AI Agents Tự Hành: Làn Sóng Tự Động Hóa Kế Tiếp Đang Thay Đổi Doanh Nghiệp Việt Nam và Đông Nam Á — Từ Hứa Hẹn Đến Hiện Thực Nghiệt Ngã

The enterprise love affair with Generative AI (GenAI) is hitting a rocky patch. While nearly nine out of ten organizations now employ AI in some capacity, a staggering 70-85% of these initiatives fail to deliver their expected value. The honeymoon phase of pilots and proofs-of-concept is giving way to the harsh realities of integration, elusive ROI, and profound organizational friction. In 2026, the real challenge isn't just adopting AI tools, but strategically scaling them to achieve tangible, measurable business transformation beyond superficial efficiency gains.

Why This Matters Now: The Enterprise's Generative AI Dilemma

The year 2026 was supposed to be the inflection point for enterprise Generative AI. Following an explosion of interest in 2023 and 2024, and a flurry of pilot projects in 2025, the market anticipated widespread, transformative adoption. Indeed, worldwide AI spending is projected to reach an astounding $2.59 trillion in 2026, a 47% increase over 2025 according to Gartner. IDC forecasts AI infrastructure spending alone to hit $497 billion in 2026, a 56% year-over-year growth. Worker access to sanctioned AI tools has expanded by 50% in just one year, now reaching approximately 60% of the workforce. McKinsey's "State of AI in 2026" survey reveals that 44% of organizations are now scaling AI across their enterprise, up from 38% a year ago. Large enterprises, with annual revenues exceeding $1 billion, are leading this charge, with 40% scaling AI agents, a significant jump from 27% in the previous year.

Yet, beneath these impressive figures lies a sobering truth: the vast majority of enterprise AI initiatives are not delivering on their promise. Multiple independent research organizations—Gartner, MIT, RAND Corporation, BCG, and McKinsey—published AI project failure studies in 2025-2026, all converging on the conclusion that 70-85% of enterprise AI initiatives fail to deliver their expected value. MIT's Project NANDA, focusing specifically on generative AI, found that 95% of GenAI pilots delivered zero measurable profit and loss (P&L) return, with only about 5% capturing value at scale. This isn't merely a measurement error; it's a systemic problem indicating a profound disconnect between ambition and activation.

The current landscape is characterized by a "two-speed race" in AI adoption, where large firms accelerate, while smaller organizations lag. This disparity is not just about resources; it's about the fundamental challenges of moving beyond isolated experiments to embedding AI as a core operational capability. The shift from experimenting with generative AI to deploying more complex agentic AI systems, predicted to be embedded in 33% of enterprise software applications by 2028 (up from less than 1% in 2024), amplifies these challenges. The imperative for enterprises in 2026 is clear: understand the unseen hurdles that turn promising pilots into expensive failures, and forge a clear path to genuine, profitable scale.

The Illusion of Seamless Integration: Unpacking the Technical Debt of AI Tools

The promise of plug-and-play AI tools often collides with the gritty reality of enterprise IT. Integrating novel Generative AI capabilities into existing, often decades-old, legacy systems is proving to be a formidable barrier. Legacy systems frequently lack modern APIs, real-time data access, clean documentation, and scalable infrastructure, making it arduous to connect AI models with the very systems where business work happens. Enterprises that attempt to overlay AI without addressing this foundational technical debt face significant friction and escalating costs. According to IBM, data readiness has become one of the largest barriers to enterprise AI adoption.

Data Quality and Governance: The Unseen Foundation

AI models are only as good as the data they are trained on, and enterprise data is rarely "AI-ready." Poor data quality and fragmentation are consistently cited as major challenges. Gartner predicts that 60% of AI projects lacking AI-ready data will be abandoned through 2026. Indeed, 42% of U.S. enterprises have already abandoned at least one AI project due to data issues, with problems stemming from insufficient volume (33%), poor quality (38%), and accessibility (29%). Without robust data governance – identifying existing data assets, establishing access protocols, and ensuring data accuracy, consistency, and security – enterprises risk inaccurate recommendations, compliance issues, security exposure, and low trust in AI-driven decisions.

Security, Compliance, and the Expanding Attack Surface

The rapid adoption of GenAI tools has expanded the enterprise attack surface in ways traditional cybersecurity controls were not designed to manage. Sensitive data leakage through AI prompts is a pervasive risk, as employees unintentionally share confidential information, source code, or intellectual property with shadow AI tools not approved to process such data. Shadow AI, the AI-era evolution of shadow IT, bypasses crucial controls, creating visibility gaps and making insider risk harder to detect. Other critical security risks include prompt injection attacks, malicious AI-generated code, hallucinations leading to security failures, and supply chain vulnerabilities.

Regulatory and compliance exposure is also a core component of GenAI security risks. Existing privacy regulations like GDPR, CCPA, and the EU AI Act amplify the need to apply consistent privacy principles across new AI-driven data flows. PwC highlights that GenAI applications exacerbate data and privacy risks, as models use and create massive amounts of data vulnerable to bias, poor quality, unauthorized access, and loss. Employees entering sensitive data into public generative AI models is a significant problem, potentially contravening privacy regulations that restrict secondary uses of personal data.

The Hidden Costs of AI at Scale

Beyond initial development, the ongoing costs of enterprise GenAI are substantial and often underestimated. Building a custom, multi-system platform with fine-tuned models and agentic workflows can range from $500,000 to over $1 million. However, the "real cost" extends far beyond the build. First-year run costs can approximate the build cost, with a three-year total often 2-3 times the initial investment, due to inference, governance, evaluation, and maintenance. Inference costs, in particular, can be a budget killer at enterprise volume, with monthly API bills running into thousands of dollars per active workload. Other operational costs include data pipeline development, specialized hardware (e.g., GPUs), software licensing fees, and continuous monitoring and optimization strategies. IDC projects AI infrastructure spending to surpass $1 trillion in 2029, reaching $1.08 trillion, before climbing to $1.21 trillion in 2030, at a compound annual growth rate of approximately 30% from 2025, indicating that AI capacity is becoming a structural cost of doing business at scale.

Beyond the Hype: The Elusive ROI of Enterprise Generative AI

Despite the immense investment and widespread adoption, quantifiable financial returns from enterprise AI remain stubbornly elusive for many. McKinsey's 2026 State of AI survey found that only 37% of organizations report any positive EBIT contribution from AI, a figure essentially flat compared to the previous year. Furthermore, only 6% of organizations qualify as "AI high performers," attributing 5% or more of EBIT to AI, a number that also remained flat year-over-year despite continued spending increases. This indicates a widening gap where spending accelerates faster than results materialize.

While 80% of survey respondents report that AI has improved their individual productivity, these gains have not translated into measurable organizational financial benefit at anywhere near the same rate. This "AI productivity paradox" highlights that simply layering AI tools onto existing processes is insufficient; true value creation requires fundamentally redesigning workflows around AI. Companies are focused on building AI fluency, but often neglect the more challenging task of redesigning work itself.

The shift to agentic AI, where AI systems make decisions, coordinate tasks, and complete multi-step workflows with limited human involvement, presents new challenges for ROI quantification. PwC predicts that measurable ROI quantification will become standard practice in 2026, with qualitative claims about 'efficiency gains' losing credibility in favor of specific metric improvements. However, the complexity of agentic workflows means that proving direct financial impact requires rigorous tracking and attribution, which many organizations are still ill-equipped to perform.

The Talent Gap and Organizational Friction: People, Process, and the Pace of Change

The challenges of enterprise AI adoption are not solely technical; they are deeply rooted in organizational readiness, talent shortages, and human resistance to change. IBM identifies skills gaps and organizational change as critical challenges for AI adoption in 2026.

The Scarcity of AI-Ready Talent

The demand for skilled AI professionals far outstrips supply. Enterprises struggle to find and retain AI specialists, data scientists, machine learning engineers, and prompt engineers who command premium salaries. This talent gap extends to implementation teams, project managers, and even legal and compliance experts who understand both the technology and its complex implications. PwC's 2026 predictions highlight the "Rise of the AI Generalist," indicating a future workforce where AI-literate individuals capable of supervising agents, interpreting outputs, and ensuring alignment with business objectives will be in high demand. However, organizations are often behind in developing this new AI-native workforce.

Resistance to Change and Leadership Alignment

AI adoption is not just a technological shift; it's a profound organizational transformation that often meets internal resistance. Employees may fear job displacement, distrust model outputs, or lack adequate training. A 2026 survey found that 54% of C-suite executives admit that adopting AI is "tearing their company apart," a double-digit increase from 2025. Nearly two-thirds (64%) of executives fear losing their job if they fail to lead the AI transition, yet three-quarters (75%) admit their company's AI strategy is "more for show" than actual internal guidance. This disconnect between leadership rhetoric and tangible action fosters skepticism and hinders genuine adoption.

Many companies make the mistake of adopting a ground-up, crowdsourced approach to AI initiatives, which rarely leads to transformation. Instead, PwC suggests that successful organizations in 2026 are shifting to top-down, leadership-driven programs that concentrate resources on a small number of high-value workflows. Without clear ownership, a disciplined strategy, and effective change management, AI projects are destined to remain isolated wins rather than enterprise-wide successes.

From Lab to Enterprise: Examining GenAI in Action – The Klarna Case Study

The journey from pilot to profitable scale is fraught with unforeseen challenges, as exemplified by Klarna, the Swedish buy-now-pay-later fintech. In early 2024, Klarna made headlines for aggressively deploying an AI assistant, built with OpenAI, to handle customer service interactions. The company proudly announced that the AI agent performed the work of 700 human agents in its first month, processing 2.3 million conversations and cutting resolution time from an average of 11 minutes to under 2. This move was projected to save Klarna $40 million in 2024 and was hailed as a benchmark for AI-driven white-collar job displacement. CEO Sebastian Siemiatkowski publicly declared that "AI can already do all of the jobs we humans do".

The Bright Side: Initial Efficiency and Cost Savings

• Rapid Resolution: The AI assistant significantly reduced average resolution times, from 11 minutes to under 2 minutes.
• High Volume Handling: In its first month, the AI processed 2.3 million conversations, handling two-thirds of all customer service inquiries across 23 markets and 35+ languages.
• Projected Cost Savings: Klarna anticipated $40 million in profit improvement for 2024 due to reduced headcount and increased efficiency.
• Reduced Repeat Inquiries: The AI achieved 25% fewer repeat inquiries, suggesting effective first-contact resolution for routine issues.

The Dark Side: Quality Erosion and the Human Element

However, by early 2026, the narrative shifted dramatically. Klarna was quietly reversing course and began rehiring human customer service staff. The initial dashboard metrics, which focused on speed and volume, failed to capture a critical dimension: quality and customer satisfaction.

• Declining Customer Satisfaction: Customer satisfaction scores deteriorated, particularly for complex service interactions, emotionally charged issues, and multi-step problem resolution. Customers complained about robotic, inflexible, and generic responses, leading to frustration and operational hiccups.
• Unrealized Cost Savings: The projected cost savings did not fully materialize because handling the quality issues and subsequent customer churn consumed more than was saved. Research shows that acquiring a new customer can cost up to seven times more than retaining an existing one.
• Rehiring Whiplash: Reversing the layoffs incurred significant, often unbudgeted, costs related to recruiting, onboarding, and training new customer service staff. CEO Sebastian Siemiatkowski admitted that Klarna "went too far," acknowledging that an overemphasis on efficiency and cost ultimately reduced service quality and eroded customer trust.
• Limitations in Nuance and Empathy: The Klarna case starkly highlighted the limitations of full AI automation in roles requiring human judgment, empathy, and relationship management. The company is now piloting a blended "Uber-style" workforce model, combining human agents with AI to leverage the strengths of both.

The Klarna reversal has become a canonical cautionary tale for enterprises in 2026, underscoring that while AI excels at routine tasks, the human element remains indispensable for complex, nuanced, and empathetic customer interactions.

What Nobody Is Saying: The Uncomfortable Truths of Enterprise GenAI in 2026

Amidst the relentless optimism surrounding AI, several critical, less-talked-about truths are shaping the enterprise landscape in 2026. These are the aspects that demand skepticism and a clear-eyed assessment, lest organizations fall prey to the next wave of tech hype.

The Illusion of AI Model Agnosticism and Deep Vendor Lock-in

For years, enterprises have strived for vendor agnosticism to avoid costly lock-in. However, Generative AI fundamentally challenges this paradigm. As Gartner flags, vendor lock-in is a "blind spot" most CIOs miss until it becomes expensive. Enterprises, eager to move fast, often pick a single vendor for speed and simplicity, inadvertently tying their data, models, and workflows to proprietary APIs and platform tools. Unlike previous technology waves where interoperability standards allowed for easier switching, there's almost no such thing as an LLM-agnostic application. Building an application where one model can be easily swapped for another without significant re-work is a "sisyphean effort". This deep integration means any transition involves substantial redevelopment, not just a simple plug-and-play.

The problem is compounded by the accumulation of "AI memory"—the prompts, corrections, workflows, and institutional knowledge fed into a vendor's system. This context, which becomes an organization's competitive advantage, is often trapped in the vendor's database, making switching not just difficult but potentially sacrificing years of accumulated intelligence. The default state of enterprise AI in 2026 is often a fragmented landscape of multiple vendor lock-ins, one per department, lacking an overarching layer for governance and visibility.

The Environmental Footprint of Large-Scale AI

While often overshadowed by efficiency narratives, the environmental cost of large language models (LLMs) is a growing concern. Training and running these models require massive amounts of computing power, primarily from energy-intensive data centers. The carbon footprint of a single large LLM can be equivalent to that of several cars over their lifetime, or even an entire building's energy consumption for months. As enterprises scale their GenAI deployments, the cumulative energy demand will only increase. This often goes unmentioned in ROI calculations, yet it represents an unsustainable externality that will eventually need to be accounted for, either through increased regulation or public pressure. The shift to AI infrastructure as a structural cost of doing business at scale implicitly means a structural increase in energy consumption, a point rarely highlighted in optimistic adoption forecasts.

The Looming AI Disillusionment and the "Strategy for Show"

The current phase of enterprise AI adoption, characterized by high investment and low measurable ROI, is setting the stage for a significant "AI disillusionment" phase. Many C-suite executives, facing pressure to adopt AI, admit their company's AI strategy is "more for show" than actual internal guidance. Nearly half (48%) call AI adoption a "massive disappointment," up from 34% last year. This performative adoption, driven by fear of missing out rather than a clear business case, leads to scattered pilot projects without clear ownership or integration into core business processes.

When leadership prioritizes impressive adoption numbers over meaningful business outcomes, the result is a proliferation of tools without systems, pilots without production, and enthusiasm without architecture. The Klarna case is a stark reminder that even seemingly successful initial metrics can mask deeper quality issues and ultimately lead to costly reversals. The industry needs to brace for a period where the initial hype cycle gives way to a more pragmatic, critical assessment of AI's true value proposition, potentially leading to a market correction in AI spending and a greater emphasis on demonstrable, long-term impact.

Charting a Path Forward: Strategic Recommendations for Enterprise AI Adoption

To navigate the chasm between ambitious pilots and profitable enterprise scale, organizations must adopt a strategic, evidence-based approach, prioritizing prudence over hype.

The Immediate Future: Consolidation and Specialization

The AI tools market will likely see significant consolidation and specialization. Niche AI tools, hybrid models combining the best of AI and human intelligence, and an increased emphasis on explainable AI will gain traction. Enterprises should prepare for a future where:

• Focus on Agentic Workflows: Agentic AI, capable of automating complex, high-value workflows beyond basic analysis, will move from pilot to production at scale. However, robust governance and human oversight will be critical.
• Data as a Strategic Asset: Data quality, readiness, and robust governance will become non-negotiable prerequisites for any successful AI initiative. Organizations must treat data modernization as a strategic initiative, not just a technical task.
• Hybrid Architectures: A blend of cloud-based AI services, on-premise deployments, and specialized edge AI will emerge to balance cost, performance, and data sovereignty requirements.
• Ethical AI by Design: Responsible AI will move from discussion to measurable traction in operations, with a focus on fairness, transparency, and accountability embedded from the outset.

Actionable Insights for Leaders

For CIOs, CTOs, and business leaders grappling with AI adoption in 2026, here are concrete recommendations:

1. Define Value, Not Just Velocity: Shift focus from simply deploying AI tools or running pilots to quantifying measurable business outcomes. PwC predicts that in 2026, companies will stop measuring AI progress by the number of tools deployed and start measuring it by the number of high-value problems actually solved. Establish clear, specific success metrics (e.g., EBIT impact, customer retention, measurable quality improvements) *before* investment, not after.
2. Prioritize Data Foundations: Before investing heavily in advanced models, invest in cleaning, unifying, and governing your enterprise data. AI-ready data is the bedrock of successful deployments. Implement strong data classification, access controls, and visibility for all data interacting with AI systems.
3. Invest in People and Process Redesign: Recognize that AI success is 80% about redesigning work and only 20% about technology. Prioritize talent development, upskilling existing employees, and implementing robust change management programs to foster AI literacy and trust. Develop a clear, top-down AI strategy that aligns with core business objectives, rather than a fragmented, crowdsourced approach.
4. Mitigate Vendor Lock-in Actively: Demand open standards, open APIs, and modular architecture in every GenAI pilot. Focus on building internal self-reliance by owning your unique use cases, business logic, and internal data assets, rather than solely relying on vendor-specific model implementations. Implement an enterprise-wide layer to authenticate agents, manage permissions, and log actions, providing centralized governance across multiple tools and vendors.
5. Embrace a Hybrid Human-AI Model: Learn from the Klarna case. For customer-facing or high-stakes internal processes, a hybrid human-AI model consistently outperforms full automation. Leverage AI for routine tasks and augmentation, freeing human agents for complex, empathetic, and judgment-intensive interactions.
6. Build a Robust AI Governance Framework: Implement comprehensive policies for security, privacy, compliance, and ethical AI from day one. This includes monitoring dashboards, error-routing pathways, regular reviews, and clear accountability for AI outputs. Over-reliance on AI for critical business decisions without human review can lead to dangerous dependencies and compliance problems.
7. Question the Hype: Maintain a healthy skepticism towards aggressive projections and PR-driven announcements. Demand verifiable production references, not just pilot decks. Understand the hidden costs, long timelines, and high failure rates that are often omitted from vendor pitches.

The Road Ahead: Prudence Over Hype

The year 2026 marks a critical juncture for enterprise AI. The initial wave of experimentation and hype is receding, revealing the complex, often unglamorous work required to translate AI's potential into sustained business value. The data is unequivocal: a significant majority of AI projects falter, not due to technological shortcomings, but due to organizational, data, and strategic missteps.

The organizations that will truly thrive in this new era are not those chasing every shiny new tool, but those that approach AI with a disciplined strategy, a clear understanding of its limitations, and a steadfast commitment to building robust data foundations, fostering human-AI collaboration, and establishing rigorous governance. The path to profitable enterprise AI scale is not a sprint, but a marathon demanding strategic foresight, operational excellence, and a healthy dose of skepticism. The future of AI in the enterprise belongs to the pragmatic, not the prone-to-hype.

Sources & References

• Deloitte US, "The State of AI in the Enterprise - 2026 AI report," August 2025.
• Swift Headway AI, "PwC 2026 AI Predictions — SMB Owner Action Plan," May 21, 2026.
• IBM, "The Biggest AI Adoption Challenges for 2026," February 14, 2025 (updated May 29, 2026).
• Ivanti, "5 Generative AI Security Risks Compliance Teams Need to Know," July 27, 2026.
• Cyberhaven, "Top Generative AI Security Risks in the Enterprise," March 25, 2026.
• Deloitte Canada, "The State of AI in the Enterprise - 2026 AI report," August 2025.
• Welcome.AI, "McKinsey Report Highlights Growing Divide in Enterprise AI Adoption," September 3, 2026.
• PwC, "PwC's 2026 AI Business Predictions: Business Strategies & Trends," February 26, 2026.
• 80% of AI Projects Fail: The Hidden Cause, "80% of AI Projects Fail: The Hidden Cause," August 1, 2026.
• DataTribes, "PwC's 2026 AI Business Predictions - DataTribes," June 29, 2026.
• Coworker AI, "Why Enterprise AI Fails: The Real Failure Rate and Root Causes (2026)," June 30, 2026.
• Connected Paths, "AI Project Failure Statistics - everything you need in 2026," September 1, 2026.
• PwC, "2026 AI Business Predictions."
• "Vendor lock-in is the blind spot most CIOs are missing in the AI strategy," August 13, 2026.
• InData Labs, "The cost of generative AI and strategies for managing it," May 7, 2026.
• PwC, "AI and the future of work," December 9, 2025.
• McKinsey, "The State of AI: Global Survey 2026," August 25, 2026.
• IDC, "IDC's Worldwide AI and Generative AI Spending Guide Taxonomy, 2026: Release V2, 2026."
• A10 Networks, "Top 9 Generative AI Security Risks in 2026," March 18, 2026.
• Digital Applied, "Klarna Reverses AI Layoffs: Why Replacing 700 Failed," March 9, 2026.
• Deloitte, "The State of AI in the Enterprise," January 21, 2026.
• Talyx, "Why 90% of Enterprise AI Implementations Fail (2026)," January 27, 2026.
• Deloitte AI Institute, "From Ambition to Activation: Organizations Stand at the Untapped Edge of AI's Potential, Reveals Deloitte Survey," January 21, 2026.
• "Klarna CEO admits aggressive AI job cuts went too far, starts hiring again after US IPO," October 10, 2025.
• Tommaso Maria Ricci, "Enterprise AI Adoption 2026: 88% Use AI, 39% See ROI," August 26, 2026.
• Value Add VC, "McKinsey: Enterprise AI ROI Still Lags Adoption Hype," August 25, 2026.
• Digital Applied, "AI Adoption Statistics 2026: Business & Enterprise Data," August 1, 2026.
• Deloitte UK, "The State of AI in the Enterprise - 2026 AI report," August 2025.
• InData Labs, "Getting ROI From Your GenAI: A Look at The High Cost of Generative AI for Enterprises," September 2, 2024.
• Nomtek, "What Enterprise AI Actually Costs and How Long It Takes to Reach Production," May 28, 2026.
• Writer, "Navigating the challenges of generative AI and "vendor lock-in" in enterprises," December 4, 2024.
• XTrace, "AI Vendor Lock-In: Who Really Owns Your Competitive Advantage?," February 9, 2026.
• Pertama Partners, "AI Project Failure Rate 2026: 80% Fail," February 8, 2026 (updated June 15, 2026).
• SAP LeanIX, "AI Vendor Lock-In: Building Your House On Sand," March 21, 2024.
• Proofpoint US, "Generative AI Security Risks: What Enterprises Need to Know," December 29, 2025.
• Truefoundry, "The Hidden Costs of GenAI and How to Control Them," March 25, 2026.
• Forbes, "Klarna Reverses AI Push, Says Customers Prefer Human Support," May 18, 2025.
• AlphaBOLD, "Cost Analysis: Implementing Generative AI in Your Organization," April 29, 2025.
• Reworked, "Klarna Claimed AI Was Doing the Work of 700 People. Now It's Rehiring," May 19, 2025.
• IDC, "AI Infrastructure Spending Holds Near $90 Billion in Q1 2026 as ARM Overtakes x86 in Accelerated Servers," July 21, 2026.
• AsisteClick, "Klarna laid off 700 humans for AI. Now it's rehiring.," May 2, 2026.
• IDC, "Worldwide AI and Generative AI Spending Guide."
• ISHIR, "Enterprise AI Adoption Challenges in 2026 | Pitfalls, Strategies and Solutions."
• Digital Applied, "AI Spending Forecasts 2026: Gartner, IDC & Stanford," June 12, 2026.
• "Enterprise AI adoption in 2026: Why 79% face challenges despite high investment," April 7, 2026.
• Appinventiv, "11+ key AI adoption challenges for enterprises to resolve," June 11, 2026.
• PwC, "Managing the risks of generative AI."
• NineTwoThree, "AI Adoption Challenges: What a $20M Company Really Needs Before AI," August 26, 2026.
• MegaRouter, "How Can Enterprise AI Reduce Vendor Lock-In? From Single-Model Dependence to Multi-Model Architecture," August 28, 2026.
• IDC, "Spending Archives," April 13, 2026.