Volume 4: An Exclusive Feature on Why AI Needs a New Operating Model

Most enterprise AI fails because the architecture is wrong, not the technology. Discover the new operating model built around outcomes, not pilots.

Most enterprise AI fails because the architecture is wrong, not the technology. Discover the new operating model built around outcomes, not pilots.
May 20, 2026
Ritesh Idnani
Chief Executive Officer & Managing Director, Firstsource
Download now
Most enterprise AI fails because the architecture is wrong, not the technology. Discover the new operating model built around outcomes, not pilots.

Around 300 BCE, the rulers of Alexandria made a decision that would define their civilization's ambition. They would collect every book ever written. Ships arriving at the harbor were boarded, and scrolls were confiscated and copied. Scholars were summoned from across the known world. At its peak, the Library of Alexandria held up to 700,000 scrolls: Aristotle, Euclid, Hippocrates, Archimedes. The sum of human knowledge, gathered under one roof.

And then it faded. Not in a single dramatic fire, as legend prefers. But slowly, through neglect, through declining patronage, through the gradual failure of a model that was magnificent at collecting intelligence and had no mechanism to operate it. The Library of Alexandria was the greatest repository the ancient world had ever assembled. It just had no way to make that knowledge move.

I have spent 25 years in an industry built on a version of the same idea. For much of that time, it worked. But what I see happening across enterprise AI today is the same architectural mistake at enormous scale and speed. Not a failure of ambition. A failure of distribution.

The world we're operating in

The forces reshaping business today are unlike anything we have navigated before. Geopolitical tensions are redrawing supply chains. Five generations are in the workforce simultaneously. Economic nationalism is pulling global operating models toward local accountability. And AI is advancing at a pace that makes last year's capabilities look primitive. What makes this moment unprecedented is not a single force. It is the convergence of all of them, with no sequential relief.

And underneath all of it, something specific is happening in the industry I have spent 25 years building in. The model is dying. Not failing gradually. Dying.

The business services industry was built across three distinct eras. The first was traditional outsourcing, reliability delivered by specialist providers. The second was labor arbitrage: skilled talent in lower-cost locations executing complex processes at a fraction of domestic cost. This era created enormous value. It built entire economies. But it was built on assumptions that AI is now dismantling. It assumed human labor was the primary scarce resource. It assumed processes were stable enough to define, document, and delegate. It assumed value lived in the SOP, the training manual, and the quality scorecard.

AI does not just automate those processes. It dissolves the assumptions underneath them. The question shifts from "where are your people?" to "how smart is your system, and does it get smarter over time?"

We are now in the third era: the intelligence era. The established categories that defined the previous two eras were BPO, IT services, analytics, and consulting; categories built for a different world. They describe how work is delivered. Not what outcomes clients receive. Not who is accountable when the intelligence doesn't perform.

We saw this coming at Firstsource. It is why we launched UnBPO™, a deliberate act of breaking the frame we had operated inside for two decades. The clients who would win the next decade were not looking for cheaper execution. They were looking for smarter outcomes.

The repository trap

This is where the Alexandria parallel becomes urgent.

Goldman Sachs estimates global AI spending could reach $1 trillion by 2030.¹ Worldwide AI spending will total $2.5 trillion in 2026, according to Gartner.² And yet, the same analysts note that most of this investment is being made by organizations fitting intelligence into existing architecture rather than building architecture designed for intelligence.

MIT research tells us that 95% of enterprise AI pilots fail to deliver measurable financial returns.³ The share of companies abandoning AI initiatives more than doubled in a single year. Boards are mandating strategies. Investment is at record levels. And most of it is not working.

Not because the technology is inadequate, but because the architecture is wrong. Organizations are building, at enormous expense, a digital Library of Alexandria. And like those scrolls, the intelligence sits, waiting to be consulted rather than designed to run.

The problem is not AI capability. It is AI continuity. The gap between what AI can do and what enterprises are realizing in value does not close with better models. It closes with a different operating model entirely.

A new category

Intelligence That Operates is not a product or a service line. It is a new category: one that makes the conventional labels obsolete.

Consulting firms will design your transformation. Software companies will sell you the tools. AI-native startups will build impressive proofs of concept. None of them will hold all three because holding all three requires something none of them have built: domain intelligence accumulated over decades, a full-stack delivery model designed around it, and the institutional commitment to put outcomes on the line.

Intelligence That Operates means a partner who designs the transformation, implements the architecture, and manages it in production, with full accountability for what it delivers. Not hours. Not milestones. Outcomes. One continuous motion.

This is what defines the category. Five principles, each of which the old model cannot satisfy:

Domain intelligence: Not generic AI bolted onto legacy processes, but intelligence trained on the specific contexts, decisions, and exceptions of each industry. The healthcare revenue cycle operates differently from financial services collections, which operate differently from insurance claims. The intelligence must know the difference: not just the process, but the intent behind it, and what happens when reality diverges from the design.

Full-stack delivery: Design, implement, and manage: not three separate engagements handed off between vendors, but one continuous motion. Every handoff is where accountability disappears. Every reset is where intelligence stops compounding. The partner who owns the full arc owns the outcome.

Compounding intelligence: AI models commoditize; every differentiated capability becomes a baseline expectation within 18 months. The model is not the moat. What does not commoditize is domain intelligence that gets sharper with every engagement, every decision, every outcome.

Outcome accountability: If we cannot underwrite the outcome, we have not earned the right to manage it. That is not a differentiator. It is the standard the entire industry needs to be held to. The metrics that matter are not cost-per-transaction or handle time. They are claims denial rates, customer lifetime value, and regulatory audit outcomes. The numbers that actually move the P&L.

Governed autonomy: Trust in AI is not declared. It is earned through transparency, auditability, and proven performance at every stage. Governed autonomy means the intelligence operates with increasing independence as that trust compounds, and never beyond the boundaries that have been earned. Autonomy without governance is risk. Governance without autonomy is just automation.

What this looks like in practice

These capabilities are already running in production across multiple industries like healthcare, financial services, and communications, delivering outcomes that the old model could not underwrite, let alone measure.

Twenty-five years ago, Firstsource began building operational expertise in the world's most complex, regulated industries. That accumulated intelligence is now encoded into our agentic operating architecture, not assembled in a deployment cycle or purchased off a shelf. We have built it over two and a half decades of learning how these industries work, and not how they are designed to work.

The engine, not the repository

The Library of Alexandria's failure was architectural. The assumption was that if you built the greatest repository of knowledge the world had ever seen, the world would come to it and be changed.

Enterprise AI is at precisely this inflection point. The industry model that served business services for 30 years faces the same reckoning. Traditional outsourcing delivered reliability. Labor arbitrage delivered scale. The intelligence era demands something neither could provide: intelligence that runs, learns, compounds, and is accountable for what it delivers.

At Firstsource, we broke the frame with UnBPO™. We moved from provocation to declaration. Intelligence That Operates is not an evolution of the old model. It is a replacement for it.

The Library of Alexandria had all the knowledge in the world. What it lacked was the engine to make that knowledge run.

The intelligence era will not wait for organizations to be ready. Build the engine. Or become the archive.

References:

Download UnBPOTM Quarterly