HaxiTAG: Yueli KGM

This article anchors itself in MIT’s The GenAI Divide: State of AI in Business 2025 and integrates HaxiTAG’s public discourse and product practices (EiKM, ESG Tank, Yueli Knowledge Computation Engine, etc.). It systematically dissects the core insights and methodological implementation pathways for AI and generative AI in enterprise applications, providing actionable guidance and risk management frameworks. The discussion emphasizes professional clarity and authority. For full reports or HaxiTAG’s white papers on generative AI applications, contact HaxiTAG.

Introduction

The most direct—and potentially dangerous—lesson for businesses from the MIT report is: widespread GenAI adoption does not equal business transformation. About 95% of enterprise-level GenAI pilots fail to generate measurable P&L impact. This is not primarily due to model capability or compliance issues, but because enterprises have yet to solve the systemic challenge of enabling AI to “remember, learn, and integrate into business processes” (the learning gap).

Key viewpoints and data insights in the research report: MIT's NANDA's 26-page "2025 State of Business AI" covers more than 300 public AI programs, 52 interviews, and surveys of 153 senior leaders from four industry conferences to track adoption and impact.

- 80% of companies "surveyed" "general LLMs" (such as ChatGPT, Copilot), but only 40% of companies "successfully implemented" (in production).

- 60% "surveyed" customized "specific task AI," 20% conducted pilots, and only 5% reached production levels, partly due to workflow integration challenges.

- 40% purchased official LLM subscriptions, but 90% of employees said they used personal AI tools at work, fostering "shadow AI."

- 50% of AI spending was on sales and marketing, although backend programs typically generate higher return on investment (e.g., through eliminating BPO).

External partnerships "purchasing external tools, co-developed with suppliers" outperformed "building internal tools" by a factor of 2.

HaxiTAG has repeatedly emphasized the same point in enterprise AI discussions: organizations need to shift focus from pure “model capability” to knowledge engineering + operational workflows + feedback loops. Through EiKM enterprise knowledge management and dedicated knowledge computation engine design, AI evolves from a mere tool into a learnable, memorizable collaborative entity.

Key Propositions and Data from the MIT Report

High proportion of pilots fail to translate into productivity: Many POCs or demos remain in the sandbox; real-world deployment is rare. Only about 5% of enterprise GenAI projects yield sustained revenue or cost improvements. 95% produce no measurable P&L impact.
The “learning gap” is critical: AI repeatedly fails in enterprise workflows because systems cannot memorize organizational preferences, convert human review into iterative model data, or continuously improve across multi-step business processes.
Build vs. Buy watershed: Projects co-built or purchased with trusted external partners, accountable for business outcomes (rather than model benchmarks), have success rates roughly twice that of internal-only initiatives. Successful implementations require deep customization, workflow embedding, and iterative feedback, significantly improving outcomes.
Back-office “silent gold mines”: Financial, procurement, compliance, and document processing workflows yield faster, measurable ROI compared to front-office marketing/sales, which may appear impactful but are harder to monetize quickly.

Deep Analysis of MIT Findings and Enterprise AI Practice

The Gap from Pilot to Production

Assessment → Pilot → Production drops sharply: Embedded or task-specific enterprise AI tools have a ~5% success rate in real deployment. Many projects stall at the POC stage, failing to enter the “sustained value zone” of workflows.

Enterprise paradox: Large enterprises pilot the most aggressively and allocate the most resources but lag in scaling success. Mid-sized enterprises, conversely, often achieve full deployment from pilot within ~90 days.

Typical Failure Patterns

“LLM Wrappers / Scientific Projects”: Flashy but disconnected from daily operations, fragile workflows, lacking domain-specific context. Users often remark: “Looks good in demos, but impractical in use.”
Heavy reconfiguration, integration challenges, low adaptability: Require extensive enterprise-level customization; integration with internal systems is costly and brittle, lacking “learn-as-you-go” resilience.
Learning gap impact: Even if frontline employees use ChatGPT frequently, they abandon AI in critical workflows because it cannot remember organizational preferences, requires repeated context input, and does not learn from edits or feedback.
Resource misallocation: Budgets skew heavily to front-office (sales/marketing ~50–70%) because results are easier to articulate. Back-office functions, though less visible, often generate higher ROI, resulting in misdirected investments.

The Dual Nature of the “Learning Gap”: Technical and Organizational

Technical aspect: Many deployments treat LLMs as “prompt-to-generation” black boxes, lacking long-term memory layers, attribution mechanisms, or the ability to turn human corrections into training/explicit rules. Consequently, models behave the same way in repeated contexts, limiting cumulative efficiency gains.

Organizational aspect: Companies often lack a responsibility chain linking AI output to business KPIs (who is accountable for results, who channels review data back to the model). Insufficient change management leads to frontline abandonment. HaxiTAG emphasizes that EiKM’s core is not “bigger models” but the ability to structure knowledge and embed it into workflows.

Empirical “Top Barriers to Failure”

User and executive scoring highlights resistance as the top barrier, followed by concerns about model output quality and poor UX. Underlying all these is the structural problem of AI not learning, not remembering, not fitting workflows.
Failure is not due to AI being “too weak” but due to the learning gap.

Why Buying Often Beats Building

External vendors typically deliver service-oriented business capabilities, not just capability frameworks. When buyers pay for business outcomes (BPO ratios, cost reduction, cycle acceleration), vendors are more likely to assume integration and operational responsibility, moving projects from POC to production. MIT’s data aligns with HaxiTAG’s service model.

HaxiTAG’s Solution Logic

HaxiTAG’s enterprise solution can be abstracted into four core capabilities: Knowledge Construction (KGM) → Task Orchestration → Memory & Feedback (Enterprise Memory) → Governance/Audit (AIGov). These align closely with MIT’s recommendation to address the learning gap.

Knowledge Construction (EiKM): Convert unstructured documents, rules, and contracts into searchable, computable knowledge units, forming the enterprise ontology and template library, reducing contextual burden in each query or prompt.

Task Orchestration (HaxiTAG BotFactory): Decompose multi-step workflows into collaborative agents, enabling tool invocation, fallback, exception handling, and cross-validation, thus achieving combined “model + rules + tools” execution within business processes.

Memory & Feedback Loop: Transform human corrections, approval traces, and final decisions into structured training signals (or explicit rules) for continuous optimization in business context.

Governance & Observability: Versioned prompts, decision trails, SLA metrics, and audit logs ensure secure, accountable usage. HaxiTAG stresses that governance is foundational to trust and scalable deployment.

Practical Implementation Steps (HaxiTAG’s Guide)

For PMs, PMO, CTOs, or business leaders, the following steps operationalize theory into practice:

Discovery: Map workflows by value stream; prioritize 2 “high-frequency, rule-based, quantifiable” back-office scenarios (e.g., invoice review, contract pre-screening, first-response service tickets). Generate baseline metrics (cycle time, labor cost, outsourcing expense).
Define Outcomes: Translate KRs into measurable business results (e.g., “invoice cycle reduction ≥50%,” “BPO spend down 20%”) and specify data standards.
Choose Implementation Path: Prefer “Buy + Deep Customize” with trusted vendors for MVPs; if internal capabilities exist and engineering cost is acceptable, consider Build.
Rapid POC: Conduct “narrow and deep” POCs with low-code integration, human review, and metric monitoring. Define A/B groups (AI workflow vs. non-AI). Aim for proof of business value within 6–8 weeks.
Embed Learning Loop: Collect review corrections into data streams (tagged) and [enable small-batch fine-tuning, prompt iteration, or rule enhancement for explicit business evolution].
Governance & Compliance (parallel): Establish audit logs, sensitive information policies, SLAs, and fallback mechanisms before launch to ensure oversight and intervention capacity.
KPI Integration & Accountability: Incorporate POC metrics into departmental KPIs/OKRs (automation rate, accuracy, BPO savings, adoption rate), designating a specific “AI owner” role.
Replication & Platformization (ongoing): Abstract successful solutions into reusable components (knowledge ontology, API adapters, agent templates, evaluation scripts) to reduce repetition costs and create organizational capability.

Example Metrics (Quantifying Implementation)

Efficiency: Cycle time reduction n%, per capita throughput n%.
Quality: AI-human agreement ≥90–95% (sample audits).
Cost: Outsourcing/BPO expenditure reduction %, unit task cost reduction (¥/task).
Adoption: Key role monthly active ≥60–80%, frontline NPS ≥4/5.
Governance: Audit trail completion 100%, compliance alert closure ≤24h.

Baseline and measurement standards should be defined at POC stage to avoid project failure due to vague results.

Potential Constraints and Practical Limitations

Incomplete data and knowledge assets: Without structured historical approvals, decisions, or templates, AI cannot learn automatically. See HaxiTAG data assetization practices.
Legacy systems & integration costs: Low API coverage of ERP/CRM slows implementation and inflates costs; external data interface solutions can accelerate validation.
Organizational acceptance & change risk: Frontline resistance due to fear of replacement; training and cultural programs are essential to foster engagement in co-intelligence evolution.
Compliance & privacy boundaries: Cross-border data and sensitive clauses require strict governance, impacting model availability and training data.
Vendor lock-in risk: As “learning agents” accumulate enterprise memory, switching costs rise; contracts should clarify data portability and migration mechanisms.

Three Recommendations for Enterprise Decision-Makers

From “Model” to “Memory”: Invest in building enterprise memory and feedback loops rather than chasing the latest LLMs.
Buy services based on business outcomes: Shift procurement from software licensing to outcome-based services/co-development, incorporating SLOs/KRs in contracts.
Back-office first, then front-office: Prioritize measurable ROI in finance, procurement, and compliance. Replicate successful models cross-departmentally thereafter.

Digital Maturity Assessment

Before embarking on AI adoption, companies must conduct a comprehensive evaluation of their digital maturity to accurately locate both challenges and opportunities. AI maturity models provide a strategic roadmap for procurement leaders by assessing the current state of technological infrastructure, team capabilities, and process digitalization. These insights help define a realistic evolution path based on gaps and readiness.

McKinsey recommends a dual-track approach—rapidly deploying AI and analytics use cases that generate quick wins, while simultaneously building a scalable data platform to support long-term needs. Similarly, DNV’s AI maturity framework emphasizes benchmarking organizational vision against industry standards to help companies set priorities from a holistic perspective and avoid becoming isolated “technology islands.”

Technology: Buy or Build?

One of the most strategic decisions in implementing AI is choosing between purchasing ready-made solutions or building custom systems. Off-the-shelf solutions offer faster time-to-value, mature interfaces, and lower technical entry barriers—but they often fall short in addressing the unique nuances of procurement functions.

Conversely, organizations with greater AI ambitions may opt to build proprietary systems to achieve deeper control over spend transparency, contract optimization, and ESG goal alignment. However, this approach demands significant in-house capabilities in data engineering and algorithm development, along with careful consideration of long-term maintenance costs versus strategic benefits.

Forbes emphasizes that AI success hinges not only on the technology itself but also on factors such as user trust, ease of adoption, and alignment with long-term strategy—key dimensions that are frequently overlooked in the build-vs-buy debate. Additionally, the initial cost and future iteration expenses of AI solutions must be factored into decision-making to prevent unmanageable ROI gaps later on.

Upskilling the Team

AI doesn't just accelerate existing procurement processes—it redefines them. As such, upskilling procurement teams is paramount. According to BCG, only 10% of AI’s value comes from algorithms, 20% from data and platforms, and a staggering 70% from people adapting to new ways of working and being motivated to learn.

Economist Impact reports that 64% of enterprises have already adopted AI tools in procurement. This transformation requires current employees to gain proficiency in data analytics and decision support, while also bringing in new roles such as data scientists and AI engineers. Leaders must foster a culture of experimentation and continuous learning through robust change management and transparent communication to ensure skill development is fully realized.

The Hackett Group further notes that the most critical future skills for procurement professionals include advanced analytics, risk assessment, and cross-functional collaboration. These competencies will empower teams to excel in complex negotiations and supplier management. Supply Chain Management Review highlights that AI also democratizes learning for budget-constrained companies, enabling them to adopt and refine new technologies through hands-on experience.

Capturing Value from Suppliers

The ultimate goal of AI adoption in procurement is to translate technical capabilities into measurable business value—generating negotiation insights through advanced analytics, optimizing contract terms, and even encouraging suppliers to adopt generative AI to reduce total supply chain costs.

BCG’s research shows that a successful AI transformation can yield cost savings of 15% to 45% across select categories of products and services. The key lies in seamlessly integrating AI into procurement workflows and delivering an exceptional initial user experience to drive ongoing adoption and scalability. Sustained value capture also depends on strong executive commitment, regular KPI evaluation, and active promotion of success stories—ensuring that AI transformation becomes an enduring engine of enterprise growth.

Conclusion

In today’s hypercompetitive market landscape, AI-driven procurement transformation is no longer optional—it is essential. It offers a vital pathway to securing future competitive advantages and building core capabilities. At Hashitag, we are committed to guiding procurement teams through every stage of the transformation journey, from maturity assessment and technology decisions to workforce enablement and continuous value realization. We hope this four-step framework provides a clear roadmap for organizations to unlock the full potential of intelligent procurement and thrive in the digital era.

Menu

HaxiTAG

Your Trusted Partner for Intelligent Transformation and AI Industry Solutions

Get GenAI guide

Monday, October 6, 2025

From “Can Generate” to “Can Learn”: Insights, Analysis, and Implementation Pathways for Enterprise GenAI