop genuine domain fluency. When the AI COE moment arrives, the differentiator will not be which team has the best engineers. It will be which team understands the enterprise well enough to know what problems are actually worth solving, what data actually reflects reality, and what a good answer actually looks like. That is a capability built over years, not quarters, and the time to build it is now.
-
Clean the data estate. This is unglamorous, difficult, and essential. Without clean, governed, accessible data, AI is a marketing claim. GCCs are often well-positioned to lead this work; they have the technical depth, the process proximity, and the organisational incentive to build unified data platforms, establish master data governance, and create API-accessible data layers that the enterprise will depend on for every AI initiative to follow. This work has value independent of AI. But it is the foundation without which AI cannot scale.
-
Build AI literacy across the function, at every level. Not AI expertise. Literacy. GCC leaders should be ensuring that their teams, from analysts to senior managers, understand what AI tools can and cannot do, where they add value in daily work, and how to work productively alongside them. This is a training investment that pays immediate dividends in productivity and begins to create the demand-pull for AI capability that makes a future COE viable.
-
Run structured, well-governed pilots. Choose two or three use cases where the business pain is real, the data is reasonably clean, and the stakeholder in the home market is genuinely motivated. Stand up a small, senior team around those use cases, not a permanent COE, but a dedicated pilot squad. Deliver measurable outcomes. Document the learnings. Use the results to build the institutional knowledge and the internal credibility that will eventually justify sustained investment.
-
Develop an AI vendor map, and stay current on it. Rather than committing to a stack, commit to understanding the landscape. Assign someone, or a small team, with the mandate to track platform developments, evaluate emerging tools against enterprise requirements, and maintain an informed view of where the market is heading. This is not a COE function. It's a strategy and architecture function that every serious GCC should have. The enterprise that understands the vendor landscape when it's ready to decide will make far better choices than the enterprise that has to learn it under time pressure.
-
Start the data sovereignty and compliance work now. Map your data flows. Understand what categories of enterprise data the GCC handles, where they originate, what regulatory regimes apply to them, and what constraints those regimes place on AI processing. This work is tedious, cross-functional, and slow, which is exactly why it should begin before the AI COE ambition creates pressure to cut corners.
-
Build relationships with the AI ecosystem in India. The Indian AI ecosystem such as academic institutions, AI-native startups, research labs, open-source communities is genuinely world-class and deepening rapidly. GCC leaders who invest in these relationships now, through hiring pipelines, research partnerships, hackathons, and community engagement, will have a significant sourcing and collaboration advantage when they are ready to scale. Ecosystem relationships take years to build; they cannot be procured when they are suddenly needed.
-
Influence the enterprise AI governance agenda. GCC leaders are often closer to implementation reality than the home-market teams drafting AI policy. Use that proximity. Participate actively in the enterprise's AI governance process, not just as an implementing body but as a voice shaping the framework. The governance structures that will govern the AI COE are being designed right now, in many enterprises. GCCs that engage early will get frameworks that work operationally, rather than frameworks that were designed in a boardroom and have to be worked around.
The cumulative effect of these moves is significant. A GCC that has clean data, an AI-literate workforce, a handful of proven pilots, a current understanding of the vendor landscape, and resolved data sovereignty questions is not just waiting for the AI COE moment — it is actively creating it. When the enterprise is ready, this GCC will be able to move with speed and confidence. The one that waited passively will spend the first eighteen months of its AI COE journey doing the foundational work it should have done earlier.
The Honest Forecast
The AI COE inside the GCC is not a mirage. It will become real, for many enterprises, it is genuinely a matter of when, not if. India's talent depth, cost structure, time zone complementarity, and growing AI ecosystem make it a logical home for enterprise AI capability over time.
But "over time" is doing a lot of work in that sentence. For most enterprises, the path runs through data readiness, governance maturity, geopolitical clarity, vendor landscape stabilisation, leadership fluency, and sustained organisational change, not through a headline announcement and a budget allocation.
The world is not standing still while enterprises sort this out. Geopolitical fault lines are shifting. Platform providers are expanding their surface area. Regulatory frameworks are taking shape. The AI COE that gets built on a foundation of genuine readiness will look very different, and perform very differently, from one announced to satisfy a board, that read a compelling analyst report or made splashy headlines.
The CXOs who will look smart in five years are not the ones who stood up the biggest AI COE the fastest. They are the ones who built the foundations that allowed their AI investments to actually land, and who used the intervening time wisely rather than waiting for permission to act.
That is a less exciting story to tell at an industry conference. But it is the right one. And in a market full of noise, it is the story that practitioners who have seen these cycles play out will recognise as true.