The End of the Transformation Program: Why AI Demands a New Operating Discipline

In the last three decades, every major enterprise technology wave has followed a simple pattern: it came with a beginning and an end. When enterprise resource planning (ERP) arrived, it had go-live dates. When we migrated to the cloud, we celebrated the completion of the "lift and shift." With digital transformation, it always came with a program charter and an eventual closing report. In short, the change was episodic.

This episodic change had a structural scaffolding, such as governance, which was also episodic. As leaders, we reflexively reach out for structures that are rational and have worked so far in terms of delivering ROI, satisfying auditors, and providing a sense of completion and closure that has historically served us well. But with AI, this organizational memory is now a liability. Looking at telecom operators across the world, I see everyone experimenting with AI, yet most do not see a measurable impact on their earnings. And no one seems clear on what the destination looks like, to answer, “Are we there yet?”

We are on the verge of realizing that we may never "complete" an AI transformation journey.

Why AI Differs Structurally

The framing around AI is that it is much faster and bigger than previous tech waves. The structural difference runs deeper. AI has no settled application layer.

Cloud, ERP and digital transformation eventually produced stable functional categories like general ledgers, compute instances, e-commerce checkout flows, to govern.¹ But AI has not settled, to be able to produce these categories yet. The length of tasks AI can reliably complete has doubled approximately every seven months since 2019. Since 2024, that interval has been compressed to every three or four months.² Any transformation scoped today is already operating on obsolete assumptions by its midpoint. Hence for technology leaders, unlike prior waves, the "go-live" never becomes a destination but a moving reference point that invalidates the preceding plan before it even completes.

Disaggregating the Enterprise AI Layers

The question that emerges now is around how we govern something that never settles.

To start with, we need to stop treating enterprise AI deployment as a monolithic application. In reality, this is an integrated stack with four distinct layers, each operating on a different cycle time. Governance failure occurs when an organization applies a rigid compliance process to a sublayer that changes in real time. It is important to understand them individually to govern them effectively:

Infrastructure layer forms the foundational compute and networking backbone. Decisions made here have long cycle times and high switching costs, requiring significant lead time for capacity planning, hardware provisioning, and regional deployment.

Platform layer brings together intelligence (foundation models, fine-tuning capabilities), data (data lakes, retrieval pipelines, vector databases), and orchestration (workflow engines, agent frameworks, observability). This layer combines durable data infrastructure with fast-evolving model and orchestration capabilities, requiring deliberate decoupling to balance the need for stability (on the data side) and continuous integration (on the model/orchestration interface side).

Service layer encompasses agents, workflows, and application logic that translates AI capabilities into business functions. This layer has the highest rate of change, with new patterns emerging every few months, demanding rapid prototyping and agile deployment cycles.

Customer segments and industry layer sit on top, where AI solutions are tailored to specific verticals, use cases, and client needs. This layer needs deep domain expertise, cross-layer reach and continuous calibration as regulatory requirements and customer expectations shift.

A single steering committee with one mandate is not enough to govern these layers with different cycle times. The architecture for governance needs to mirror the architecture of the problem.

Navigating the Sovereign Complexity

Now add sovereignty to the picture. When faced with this complexity, the natural instinct is to simplify by choosing a frontier model provider and rolling it out globally. But this instinct collides with geopolitical reality. Governments are increasingly signaling a preference for local AI capabilities , particularly when it comes to national infrastructure like telecom networks. Outright mandates remain rare, but there is growing regulatory pressure. Telecom operators can meet this demand but will need to navigate a shifting landscape of incentives, data residency requirements, and national security considerations.

By 2028, 60% of multinational firms are likely to split AI stacks across sovereign zones, with integration costs potentially tripling.³ For global enterprises, this may even mean running regionally differentiated stacks, each evolving on its own regulatory timeline.

This makes any global AI transformation program structurally incoherent as a singular initiative.

Addressing the Governance Mismatch

Current enterprise AI governance structures have largely been inherited from prior technology programs or assembled reactively. In fact, 70% of Fortune 500 executives report having AI risk committees, and only 41% have a dedicated AI governance team.⁴ Operational reality consistently lags board-level intent. This in my view, points to a design flaw.

The consequences of which are clear from the fact that AI use cases in production doubled between 2024 and 2025, yet only one in four initiatives met revenue impact expectations. At an average spend of $1.3 million per use case, this is a capital allocation problem.⁵

Traditional project budgeting relies on a defined scope and duration. These conditions do not apply to AI, wherein the capability frontier moves faster than the project timeline.

Steering committees designed for finite programs typically ask, "Is this on track?" "Will it complete on time?" The assumption being there is a destination. The right questions are different and more nuanced. Is our capacity to absorb change keeping pace with the capability frontier? Are our data foundations ahead of our model ambitions? What did the last model generation reveal that our current process design cannot absorb?

Moreover, ROI per initiative cannot capture systemic compounding value. AI-rich enterprises are pulling away from peers, and initiative-level ROI measurement misses this divergence. 88% of organizations use AI, yet only 6% report meaningful enterprise-wide financial impact.⁶ This is a governance gap that is widening.

Building AI Absorption Capacity

If governance is the problem, what can be the solution? This depends on who is able to build organizational capacity to absorb each successive capability shift faster than their competitors. Organizations that are able to invest ahead of model capability in data foundations can fine-tune smaller, domain-specific models on proprietary data.

This is the type of advantage that can only accumulate through discipline over time. In financial services, it's transaction data. In manufacturing, it's sensor data. In telecom, it is the network itself; call detail records, tower telemetry, customer interaction logs. These assets grow more valuable with each new model generation.

Success here will require more than a fundamental shift. It will comprise:

• Restructuring the investment category: AI needs a hybrid budget that funds both current deployment and next-generation readiness. But this cannot happen without a CEO mandate as a prerequisite for success. CEOs will need to set the tone from the top, establishing that AI investment is a strategic priority that spans fiscal years and functional boundaries. Without this mandate, CFOs demanding project-level ROI will consistently underfund the data foundations that determine long-term competitiveness.
• Redesigning governance for four cycles: A senior, cross-functional body must disaggregate its mandate; governing data on a multi-year horizon, processes on a model-release cadence, organizational structure as responsive to change, and commercial models on a near-real-time basis.
• Developing organizational absorption capacity as the primary constraint: Technology moves faster than the organization can absorb it. The rate limiter is the ability to redesign processes, retrain people, and update models in response to each shift.

The Discipline With No Finish Line

If transformation is permanent, the mandate is to build and run the organizational infrastructure that continuously absorbs capability shifts.

Organizations need to stop making the structural error of treating AI as one more episode. Over the next decade, the decisive advantage will be with those that can build the capacity to absorb change across all four layers and convert each new capability shift faster than their competitors. While most enterprises today can describe their AI program, only a few show how they have built the operating capacity to keep evolving with this technology. Organizations must also not fall into the trap of just measuring AI adoption metrics or using conventional productivity metrics to measure ROI. As our CEO Mohit Joshi says, they should be asking what AI makes newly possible. He also emphasizes that AI must be a Board agenda, recognizing it as a fundamental driver of risk management, competitiveness, and long-term value creation.

Those organizations pulling ahead are not waiting for AI to settle. Instead, they have already started with data, redesigned the processes they can influence, and built the organizational muscle to keep up with change. They are moving from AI adoption to AI advantage.

We already know which side of that divide we want to be on. The real test is whether we have the discipline to govern for a race that has no finish line, and the much-needed urgency to start building that capacity now.