HaxiTAG Case Investigation & Analysis: How an AI Decision System Redraws Retail Banking’s Cognitive Boundary

Structural Stress and Cognitive Bottlenecks in Finance

Before 2025, retail banking lived through a period of “surface expansion, structural contraction.” Global retail banking revenues grew at ~7% CAGR since 2019, yet profits were eroded by rising marketing, compliance, and IT technical debt; North America even saw pre-tax margin deterioration. Meanwhile, interest-margin cyclicality, heightened deposit sensitivity, and fading branch touchpoints pushed many workflows into a regime of “slow, fragmented, costly.” Insights synthesized from the Retail Banking Report 2025.

Management teams increasingly recognized that “digitization” had plateaued at process automation without reshaping decision architecture. Confronted by decision latency, unstructured information, regulatory load, and talent bottlenecks, most institutions stalled at slogans that never reached the P&L. Only ~5% of companies reported value at scale from AI; ~60% saw none—evidence of a widening cognitive stratification. For HaxiTAG, this is the external benchmark: an industry in structural divergence, urgently needing a new cost logic and a higher-order cognition.

When Organizational Mechanics Can’t Absorb Rising Information Density

Banks’ internal retrospection began with a systematic diagnosis of “structural insufficiencies” as complexity compounded:

• Cognitive fragmentation: data scattered across lending, risk, service, channels, and product; humans still the primary integrators.

• Decision latency: underwriting, fraud control, and budget allocation hinging on batched cycles—not real-time models.

• Rigid cost structure: compliance and IT swelling the cost base; cost-to-income ratios stuck above 60% versus ~35% at well-run digital banks.

• Cultural conservatism: “pilot–demo–pause” loops; middle-management drag as a recurring theme.

In this context, process tweaks and channel digitization are no longer sufficient. The binding constraint is not the application layer; the cognitive structure itself needs rebuilding.

AI and Intelligent Decision Systems as the “Spinal Technology”

The turning point emerged in 2024–2025. Fintech pressure amplified through a rate-cut cycle, while AI agents—“digital labor” that can observe, plan, and act—offered a discontinuity.

Agents already account for ~17% of total AI value in 2025, with ~29% expected by 2028 across industries, shifting AI from passive advice to active operators in enterprise systems. The point is not mere automation but:

• Value-chain refactoring: from reactive servicing to proactive financial planning;

• Shorter chains: underwriting, risk, collections, and service shift from serial, multi-team handoffs to agent-parallelized execution;

• Real-time cadence: risk, pricing, and capital allocation move to millisecond horizons.

For HaxiTAG, this aligns with product logic: AI ceases to be a tool and becomes the neural substrate of the firm.

Organizational Intelligent Reconstruction: From “Process Digitization” to “Cognitive Automation”

1) Customer: From Static Journeys to Live Orchestration

AI-first banks stop “selling products” and instead provide a dynamic financial operating system: personalized rates, real-time mortgage refis, automated cash-flow optimization, and embedded, interface-less payments. Agents’ continuous sensing and instant action confer a “private CFO” to every user.

2) Risk: From Batch Control to Continuous Control

Expect continuous-learning scoring, real-time repricing, exposure management, and automated evidence assembly with auditable model chains—shifting risk from “after-the-fact inspection” to “always-on guardianship.”

3) Operations: Toward Near-Zero Marginal Cost

An Asian bank using agent-led collections and negotiation cut costs 30–40% and lifted cure rates by double digits; virtual assistants raised pre-application completion by ~75% without harming experience. In an AI-first setup:

• ~80% of back-office flows can run agent-driven;

• Mid/back-office roles pivot to high-value judgment and exception handling;

• Orgs shrink in headcount but expand in orchestration capacity.

4) Tech & Governance: A Three-Layer Autonomy Framework

Leaders converge on three layers:

1. Agent Policy Layer — explicit “can/cannot” boundaries;

2. Assurance Layer — audit, simulation, bias detection;

3. Human Responsibility Layer — named owners per autonomous domain.

This is how AI-first banking meets supervisory expectations and earns customer trust.

Performance Uplift: Converting Cognitive Dividends into Financial Results

Modeled outcomes indicate 30–40% lower cost bases for AI-first banks versus baseline by 2030, translating to >30% incremental profit versus non-AI trajectories, even after reinvestment and pricing spillbacks. Leaders then reinvest gains, compounding advantage; by 2028 they expect 3–7× higher value capture than laggards, sustained by a flywheel of “investment → return → reinvestment.”

Concrete levers:

• Front-office productivity (+): dynamic pricing and personalization lift ROI; pre-approval and completion rates surge (~75%).

• Mid/back-office cost (–): 30–50% reductions via automated compliance/risk, structured evidence chains.

• Cycle-time compression: 50–80% faster across lending, onboarding, collections, AML/KYC as workflows turn agentic.

On the macro context, BAU revenue growth slows to 2–4% (2024–2029) and 2025 savings revenues fell ~35% YoY, intensifying the necessity of AI-driven step-changes rather than incrementalism.

Governance and Reflection: The Balance of Smart Finance

Technology does not automatically yield trust. AI-first banks must build transparent, regulator-ready guardrails across fairness, explainability, auditability, and privacy (AML/KYC, credit pricing), while addressing customer psychology and the division of labor between staff and agents. Leaders are turning risk & compliance from a brake into a differentiator, institutionalizing Responsible AI and raising the bar on resilience and audit trails.

Appendix: AI Application Utility at a Glance

Application Scenario	AI Capability Used	Practical Utility	Quantified Effect	Strategic Significance
Example 1	NLP + Semantic Search	Automated knowledge extraction; faster issue resolution	Decision cycle shortened by 35%	Lowers operational friction; boosts CX
Example 2	Risk Forecasting + Graph Neural Nets	Dynamic credit-risk detection; adaptive pricing	2-week earlier early-warning	Strengthens asset quality & capital efficiency
Example 3	Agent-Based Collections	Automated negotiation & installment planning	Cost down 30–40%	Major back-office cost compression
Example 4	Dynamic Marketing Optimization	Agent-led audience segmentation & offer testing	Campaign ROI +20–40%	Precision growth and revenue lift
Example 5	AML/KYC Agents	Automated evidence chains; orchestrated case-building	Review time –70%	Higher compliance resilience & auditability

The Essence of the Leap: Rewriting Organizational Cognition

The true inflection is not the arrival of a technology but a deliberate rewriting of organizational cognition. AI-first banks are no longer mere information processors; they become cognition shapers—institutions that reason in real time, decide dynamically, and operate through autonomous agents within accountable guardrails.

For HaxiTAG, the implication is unequivocal: the frontier of competition is not asset size or channel breadth, but how fast, how transparent, and how trustworthy a firm can build its cognition system. AI will continue to evolve; whether the organization keeps pace will determine who wins. 

Related Topic

Generative AI: Leading the Disruptive Force of the Future
HaxiTAG EiKM: The Revolutionary Platform for Enterprise Intelligent Knowledge Management and Search
From Technology to Value: The Innovative Journey of HaxiTAG Studio AI
HaxiTAG: Enhancing Enterprise Productivity with Intelligent Knowledge Management Solutions
HaxiTAG Studio: AI-Driven Future Prediction Tool
A Case Study:Innovation and Optimization of AI in Training Workflows
HaxiTAG Studio: The Intelligent Solution Revolutionizing Enterprise Automation
Exploring How People Use Generative AI and Its Applications
HaxiTAG Studio: Empowering SMEs with Industry-Specific AI Solutions
Maximizing Productivity and Insight with HaxiTAG EIKM System

Read more →

Upgrading Personal Global Asset Allocation in the Age of AI

An asset allocation brief from HSBC Singapore looks, on the surface, like just another routine “monthly outlook”: maintain an overweight to the US but trim it slightly, increase exposure to Asia and gold, prefer investment-grade bonds over high-yield bonds, and emphasize that “AI adoption and local consumption are the twin engines for Asia’s growth.” ([HSBC China][1])

Yet for an ordinary high-net-worth individual investor, what this letter really exposes is another layer of reality: global asset pricing is increasingly being simultaneously reshaped by three forces—AI investment, regional growth divergence, and central bank policy. Under such complexity, the traditional personal investing style of “experience + hearsay” can hardly support rational, stable, and reviewable decisions.

This article focuses on a single question: In an era of AI-driven global asset repricing, how can individuals use AI tools to rebuild their capability for global asset allocation?

---

From Institutional Perspective to Individual Dilemma: Key Challenges of Asset Allocation in the AI Era

The Macro Narrative: AI and the Dual Reshaping of “Geography + Industry”

According to HSBC’s latest global investment outlook, US equities remain rated “overweight” thanks to the AI investment boom, expanding tech earnings, and fiscal support. However, due to valuation and policy uncertainty, HSBC is gradually shifting part of that weight toward Asian equities, gold, and hedge funds, while on the bond side preferring investment-grade credit over high-yield bonds. ([HSBC China][1])

Beyond the US, HSBC defines Asia as a region enjoying a “twin tailwind” of AI ecosystem + local consumption:

• On one hand, Asia is expected to outperform global peers between 2025–2030 in areas such as data-center expansion, semiconductors, and compute infrastructure. ([HSBC Global Private Banking][2])

• On the other hand, resilient local consumption, supported by policy stimulus in multiple countries and ongoing corporate governance reforms, underpins expectations for improving regional return on equity (ROE). ([HSBC Bank Malaysia][3])

This is a highly structured, cross-regional asset-allocation narrative with AI as one of the core variables. The typical institutional logic can be summarized as:

“Amid the tension between the AI investment wave and regional fundamental divergences, use a multi-region, multi-asset portfolio to hedge single-market risk while sharing in the structural excess returns brought by AI.”

The Ground Reality: Four Structural Challenges Facing Individual Investors

If we “translate” this letter down to the individual level, a typical compliant investor (for example, someone working in Singapore and holding multi-regional assets) is confronted with four practical challenges:

1. Information Hierarchy Gap

   • Institutions have access to multi-regional data, research teams, industry dialogues, and quantitative tools.

   • Individual investors usually only see information that has been “compressed several times over”: marketing materials, media summaries, and fragmented social media opinions—making it hard to grasp the underlying reasoning chain.

2. Cross-Market Complexity and Asymmetric Understanding

   • The brief covers multiple regions: the US, Asia (Mainland China, Singapore, Japan, South Korea, Hong Kong), the UK, each with different currencies, rate cycles, valuation regimes, and regulatory environments.

   • For an individual, it is difficult to understand within a unified framework how “US AI equities, high-dividend Asian stocks, investment-grade USD bonds, gold, and hedge funds” interact with each other.

3. Uncertainty Within the AI Investment Narrative Itself

   • The OECD and other research bodies estimate that AI could add 0.5–3.5 percentage points per year to labor productivity over the next decade, but the range is wide and highly scenario-dependent. ([OECD][4])

   • At the same time, recent outlooks caution that AI-driven equity valuations may contain bubble risks; if sentiment reverses, the resulting correction could drag on both economies and markets. ([Axios][5])

4. Tight Coupling Between Individual Decisions and Emotions

   • Under the multi-layered narrative of “AI leaders + high valuations + global rate shifts + regional rotation,” individuals are easily swayed by short-term price moves and headline news, ending up with momentum-chasing and panic-selling instead of following a life-cycle-based strategic framework.

In short: Institutions are using AI and multi-asset models to manage portfolios, while individuals are still relying on “visual intuition, gut feel, and fragmented information” to make decisions—that is the structural capability gap we face today.

---

AI as a “Personal CIO”: Three Anchors for Upgrading Asset Allocation Capability

Against this backdrop, if individuals only see AI as a chatbot that “answers market questions,” their decision quality will hardly improve. What truly matters is embedding AI into the three critical stages of personal asset allocation: cognition, analysis, and execution.

Cognitive Upgrade: From “Listening to Conclusions” to “Reading the Originals + Cross-Checking Sources”

Institutional judgments—such as “Asia benefits from the twin tailwind of AI and local consumption” and “the US remains overweight but should gradually diversify”—are, by nature, compressed syntheses of massive underlying facts. ([HSBC China][1])

Once LLM/GenAI enters the picture, individual investors can construct a new cognitive pathway:

1. Automatically Collect Source Materials

   • Use agents to automatically fetch public information from: HSBC’s official website, central-bank statements, OECD reports, corporate earnings summaries, etc.

   • Tag and organize this content by region (US, Asia, UK) and asset class (equities, bonds, gold, hedge funds).

2. Multi-Source Reading Comprehension and Bias Detection

   • Apply long-form reading and summarization capabilities to compress each institutional view into a four-part structure: “background – logic – conclusion – risks.”

   • Compare differences across institutions (e.g., OECD, commercial banks, independent research houses) on the same topic, such as:

     • The projected range of AI’s contribution to productivity growth;

     • How they assess AI bubble risks and valuation pressures. ([OECD][6])

3. Build a “Personal Facts Baseline”

   • Let AI help classify: which points are hard facts broadly agreed upon across institutions, and which are specific to a particular institution’s stance or model assumptions.

   • On this basis, evaluate the strength of any given investment brief’s arguments instead of accepting them unquestioningly.

Analytical Upgrade: From “Vague Impressions” to “Visualized Scenarios and Stress Tests”

Institutions use multi-asset models, scenario analysis, and stress testing—individuals can build a lightweight version of these with AI:

1. Scenario Construction

   • Ask an LLM, using public data, to construct several macro scenarios:

     • Scenario A: AI investment remains strong without a bubble burst; the Fed cuts rates as expected.

     • Scenario B: AI valuations correct by 20–30%; the pace of rate cuts slows.

     • Scenario C: Asian local consumption softens, but AI-related exports stay robust.

   • For each scenario, generate directional views on regional equities, bond yields, and FX, and clearly identify the “core drivers.”

2. Parameterised Portfolio Analysis

   • Feed an individual’s existing positions into an AI-driven allocation tool (e.g., 60% US equities, 20% Asian equities, 10% bonds, 10% cash).

   • Let the system estimate portfolio drawdown ranges, volatility, and expected return levels under those scenarios, and present them via visual charts.

3. Risk Concentration Detection

   • Using RAG + LLM, reclassify holdings by industry (IT, communications, financials), theme (AI ecosystem, high dividend, cyclicals), and region (US, Asia, Europe).

   • Reveal “nominal diversification but actual concentration”—for example, when multiple funds or ETFs all hold the same set of AI leaders.

With these capabilities, individuals no longer merely oscillate between “the US feels expensive and Asia looks cheaper,” but instead see quantified scenario distributions and risk concentrations.

Execution Upgrade: From “Passive Following” to “Rule-Based + Semi-Automated Adjustments”

The institutional call to “trim US exposure and add to Asia and gold” is, in essence, a disciplined rebalancing and diversification process. ([HSBC Bank Malaysia][3])

Individuals can use AI to build their own “micro execution engine”:

1. Rules-Based Investment Policy Statement (IPS) Template

   • With AI’s assistance, draft a quantitative personal IPS, including target return bands, maximum acceptable drawdown, and tolerance ranges for regional and asset allocations.

   • For example:

     • US equities target range: 35–55%;

     • Asian equities: 20–40%;

     • Defensive assets (investment-grade bonds + gold + cash): at least 25%.

2. Threshold-Triggered Rebalancing Suggestions

   • Via broker/bank open APIs or semi-manual data import, let AI periodically check whether the portfolio has drifted outside IPS ranges.

   • When deviations exceed a threshold (e.g., US equity weight 5 percentage points above the upper bound), automatically generate a rebalancing proposal list, with estimated transaction costs and tax implications.

3. “AI as Watchtower,” Not “AI as Commander”

   • AI does not replace the final decision-maker. Instead, it is responsible for:

     • Continuously monitoring the Fed, OECD, major economies’ policies, and structural changes in the AI market;

     • Flagging risk events and rebalancing opportunities relevant to the individual’s IPS;

     • Translating complexity into “the three things you need to pay attention to this week.”

---

The Incremental Value of AI for Personal Asset Allocation: From Qualitative to Quantitative

Drawing on HSBC’s research structure and public data, we can break down AI’s contribution to personal asset-allocation capability into several measurable, comparable dimensions.

Multi-Stream Information Integration

• Traditional approach:

  • Mostly depends on a single bank/broker’s monthly reports plus headline news;

  • Individuals find it hard to understand systematically why the portfolio is overweight the US and why it is adding to Asia.

• With AI embedded:

  • Multiple institutional views (HSBC, OECD, other research institutions, etc.) can be integrated in minutes and summarized using a unified template. ([HSBC China][1])

  • The real improvement lies in “breadth × structuredness of information,” rather than simply piling up more content.

Scenario Simulation and Causal Reasoning

• Both HSBC and the OECD highlight in their outlooks that AI investment simultaneously supports productivity and earnings expectations and introduces valuation and macro-volatility risks. ([Axios][5])

• Relying on intuition alone, individuals rarely connect “AI bubble risk” with the Fed’s rate path or regional valuations.

• LLMs can help unpack, across different AI investment scenarios, which assets benefit and which come under pressure, while providing clear causal chains and indicative ranges.

Content Understanding and Knowledge Compression

• Institutional reports are often lengthy and saturated with jargon.

• AI reading and summarization can retain key numbers, assumptions, and risk flags, while compressing them into a one-page memo that individuals can actually digest—drastically reducing cognitive load.

Decision-Making and Structured Thinking

• HSBC’s research shows that enterprises adopting AI significantly outperform non-adopters in profitability and valuation, with US corporate AI adoption around 48%, nearly twice that of Europe. ([HSBC][7])

• Transposing this structured thinking into personal asset allocation, AI tools help individuals:

  • Clarify why they are adding to a specific region or sector;

  • View risk and return at the portfolio level rather than fixating on single stocks or short-term price swings.

Expression and Reviewability

• With generative AI, individuals can record the logic behind each adjustment as a short “investment memo,” including assumptions, objectives, and risk controls.

• When they look back later, they can clearly distinguish whether gains or losses were due to random market noise or flaws in their original decision framework.

---

Building a “Personal Intelligent Asset-Allocation Workflow”

Operationally, an AI-enabled personal asset-allocation process can be decomposed into five executable steps.

Step 1: Define the Personal Problem Instead of Parroting Institutional Views

• Do not start from “Should I follow HSBC and allocate more to Asia?”

• Instead, let AI help surface:

  • Sources of income, currency exposure, and job stability;

  • Cash-flow needs and risk tolerance over the next 3–10 years;

  • Existing concentration across regions, industries, and themes.

Step 2: Build a “Multi-Source Facts Base”

• Treat HSBC’s views, OECD reports, and other authoritative studies as data sources, and let AI:

  • Distill consensus—for example, “mainstream forecast ranges for AI’s impact on productivity” and “structural differences between Asia and the US in AI investment and adoption”;

  • Highlight points of contention—such as differing assessments of AI bubble risks.

Step 3: Use AI to Design Scenarios and Portfolio Templates

• Ask AI to generate two or three candidate portfolios:

  • Portfolio A: Maintain current structure with only minor rebalancing;

  • Portfolio B: Substantially increase weights in Asia and gold;

  • Portfolio C: Increase exposure to defensive assets such as investment-grade bonds and cash.

• For each portfolio, AI provides expected return ranges, volatility, and historical analogues for maximum drawdowns.

Step 4: Make Execution Rules Explicit Instead of “One-Off Gut Decisions”

• With AI’s assistance, write down clear rules for “when to rebalance, by how much, and under which conditions to pause trading” in a one-page IPS.

• At the same time, use agents to regularly check for portfolio drift; only when thresholds are breached are action suggestions triggered—reducing emotionally driven trading frequency.

Step 5: Review in Natural Language and Charts

• Each quarter, ask AI to summarize:

  • Whether portfolio performance has stayed within the expected range;

  • The three most important external factors during the period (e.g., Fed meetings, AI valuation corrections, policy changes in Asia);

  • Which decisions reflected “disciplined persistence” and which ones were “self-persuasion” that deserve reflection.

---

Example: How a Single Brief Is Reused by a “Personal AI Workbench”

Take three key signals from this HSBC brief as an example:

1. “The US remains overweight but is slightly downgraded” →

   • AI tools interpret this as “do not go all-in on US AI assets; moderate regional diversification is necessary,” and then cross-check whether other institutions share similar views.

2. “Asia benefits from the twin tailwind of AI and local consumption, overweighting China/Hong Kong, Singapore, Japan, and South Korea” →

   • AI further breaks down cross-country differences in AI ecosystems (chips, compute, applications), consumption, and governance reforms, and presents them in tables to individual investors. ([HSBC China][1])

3. “Prefer investment-grade bonds, high-dividend stocks, and gold, while de-emphasizing high-yield bonds” →

   • AI helps screen for concrete instruments in the existing product universe (such as specific Asian investment-grade bond funds or gold ETFs) and estimates their roles given the current yield and volatility environment.

Through this series of “decompose – recombine – embed into workflow” operations, what began as a mass-distributed brief is transformed into a set of asset-allocation decision inputs conditioned on personal constraints, rather than simple “market mood guidance.”

---

From Asset Allocation to Capability Uplift: The Long-Term Significance of AI for Individual Investors

At the macro level, AI is reshaping productivity, corporate earnings structures, and capital-market valuation logic. At the micro level, financial institutions are rapidly deploying generative AI models for research, risk management, and client service. ([Reuters][8])
If individual investors remain stuck at the level of “using AI only as a Q&A gadget,” they will be persistently outpaced by institutions in terms of tools and decision frameworks for asset allocation.

Yueli AI · Unified Intelligent Workbench 

**Yueli AI is a unified intelligent workbench (Yueli Deck) that brings together the world’s most advanced AI models in one place.**

It seamlessly integrates private datasets and domain-specific or role-specific knowledge bases across industries, enabling AI to operate with deeper contextual awareness. Powered by advanced **RAG-based dynamic context orchestration**, Yueli AI delivers more accurate, reliable, and trustworthy reasoning for every task.

Within a single, consistent workspace, users gain a streamlined experience across models—ranging from document understanding, knowledge retrieval, and analytical reasoning to creative workflows and business process automation.

By blending multi-model intelligence with structured organizational knowledge, **Yueli AI functions as a data-driven, continuously evolving intelligent assistant**, designed to expand the productivity frontier for both individuals and enterprises.

Related Topic

Generative AI: Leading the Disruptive Force of the Future
HaxiTAG EiKM: The Revolutionary Platform for Enterprise Intelligent Knowledge Management and Search
From Technology to Value: The Innovative Journey of HaxiTAG Studio AI
HaxiTAG: Enhancing Enterprise Productivity with Intelligent Knowledge Management Solutions
HaxiTAG Studio: AI-Driven Future Prediction Tool
Microsoft Copilot+ PC: The Ultimate Integration of LLM and GenAI for Consumer Experience, Ushering in a New Era of AI
In-depth Analysis of Google I/O 2024: Multimodal AI and Responsible Technological Innovation Usage
Google Gemini: Advancing Intelligence in Search and Productivity Tools

Read more →

Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations

This research report, 《Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations》, authored by the Anthropic team, presents a systematic analysis of AI usage patterns in economic tasks by leveraging privacy-preserving data from millions of conversations on Claude.ai. The study aims to provide empirical insights into how AI is integrated into different occupational tasks and its impact on the labor market.

Research Background and Objectives

The rapid advancement of artificial intelligence (AI) has profound implications for the labor market. However, systematic empirical research on AI’s actual application in economic tasks remains scarce. This study introduces a novel framework that maps over four million conversations on Claude.ai to occupational categories from the U.S. Department of Labor’s O*NET database, identifying AI usage patterns and its impact on various professions. The research objectives include:

1. Measuring the scope of AI adoption in economic tasks, identifying which tasks and professions are most affected by AI.

2. Quantifying the depth of AI usage within occupations, assessing the extent of AI penetration in different job roles.

3. Evaluating AI’s application in different occupational skills, identifying the cognitive and technical skills where AI is most frequently utilized.

4. Analyzing the correlation between AI adoption, wage levels, and barriers to entry, determining whether AI usage aligns with occupational salaries and skill requirements.

5. Differentiating AI’s role in automation versus augmentation, assessing whether AI primarily functions as an automation tool or an augmentation assistant enhancing human productivity.

Key Research Findings

1. AI Usage is Predominantly Concentrated in Software Development and Writing Tasks

• The most frequently AI-assisted tasks include software engineering (e.g., software development, data science, IT services) and writing (e.g., technical writing, content editing, marketing copywriting), together accounting for nearly 50% of total AI usage.

• Approximately 36% of occupations incorporate AI for at least 25% of their tasks, indicating AI’s early-stage integration into diverse industry roles.

• Occupations requiring physical interaction (e.g., anesthesiologists, construction workers) exhibit minimal AI usage, suggesting that AI’s influence remains primarily within cognitive and text-processing domains.

2. Quantifying the Depth of AI Integration Within Occupations

• Only 4% of occupations utilize AI for over 75% of their tasks, indicating deep AI integration in select job roles.

• 36% of occupations leverage AI for at least 25% of tasks, signifying AI’s expanding role in various professional task portfolios, though full-scale adoption is still limited.

3. AI Excels in Tasks Requiring Cognitive Skills

• AI is most frequently employed for tasks that demand reading comprehension, writing, and critical thinking, while tasks requiring installation, equipment maintenance, negotiation, and management see lower AI usage.

• This pattern underscores AI’s suitability as a cognitive augmentation tool rather than a substitute for physically intensive or highly interpersonal tasks.

4. Correlation Between AI Usage, Wage Levels, and Barriers to Entry

• Wage Levels: AI adoption peaks in mid-to-high-income professions (upper quartile), such as software development and data analysis. However, very high-income (e.g., physicians) and low-income (e.g., restaurant workers) occupations exhibit lower AI usage, possibly due to:

  • High-income roles often requiring highly specialized expertise that AI cannot yet fully replace.

  • Low-income roles frequently involving significant physical tasks that are less suited for AI automation.

• Barriers to Entry: AI is most frequently used in occupations requiring a bachelor’s degree or higher (Job Zone 4), whereas occupations with the lowest (Job Zone 1) or highest (Job Zone 5) education requirements exhibit lower AI usage. This suggests that AI is particularly effective in knowledge-intensive, mid-tier skill professions.

5. AI’s Dual Role in Automation and Augmentation

• AI usage can be categorized into:

  • Automation (43%): AI directly executes tasks with minimal human intervention, such as document formatting, marketing copywriting, and code debugging.

  • Augmentation (57%): AI collaborates with users in refining outputs, optimizing code, and learning new concepts.

• The findings indicate that in most professions, AI is utilized for both automation (reducing human effort) and augmentation (enhancing productivity), reinforcing AI’s complementary role in the workforce.

Research Methodology

This study employs the Clio system (Tamkin et al., 2024) to classify and analyze Claude.ai’s vast conversation data, mapping it to O*NET’s occupational categories. The research follows these key steps:

1. Data Collection:

   • AI usage data from December 2024 to January 2025, encompassing one million interactions from both free and paid Claude.ai users.

   • Data was analyzed with strict privacy protection measures, excluding interactions from enterprise customers (API, team, or enterprise users).

2. Task Classification:

   • O*NET’s 20,000 occupational tasks serve as the foundation for mapping AI interactions.

   • A hierarchical classification model was applied to match AI interactions with occupational categories and specific tasks.

3. Skills Analysis:

   • The study mapped AI conversations to 35 occupational skills from O*NET.

   • Special attention was given to AI’s role in complex problem-solving, system analysis, technical design, and time management.

4. Automation vs. Augmentation Analysis:

   • AI interactions were classified into five collaboration modes:

     • Automation Modes: Directive execution, feedback-driven corrections.

     • Augmentation Modes: Task iteration, knowledge learning, validation.

   • Findings indicate a near 1:1 split between automation and augmentation, highlighting AI’s varied applications across different tasks.

Policy and Economic Implications

1. Comparing Predictions with Empirical Findings

• The research findings validate some prior AI impact predictions while challenging others:

  • Webb (2019) predicted AI’s most significant impact in high-income occupations; however, this study found that mid-to-high-income professions exhibit the highest AI adoption, while very high-income professions (e.g., doctors) remain less affected.

  • Eloundou et al. (2023) forecasted that 80% of occupations would see at least 10% of tasks impacted by AI. This study’s empirical data shows that approximately 57% of occupations currently use AI for at least 10% of their tasks, slightly below prior projections but aligned with expected trends.

2. AI’s Long-Term Impact on Occupations

• AI’s role in augmenting rather than replacing human work suggests that most occupations will evolve rather than disappear.

• Policy recommendations:

  • Monitor AI-driven workforce shifts to identify which occupations benefit and which face displacement risks.

  • Adapt education and workforce training programs to ensure workers develop AI collaboration skills rather than being displaced by automation.

Conclusion

This research systematically analyzes over four million Claude.ai conversations to assess AI’s integration into economic tasks, revealing:

• AI is primarily applied in software development, writing, and data analysis tasks.

• AI adoption is widespread but not universal, with 36% of occupations utilizing AI for at least 25% of tasks.

• AI usage exhibits a balanced distribution between automation (43%) and augmentation (57%).

• Mid-to-high-income occupations requiring a bachelor’s degree show the highest AI adoption, while low-income and elite specialized professions remain less affected.

As AI technologies continue to evolve, their role in the economy will keep expanding. Policymakers, businesses, and educators must proactively leverage AI’s benefits while mitigating risks, ensuring AI serves as an enabler of productivity and workforce transformation.

Related Topic

HaxiTAG Intelligent Application Middle Platform: A Technical Paradigm of AI Intelligence and Data Collaboration
RAG: A New Dimension for LLM's Knowledge Application
HaxiTAG Path to Exploring Generative AI: From Purpose to Successful Deployment
The New Era of AI-Driven Innovation
Unlocking the Power of Human-AI Collaboration: A New Paradigm for Efficiency and Growth
Large Language Models (LLMs) Driven Generative AI (GenAI): Redefining the Future of Intelligent Revolution
LLMs and GenAI in the HaxiTAG Framework: The Power of Transformation
Application Practices of LLMs and GenAI in Industry Scenarios and Personal Productivity Enhancement

Read more →