Abstract
Enterprises are accelerating investment in AI agents but struggling to convert adoption into sustained business value. Drawing on cross-industry leadership insights, this paper highlights three critical challenges - value pool targeting gaps, context leakage, and governance immaturity - in realizing the value from AI. Organizations that address these effectively can unlock scalable, trust-driven AI value and translate fragmented innovation into enterprise-wide impact.
1. The AI value gap is not a technology problem
Enterprises have entered a new phase of AI maturity. The question is no longer whether to deploy AI agents - it is why so few organizations are realizing the value they were promised. Across industries, a familiar pattern is emerging. Pilots are launched with enthusiasm. Initial results show potential. But when organizations attempt to scale, the trajectory flattens. Use cases that performed well in controlled environments fail to generalize. Agents that scored well in testing underperform in production. Adoption stalls. ROI remains elusive. Executives who championed AI programs are left explaining to boards why the returns have not materialized.
This is the AI value gap - and it is widening. EXL's research with senior leaders across insurance, healthcare, banking, manufacturing, and professional services points to three interlocking strategic themes:
(i) organizations don't know where value actually lives;
(ii) don't transfer their institutional knowledge to agents effectively; and
(iii) don't build the governance frameworks that earn the trust needed to operate at scale.
The critical insight
The AI value gap is not primarily a technology failure. Foundation models are powerful. Agent frameworks are maturing. What is missing is the organizational infrastructure to deploy them well. Organizations that address all three strategic imperatives consistently outperform those that treat them as isolated pain points.
2. Value pool targeting
Most enterprise AI programs suffer from a fundamental strategic ambiguity: they know AI is important, but they're unclear about precisely where it will generate returns. This ambiguity produces two predictable failure modes. Together, these keep AI investment trapped at the level of aspiration:
Failure mode 1: The politics portfolio: Without a structured methodology for mapping value pools, use cases emerge reactively, nominated ad hoc by whoever has the loudest voice in the room. The result is an AI portfolio that reflects organizational politics, not strategic logic.
Failure mode 2: The unfundable business case: When organizations do move forward, they consistently fail to build credible ROI narratives. Benefits are diffuse, timelines uncertain, and interaction effects between agents and human workflows are hard to model. Investment proposals stall in the approval queue, even when the underlying opportunity is real.
The hard truth for CXOs
Your competitors aren't winning with better AI models. They're winning because they know which problems are worth solving with AI, and they've built the methodology to find them systematically. Value architecture precedes agent architecture every time.
The CXO imperative: Build a value architecture before an agent architecture
Leading organizations treat value pool targeting as a living enterprise capability, not a one-time strategy exercise. This requires three interlocking practices:
Leading organizations treat value pool targeting as a living enterprise capability, not a one-time strategy exercise. This requires three interlocking practices:
CXOs who institutionalize this practice will consistently find and capture the highest-return opportunities and systematically outpace competitors who rely on reactive, ad hoc approaches to AI investment
EXL practitioner insights
Prioritization of and value realization from AI Use cases for a global insurance group: The client was struggling with where to begin its AI journey, lacking clarity on which use cases would deliver meaningful business impact at scale. EXL started the engagement with the establishment of an AI CoE designed to systematically assess and prioritize use cases across the insurance value chain - mapping AI leverage across claims, distribution, and risk functions through the lens of cost-benefit analysis, strategic impact, and implementation feasibility. Rather than pursuing AI opportunistically, we built a four-phase value architecture moving from foundational capability setup and proof-of-concept validation through scaled production deployment. Over 30 AI use cases were evaluated, with 22 prioritized for production. The discipline of structured sequencing paid compound dividends: use cases built on shared data and model foundations compressed build times and amplified returns across the portfolio. The outcomes were concrete - 14x higher detection of issues in claim files and customer interactions, and more than 50% reduction in identified leakage and non-compliance.
3. Containing enterprise context leakage
There is a paradox at the heart of enterprise AI development. The organizations most capable of building technically sophisticated agents are often the least effective at making them contextually intelligent. The processing power is there. The institutional knowledge that would make it genuinely valuable is not - because no one captured it before it was needed.
The leakage begins with how knowledge transfers from human experts to AI systems. The dominant approach - asking subject matter experts to articulate what they do - consistently underperforms. The reason is structural: the most valuable organizational knowledge is what experts have internalized so deeply they can no longer fully articulate it. The judgment calls, the exception-handling instincts, the contextual adaptations seasoned practitioners make without thinking - these are precisely the competencies that determine whether an agent performs like a senior expert or a well-briefed novice.
The leakage deepens during workflow design, where unclear process boundaries and disengaged business stakeholders leave agents operating on incomplete maps of the work they are meant to automate. Fragmented tool landscapes compound the problem further: when data pipelines are unreliable, agents operate on corrupted contextual signals, and their outputs reflect that degradation.
What is context leakage?
Enterprise Context Leakage is the systematic erosion of the tacit, nuanced organizational knowledge that distinguishes a truly intelligent agent from a sophisticated script. It is silent, invisible during development - and devastating in production.
The hard truth for CXOs
An AI agent is only as intelligent as the knowledge it was built on. If your organization hasn't treated knowledge engineering as a strategic discipline with dedicated resources, accountability, and career paths, your agents are running on institutional amnesia.
The CXO imperative: Treat knowledge engineering as a strategic asset
The solution is not better technology. It is better knowledge engineering. Leading organizations are shifting from reliance on expert articulation alone to richer methods: process mining to surface actual patterns at scale; structured observation methodologies to capture tacit judgment; capability mapping to define agent personas that reflect full human roles rather than narrow subsets of formalized tasks; and co-design workshops that embed business process owners in agent development from day one - not as reviewers of a finished product, but as co-architects.
CXOs must create organizational conditions for deep, sustained collaboration between process owners and AI development teams. Business stakeholder engagement is not optional. It is the primary mechanism through which context leakage is stopped.
EXL practitioner insights
Legal billing compliance for a professional information services firm: EXL’s engagement with the client illustrated precisely how context leakage manifests, and how disciplined knowledge engineering stops it. The client’s team manually reviewed invoice billing guidelines in PDF format and converted them into standardized rule sets, a process dependent on expert judgment, vulnerable to inconsistent interpretation, and fundamentally unscalable. EXL deployed an agentic AI solution that systematically encoded the decision logic of experienced compliance professionals: an OCR agent extracted content from source documents, a guidelines agent classified compliance rules and exceptions, and a conversational agent enabled practitioners to modify thresholds interactively. The tacit knowledge that previously lived only in expert heads was captured, structured, and transferred into agent behavior, achieving 90% accuracy in extracting guideline rules and a 60% reduction in turnaround time.
4. Governance as a competitive moat
Of all the forces slowing AI value throughput, governance is the most frequently mischaracterized. It is described as a bottleneck, a compliance tax, or as a brake on innovation. In reality, governance is the mechanism by which AI agents earn the right to operate at scale and organizations that invest in it strategically will outpace those that treat it as a checkbox exercise.
The evidence is consistent: Security, compliance, and governance reviews are among the most significant sources of deployment delay. In regulated industries - banking, insurance, healthcare - the threshold for trust is high and non-negotiable. Agents that cannot demonstrate explainability, auditability, and regulatory alignment will not be deployed in production, regardless of technical performance.
The human dimension is equally neglected. Most organizations have no structured program for reskilling or redeploying employees whose roles are affected by AI agents. This is not only a talent risk - it is a trust risk. When employees experience AI as a threat rather than a partner, adoption falters, usage is resisted, and the ROI never materializes. Responsible AI and successful AI are not competing objectives. They are the same objectives.
The invisible risk
A large proportion of organizations have no systematic approach to tracking model drift or performance degradation. Issues surface reactively - often weeks after agent outputs have declined in quality. In high-stakes domains, this reactive posture is not merely inefficient. It is a material risk to outcomes, regulatory standing, and organizational reputation.
The hard truth for CXOs
Governance isn't slowing your AI program. Your competitors' lack of governance is slowing theirs, creating an opening for you. Organizations with mature trust architectures can deploy agents faster, in more sensitive contexts, with greater stakeholder confidence, than those still retrofitting trust onto systems built without it.
The CXO imperative: Build a trust architecture, not a compliance checklist
A trust architecture operates across three reinforcing layers.
A trust architecture operates across three reinforcing layers.
EXL practitioner insights
AI governance framework for the global finance function of a Fortune 100 life sciences enterprise: Facing an expanding AI agent portfolio operating within SOX-controlled environments - where auditability, explainability, and regulatory alignment are non-negotiable - the client’s finance leadership worked with EXL to establish a dedicated governance body whose mandate explicitly connected AI deployment to measurable business value. EXL co-designed the governance architecture as a phased framework beginning with standardized intake and prioritization methodology, moving through SOX control deployment and risk classification of all deployed agents, and culminating in third-party audit readiness and scaled adoption.
Every AI agent was required to satisfy naming conventions, risk classification criteria, and role-based access control standards before production approval - all captured in catalog serving as the authoritative record of every deployed agent, its permissions, and its entitlements. Performance was tracked through a balanced scorecard measuring financial ROI realization, SOX control exceptions, compliance rates, and portfolio growth simultaneously, making the business case for governance visible in the same terms as the business case for AI itself. The result was a trust architecture that transformed governance from a sequential gate into a parallel capability, built alongside agent development, not after it.
5. Five actions that define AI leaders
The three themes converge on five concrete leadership actions. These are the moves that separate organizations accelerating AI value from those still explaining the gap to their boards.
The three themes converge on five concrete leadership actions. These are the moves that separate organizations accelerating AI value from those still explaining the gap to their boards.
6. The next decade belongs to throughput, not deployment
The first chapter of enterprise AI was about deployment, proving that agents could be built and delivered into production. That chapter is closing. The defining question of the next chapter is throughput: the ability to move from AI investment to AI value with consistency, speed, and scale across the full complexity of enterprise operations. The organizations that win this chapter will match their technical ambition with strategic precision, organizational discipline, and trust architectures that allow AI to operate responsibly in the highest-value contexts. They will know where value lives. They will preserve and encode their institutional intelligence. They will govern AI not as a constraint on progress but as the foundation on which sustainable progress is built.
The bottom line
The bottlenecks in AI value throughput are real, and they are solvable. But they require leadership decisions, not technology decisions. The CXOs who make them now will define who leads the next decade of enterprise transformation.
Written by:
Sumit Taneja
SVP & Global Head of AI Consulting & Implementation
Saurabh Mittal
VP & Head of AI & Digital Engineering
Anurag Gupta
SAVP & Lead - AI Center of Excellence