Research Background and Core Findings
The research report "Building the foundations for agentic AI at scale" published by McKinsey Technology reveals the critical challenges and opportunities facing enterprise-level AI Agent deployment. The report opens with an industry-alarming reality: nearly two-thirds of companies worldwide have experimented with AI Agents, yet fewer than 10 percent have achieved true scaled deployment generating substantial business value. This significant gap does not stem from insufficient technological maturity but rather points to a more fundamental issue—the fragility of data infrastructure. According to survey data, a striking 80 percent of companies cite data limitations as the primary obstacle constraining AI at scale.
This finding carries profound strategic implications. In recent years of AI implementation practice, companies have generally undergone an evolutionary认知 process: from initial blind worship of AI technology to gradually recognizing that data quality is the critical factor determining AI application success or failure. McKinsey's research further deepens this understanding, shifting the focus from traditional AI applications (embedding single AI capabilities into business processes) to more complex Agentic AI scenarios, emphasizing that in highly autonomous, real-time coordinated Agent work environments, the importance of data infrastructure is further amplified.
The report's core contribution lies in constructing a systematic analytical framework, including the identification of two Agent architecture patterns, the refinement of seven data architecture principles, and the design of a four-step implementation pathway. This framework not only addresses the diagnostic question of "why current Agent deployment success rates are low" but, more importantly, provides enterprises with an action guide for "how to systematically solve this challenge."
From the perspective of current industry application status, AI adoption is spreading from a few leading functions to the entire enterprise. McKinsey global survey data shows that in 2025, multiple industries have achieved routine AI usage in marketing, knowledge, and IT functions. Marketing leads AI application across nearly all surveyed industries, knowledge management ranks second, and IT also demonstrates strong adoption momentum. This distribution feature provides an empirical foundation for understanding which business scenarios are best suited for prioritizing Agent transformation.
Use Case Positioning and Technical Characteristics of Two Agent Architecture Patterns
Single Agent Workflow Pattern
The first Agent architecture pattern identified in the report is the single Agent workflow, with its core characteristic being one Agent sequentially calling multiple tools and data sources to complete complex tasks. This architecture represents an intermediate stage in AI application's evolution from "auxiliary tool" to "autonomous agent," suitable for workflow scenarios with relatively clear decision paths and definable task steps.
In a single Agent workflow, the Agent plays a highly intelligent digital assistant role. It understands user's natural language instructions, decomposes complex tasks into several sub-steps, sequentially calls corresponding tools and data sources, and progressively advances until task completion. This model's advantages manifest in several aspects: the architecture is concise and clear, facilitating debugging and troubleshooting; the causal chain is transparent, facilitating decision process tracing; and the technical implementation threshold is relatively low, making it a suitable starting choice for enterprise Agent transformation.
However, single Agent architecture also has inherent limitations. The most significant issue is that when tasks involve multiple fragmented data sources, the Agent may make contradictory decisions based on inconsistent information. The report explicitly points out this issue: single Agents can make inconsistent decisions from fragmented data. In actual enterprise operational environments, data silos are widespread—customer information may be scattered across multiple independent databases such as CRM, ERP, and customer service systems, while product data may be independently maintained by different departments. If an Agent needs to integrate data from these varying sources and standards for comprehensive judgment, single Agent's planning capability may face severe challenges.
From a use case perspective, single Agent workflows are particularly suitable for the following types of work: highly structured workflows with clear task boundaries, such as invoice processing, report generation, and data entry—repetitive tasks; scenarios requiring sequential access to multiple systems following fixed processes but with relatively simple decisions at each step, such as order fulfillment tracking and employee onboarding; and exploratory and experimental Agent application scenarios where enterprises can accumulate experience through single Agent pilots before evolving to more complex patterns.
Multi-Agent Workflow Pattern
The second Agent architecture pattern identified in the report is the multi-Agent workflow, with its core characteristic being specialized Agents collaborating through shared knowledge graphs. In this model, different Agents are endowed with different domain-specific knowledge, respectively assuming professional roles such as Finance Agent, Legal Agent, Operations Agent, and Marketing Agent—they exchange information and coordinate tasks through a common semantic layer.
The potential advantage of multi-Agent architecture lies in its ability to handle genuinely enterprise-level complex tasks, organically combining specialized depth with collaborative breadth. In a well-designed multi-Agent system, various specialized Agents can process their respective sub-tasks in parallel while coordinating progress, exchanging information, and resolving conflicts through standardized communication protocols. This architecture pattern more closely mirrors how enterprise organizations operate in reality—different functional departments each fulfilling their responsibilities and collaborating.
However, the core challenges of multi-Agent architecture lie in coordination mechanism design and error propagation control. The report provides explicit warning: multi-Agent systems could lose coordination and amplify errors. Specifically, when multiple Agents simultaneously access and modify shared data, sophisticated conflict detection and resolution mechanisms are needed; when an Agent makes an incorrect decision, preventing this error from propagating and amplifying through the Agent information chain is necessary; and when Agents have differing understandings of data semantics, ensuring they communicate effectively within a shared semantic framework is required.
Multi-Agent workflows are best suited for application scenarios including: complex decision scenarios requiring multi-domain expertise, such as market entry strategy formulation, M&A due diligence, and product portfolio optimization; workflows involving multiple business process intersection points, such as cross-departmental customer service processes and end-to-end order fulfillment processes; and dynamic scenarios requiring continuous monitoring and response to external changes, such as real-time risk monitoring and dynamic pricing management.
Synergistic Evolution of Both Architecture Patterns
The report emphasizes that the two architecture patterns are not mutually exclusive but rather suited for different scenarios and enterprise maturity stages. In fact, in many practical applications, single and multi-Agent architectures are often nested: a multi-Agent system may consist of multiple single Agents, with each single Agent responsible for processing sequential tasks within its professional domain, while different single Agents collaborate through a coordination layer.
When selecting architecture patterns, enterprises need to comprehensively consider multiple factors: task complexity and number of domains involved, team's experience in Agent development, maturity of existing data infrastructure, and requirements for system reliability and explainability. For enterprises newly exploring the Agent domain, starting with single Agent workflows and focusing on high-value, clearly bounded workflows for pilots is recommended, with gradual introduction of multi-Agent collaboration mechanisms after accumulating sufficient experience.
In-Depth Interpretation of Seven Data Architecture Principles and Use Case Mapping
The seven data architecture principles proposed in the report constitute the theoretical cornerstone supporting Agentic AI at scale, with each principle corresponding to specific technical capabilities and business value. Deep understanding of the relationships among these principles and their mapping to specific use case scenarios is crucial for enterprises formulating practical Agent deployment strategies.
Data Ingestion as Product Principle
The first principle—"Treat data ingestion like a product, making it easy and consistent for all data to enter the company once and be usable by everyone"—lays the shared foundation for data architecture. This principle's core philosophy is that data ingestion should not be viewed as a one-time technical integration project but rather managed as a continuously operating data product. This means the data ingestion process needs to possess product-like qualities: usability (users can conveniently introduce data into the system), consistency (different data sources are ingested according to unified standards), and discoverability (users can easily find the data assets they need).
In Agent application scenarios, this principle's value manifests at multiple levels. First, Agents need to access data from multiple business systems; if each system has an independent ingestion process and standards, each Agent call requires cumbersome adaptation work, severely affecting efficiency and reliability. Second, standardized data ingestion enables Agents to understand and process data from different sources consistently, reducing uncertainty in data interpretation. Third, the discoverability of data ingestion products enables Agents to autonomously discover and ingest new data sources when needed, enhancing system flexibility and adaptability.
From an implementation perspective, achieving data ingestion as a product requires establishing unified data ingestion standards and processes, providing self-service ingestion tools and platforms, and equipping a dedicated product team for continuous operation and optimization of data ingestion experience.
Semantic Unification Principle
The second principle—"Share meaning, not just data, ensuring data comes with clear, common definitions"—addresses the core of enterprise data governance—the standardization of data semantics. This principle emphasizes that data is not merely a collection of raw values but carries specific business meanings. Different systems and departments may have different understandings and definitions of the same data object. If these semantic differences are not bridged, Agents may generate interpretation biases and decision errors during multi-system collaboration.
Taking the seemingly simple concept of "customer" as an example, in different enterprise business contexts, it may refer to different entities: in CRM systems, "customer" may include all potential and existing customers; in financial systems, "customer" typically refers to transaction counterparties with accounts receivable balances; in logistics systems, "customer" may be the consignee or shipper. If an Agent needs to simultaneously process data from these three systems without unified "customer" definition, errors such as identifying the same actual entity as different customers or misidentifying different entities as the same customer may occur.
Achieving semantic unification requires establishing unified data dictionaries and business definitions at the enterprise level, ensuring key business concepts (such as customer, product, order, revenue, etc.) maintain consistent understanding across all systems. This work is typically accomplished through establishing enterprise-level data ontologies and knowledge graphs, providing Agents with a shared conceptual framework.
Unified Analytics and AI Data Foundation Principle
The third principle—"Use one data foundation for analytics and AI, building data once and using it everywhere"—emphasizes the simplicity and reusability of data architecture. In traditional enterprise data construction, analytics (BI reports and data analysis) and AI applications often each construct independent data pipelines and storage, leading to data redundancy, inconsistent standards, and high maintenance costs. This principle explicitly opposes this approach, advocating for establishing a unified data foundation platform supporting all data consumption scenarios.
For Agent applications, the unification of analytics and AI data foundations holds special importance. First, Agents need real-time or near-real-time access to business data; if AI applications and analytics use different data pipelines, inconsistencies between data Agents base their decisions on and data human managers see could occur, severely affecting decision coordination and credibility. Second, Agents' continuous learning and optimization need to be based on the same performance feedback that human analysts use; if feedback data sources are inconsistent, it will hinder Agents' improvement iteration. Third, a unified data foundation reduces data synchronization complexity and latency, enhancing Agent response timeliness.
Trust Mechanism Built-In Principle
The fourth principle—"Build trust into the platform by default, with security, access controls, privacy, and AI governance automation"—embodies the indispensable governance dimension of Agent scaled deployment. As Agents undertake increasingly autonomous decision-making tasks, ensuring Agent behavior is safe, compliant, and trustworthy becomes a core challenge. This principle advocates building trust mechanisms (security controls, access restrictions, privacy protection, AI ethics governance) as built-in capabilities of the data platform rather than post-hoc外挂 modules.
The automatic building-in of trust mechanisms means that when Agents acquire data access permissions, the system automatically performs permission verification; when processing sensitive information, automatic de-identification occurs; and when making high-risk decisions, automatic triggers for manual approval processes or additional verification steps happen. The report specifically emphasizes that Agents should not introduce new data quality or governance rules but should follow the same standards as other systems—however, the execution of these standards needs to be automated as Agent autonomy increases.
Stable Interface Principle
The fifth principle—"Expose capabilities through stable interfaces, providing clear APIs and model access points"—emphasizes the modularity and evolvability of data architecture. In the current era of rapid AI technology iteration, Agent frameworks, underlying models, and toolkits are continuously evolving. Well-designed interfaces enable enterprises to gradually upgrade various technical components without reconstructing entire systems.
The importance of stable interfaces manifests in several aspects: for Agents, reliable APIs mean they can stably call required data and services without worrying about failure due to upstream system changes; for technical teams, clear interface specifications reduce development and integration complexity and decrease maintenance burdens caused by system dependencies; and for enterprise strategy, interface standardization enables Agent assets to migrate across different platforms and technology stacks, reducing vendor lock-in risk.
Behavioral Visibility Principle
The sixth principle—"Make behavior visible and measurable, continuously tracking data quality, model performance, speed, and cost"—provides the necessary infrastructure for Agent operational management. Only when Agent behavior is observable and quantifiable can enterprises achieve effective management and continuous optimization of Agents. This principle requires establishing a comprehensive observability system covering data quality, model performance, system response speed, and cost efficiency.
Behavioral visibility is crucial for Agent governance and optimization. By continuously tracking Agent decision quality, enterprises can identify which scenarios Agents perform excellently in and which scenarios need improvement; by monitoring Agent data usage patterns, potential data quality issues or abnormal access behaviors can be discovered; and by analyzing Agent response time and resource consumption, system performance and cost efficiency can be optimized. The report points out that this observability is not only a management requirement but also the foundation for Agent continuous improvement—Agents can optimize decision strategies by analyzing their own behavioral data.
Enterprise-Level Execution Layer Principle
The seventh principle—"Provide a controlled way to run AI agents and applications, coordinating through a shared execution layer that enforces enterprise rules and guardrails"—ensures consistency between Agent behavior and enterprise policies and compliance requirements. This principle embodies the "shift-left governance" design philosophy—embedding control mechanisms into Agent execution layers rather than conducting post-hoc inspections and corrections.
Core functions of the enterprise-level execution layer include: unified identity authentication and authorization management, ensuring only authorized Agents can access specific data and execute specific operations; centralized policy execution engine, mandatorily applying enterprise compliance and security policies during all Agent operations; and standardized audit logs, recording all Agent critical behaviors for tracing and accountability.
Use Case Evolution Logic of the Four-Step Implementation Pathway
The four-step implementation pathway proposed in the report is not only a technical implementation roadmap but also an organizational capability building pathway. Each step correlates with specific use case scenarios and capability building objectives, constituting a progressive pathway from pilot to scale.
Step One: High-Value Workflow Identification and Agentification
The report recommends enterprises start with a few high-value, end-to-end business workflows for Agentification pilots rather than spreading efforts comprehensively. This strategy is based on several pragmatic considerations: when Agent capabilities are not yet mature and organizational experience is limited, concentrating resources on a few scenarios is more likely to achieve breakthroughs; experience accumulated through pilots on a few scenarios can be reused and migrated to other scenarios; and clear metrics for high-value scenarios facilitate quantitative evaluation of Agent return on investment.
The report explicitly recommends prioritizing knowledge management and marketing functions as pilot domains. This recommendation is based on two aspects: first, according to McKinsey global survey results, marketing, knowledge, and IT are the three most active functions in AI adoption, possessing good foundation and acceptance; second, these functions possess characteristics suitable for Agentification—data-intensive, relatively standardized processes, and clear efficiency improvement opportunities.
Specifically regarding workflow-level Agentification, the report recommends taking the following steps: first, completely map the end-to-end workflow, identifying various nodes and dependencies in the process; then, identify which links are suitable for Agent involvement—typically tasks that are highly repetitive, have clear rules, but require some intelligent judgment; next, evaluate what data support is needed for Agent involvement and check whether this data is available; finally, design clear quantifiable metrics and conduct small-scale pilot verification.
Reusability is the key from pilot to scale. When designing pilot use cases, consciously identifying and constructing reusable components is necessary—reusable data interfaces, reusable Agent capability modules, reusable governance rules, etc. These reusable assets will significantly accelerate subsequent scaled expansion.
Step Two: Layer-by-Layer Modernization of Data Architecture
This step corresponds to data infrastructure construction required for Agent applications. The report emphasizes progressive evolution rather than rebuilding from scratch, with modularity and evolvability as core principles. The breadth and depth of data architecture modernization directly determine the boundaries of Agent application depth and scope.
The report uses an omnichannel retail scenario as an example, showing the layered structure of modernized data architecture. At the data source layer, customer data (such as browsing history, wishlists, purchase history, and support interactions) enters the enterprise from various systems. For unstructured data, continuous ingestion, transformation, and recombination are needed, and data governance must flow along with it. Data quality checks, security controls, and lineage tracking need to be automated and directly embedded in pipelines rather than handled as one-time reviews.
At the data platform layer, data from different systems needs to be connected, making it usable for applications and AI models. This layer achieves this through orchestrating access, synchronization, and real-time cross-system interaction. Vector stores and embedding services are key components for handling unstructured data—they enable documents, images, and other unstructured content to be searchable based on meaning rather than keywords. Agent-specific interoperability standards can further automate integration and access processes, enabling structured context sharing, direct Agent-to-Agent coordination, and secure transactional exchanges.
The semantic layer transforms data into knowledge, sitting between raw data and AI applications, codifying the business meaning of data into machine-readable yet human-understandable form. This layer is most often implemented through ontologies and knowledge graphs—ontologies define how attributes and relationships combine into business reality, while knowledge graphs operationalize this vocabulary by linking real-world data across systems into connected entity networks.
The data products layer turns curated data into reusable, business-ready assets. Data products are managed with a product mindset, treating data as a performance asset that can be reused across multiple use cases and domains. Reusable data products enable Agents to draw on trustworthy predictive and generative insights at scale, while observability records show Agents use data, creating traceability needed for oversight.
Step Three: Data Quality Management Upgrade
This step upgrades data quality management from periodic batch processing mode to real-time continuous mode. The report specifically emphasizes challenges in unstructured data quality management and the viewpoint that Agent-generated new data also needs to be incorporated into the quality management system.
In the Agent era, data quality has become a strategic differentiator. The report points out that organizations with well-structured internal datasets can reduce technology investment costs by fine-tuning smaller, domain-specific models on their own data. These models are not just more cost- and resource-efficient but more resilient and compliant. This means high-quality data assets not only reduce risk in Agent applications but also reduce total technology investment costs.
Making unstructured data usable requires improving its quality through tagging, classification, vector embeddings, and graph-based structuring. This enables Agents to reliably understand entities, relationships, and context. Unstructured data must be held to the same standards as structured data.
Structured data management also needs evolution. The report advocates shifting from periodic cleanups to continuous, real-time data quality monitoring. This process is supported by AI-enabled automated validation, anomaly detection, and enrichment pipelines that prevent issues from propagating across workflows. Metadata management provides lineage and business context so that Agents can trace and justify decisions.
Finally, as Agents generate new data, the same quality, lineage, and reconciliation standards must be applied to their outputs. This includes data retrieved or written through Agent-invoked tools and APIs, which should operate through governed, reconcilable interfaces rather than bypassing enterprise quality control. Shared fit-for-purpose definitions embedded into automated quality checks ensure Agents act on reliable information at scale.
Step Four: Operating and Governance Model Establishment
After scaling, governance becomes the primary control mechanism. This step involves comprehensive design of organizational structure, policy systems, and technical tools, aiming to provide a governance framework for large-scale Agent operation.
The report emphasizes the need for clear and explicit policies defining what Agents can do, what data they can access, and when human approval is required. Access checks are automatically evaluated for each Agent based on their role and scope. This means traditional static access control lists need to evolve into dynamic, context-based permission evaluation mechanisms.
The report recommends a federated governance model: business domains own day-to-day governance of Agent-enabled workflows, including domain models and ontologies; central data and AI teams maintain shared platforms, guardrails, and oversight. This model attempts to balance domain autonomy with enterprise-wide accountability. However, the report also points out this model's implementation complexity—business domains may sacrifice data quality or governance standards for short-term business objectives, and central teams need sufficient authority and tools to ensure enterprise-level policies are executed.
At the technical implementation level, the report recommends establishing dedicated guardrail Agents operating within well-defined control functions, continuously monitoring Agent activity to ensure transparent and compliant behavior. For example, creative compliance Agents can review images and multimedia outputs for brand misrepresentation or policy violations, triggering corrective actions.
IT and governance functions must also manage the Agent lifecycle. This requires issuing credentials, tracking activity logs, monitoring performance, and enforcing policy compliance through automated checks. Agent activity is automatically captured through built-in telemetry, ensuring actions, data access, and decisions are consistently logged and traceable.
Use Case Distribution Analysis from Industry and Function Dimensions
The survey data contained in the report reveals characteristics of AI adoption distribution across different industries and functions, providing an empirical foundation for understanding Agent use case priority distribution.
Use Case Distribution from Function Dimension
According to McKinsey's global survey conducted from June to July 2025, covering 1,993 respondents from 105 countries representing all regions, industries, company sizes, functional specialties, and tenures, marketing, knowledge, and IT are the three functional areas with the most active current Gen AI adoption.
Marketing leads Gen AI usage across nearly all surveyed industries, and this phenomenon has inherent logic. Marketing workflows have several characteristics suitable for AI involvement: content creation (copy, images, videos) is highly repetitive and large-scale, making it a typical AI-generated content scenario; analytical tasks such as customer segmentation, behavior prediction, and personalized recommendation are data-intensive with quantifiable value; and marketing activity execution involves extensive cross-system coordination, suitable for Agent orchestration optimization.
Knowledge management function ranks second, reflecting enterprises' increasingly growing demand for leveraging knowledge assets. In knowledge-intensive work, effective knowledge acquisition, integration, and application are crucial for decision quality and operational efficiency. AI Agent's application value in this domain manifests in: automated knowledge acquisition and integration, semantic knowledge retrieval and Q&A, and knowledge quality lifecycle management.
IT function also demonstrates strong AI adoption momentum, partly because the IT department itself is an early adopter of AI tools, and partly because AI applications in code generation, test automation, and system monitoring are profoundly changing software development and service operation models.
Use Case Differences from Industry Dimension
Survey data shows significant differences in AI adoption depth and breadth across industries. Technology and financial services industries usually lead, possessing better data infrastructure, stronger technology acceptance, and clear efficiency improvement pressure. Manufacturing, healthcare, and retail industries are catching up, exploring AI applications in differentiated scenarios such as supply chain optimization, clinical decision support, and customer experience enhancement.
Notably, priority rankings for Agent use cases also differ across industries. The financial industry may focus more on high-value scenarios such as risk monitoring, compliance checking, and anti-fraud; the retail industry may focus more on front-end operational scenarios such as customer journey optimization and inventory management; and the manufacturing industry may focus more on back-end support scenarios such as predictive maintenance and production planning optimization.
In-Depth Reflection on Extended Use Case Scenarios
Based on deep understanding of the report's core framework, systematic reflection on Agent application scenarios can be further expanded, identifying potential use case domains not explicitly mentioned in the report but possessing significant value.
Enterprise-Level Decision Support System
One of the most strategically valuable application scenarios of the multi-Agent workflow architecture mentioned in the report is constructing an enterprise-level decision support system. In this model, Agents from different specialized domains are each responsible for their domain's analysis and can collaborate to complete cross-domain comprehensive analysis tasks.
Taking market entry decisions as an example, a well-designed multi-Agent system can collaborate: the Marketing Agent analyzes target market scale and growth potential, competitive landscape, and consumer characteristics; the Finance Agent evaluates return on investment, capital requirements, and cash flow impact; the Legal Agent assesses regulatory environment, compliance requirements, and potential legal risks; the Operations Agent analyzes execution feasibility, capability gaps, and supply chain impact; and the Human Resources Agent evaluates talent needs and organizational change impact. Each Agent's output is integrated through the semantic layer, providing decision-makers with comprehensive, balanced analysis reports.
This application scenario's core value lies in: its ability to synthesize multi-domain expertise, avoiding blind spots from single perspectives; its ability to quickly respond to external environmental changes, supporting dynamic decision adjustment; and its ability to trace decision rationale, improving decision explainability and auditability.
Intelligent Risk Management System
Agents possess enormous application potential in the risk management domain. Traditional risk management often relies on periodic reports and manual analysis, making it difficult to capture rapidly changing risk signals in a timely manner. Agents can serve as the enterprise's "neural network" for risk management, continuously monitoring all levels of enterprise operations, identifying potential risk signals, and coordinating response measures.
In credit risk management scenarios, Agents can integrate customer financial data, transaction behavior, market sentiment, industry trends, and other multi-source information in real time, continuously updating credit assessment models and providing timely warnings of potential credit risk events. In operational risk scenarios, Agents can monitor all links of the supply chain, identify potential supply disruption risks, and automatically assess impact scope and recommend response plans. In compliance risk scenarios, Agents can continuously track regulatory policy changes, automatically assess their impact on enterprise business, and track compliance remediation progress.
Intelligent Finance and Accounting Automation
The data-intensiveness and process standardization of the finance domain makes it a naturally high-quality scenario for Agent applications. Agent applications in finance functions can extend to multiple levels.
In financial analysis and report generation, Agents can aggregate data from multiple sources, perform financial analysis, and automatically generate various financial reports and analytical materials. This includes not only standard management reports but also customized analytical reports based on specific management needs. Agents can explain the driving factors behind financial data changes, predict future financial trends, and provide insights on improvement recommendations.
In accounts receivable management, Agents can achieve end-to-end automation from customer credit assessment, invoice generation and sending, payment tracking, and collection management to bad debt write-off. Particularly in collection scenarios, Agents can develop personalized collection strategies based on factors such as customer's payment history, current communication response, and financial status.
In tax compliance, Agents can continuously track changes in tax laws, automatically assess their impact on various business lines, calculate optimal tax arrangements, and ensure accuracy and timeliness of tax filings.
Supply Chain and Operations Optimization
Agent applications in the operations domain focus on improving operational visibility, predictive capability, and response efficiency. In supply chain management scenarios, Agents can integrate data from suppliers, logistics providers, retailers, and other parties, constructing an end-to-end supply chain visibility platform.
In demand forecasting, Agents can integrate various information such as sales data, historical patterns, seasonal factors, marketing activities, and macroeconomic indicators, performing more accurate demand forecasting than traditional statistical models. This forecasting focuses not only on total volume forecasting but also breaks down to SKU level, channel level, and time range level.
In inventory optimization, Agents can dynamically adjust replenishment strategies based on real-time demand signals and supply situations, achieving optimal balance between inventory holding costs and stockout risk. In supply disruption scenarios, Agents can quickly assess the impact of alternative solutions and automatically trigger emergency response processes.
Human Resources and Talent Management
Human resources management involves extensive documentation work and cross-system coordination, and Agent application potential in this domain is emerging. In recruitment scenarios, Agents can undertake multiple links from position requirement analysis, candidate persona construction, resume screening, and initial communication to interview scheduling.
Agents can not only screen resumes based on position requirements and historical successful hiring data but also provide more comprehensive candidate assessments by analyzing supplementary data sources such as candidates' public information and social media presence. In interview scheduling, Agents can coordinate availability of candidates, hiring managers, and interviewers, automatically handling schedule arrangements and change notifications.
In employee development, Agents can analyze employee learning behaviors, work performance, and career development paths, recommending personalized learning content and development plans. Agents can also proactively identify high-potential talents and retention risks, providing management with warnings and recommendations.
Customer Service Experience Upgrade
Customer service is one of the most mature domains for AI applications, but current maturity is mainly reflected at the simple FAQ Q&A level. Agent capabilities described in the report expand this domain to a higher level.
Agents can undertake the full process of complex problem handling: from problem understanding and diagnosis, multi-system data querying, solution formulation, to execution tracking. This capability is particularly valuable in B2B customer service scenarios, where problems typically require crossing multiple subsystems—orders, inventory, technical support, and finance—to resolve.
Agents can also achieve proactive customer service models. Agents proactively monitor customers' product usage situations, service expiration reminders, and potential problem signals, initiating service contact before customers even raise demands. This proactive service model can significantly enhance customer experience and loyalty.
Conclusions and Strategic Implications
The report starts from the core thesis that "shaky data is the primary obstacle to Agent scaled deployment," analyzes the common dependency of two Agent architecture patterns on data consistency, derives seven data architecture principles, and further proposes a four-step implementation pathway. This reasoning process is logically rigorous, and the connection between arguments and thesis is reasonable.
The report's diagnosis of the problem demonstrates strong insight. Attributing the root cause of Agent scaled deployment failure to data infrastructure issues rather than model capability or algorithm issues aligns closely with industry practical experience. The data cited in the report showing that 80 percent of companies cite data limitations as the primary obstacle to Scaling AI provides strong empirical support for the core thesis.
The first area is deeper discussion of Agent and human collaboration models. The human retreat to supervisor role implied by the report may be overly simplified. Different types of work tasks may require different degrees of human participation—highly structured, risk-controllable tasks may be suitable for highly autonomous Agent processing; tasks involving high uncertainty, significant value judgments, and complex interest coordination may require close human-Agent collaboration. How to design optimal human-machine collaboration models for different task types is a question worthy of in-depth research.
The second area is discussion of technical safeguards for Agent reliability and security. The report mentions the concept of guardrail Agents but elaborates little on their technical implementation. How to ensure Agent behavior conforms to expectations without producing unintended consequences? How to prevent Agents from malicious attacks or manipulation? How to design effective anomaly detection and recovery mechanisms? These questions are crucial for Agent safe and reliable operation.
The third area is reflective discussion of ethical and social impacts of Agent applications. The report primarily discusses Agent value from the perspective of enterprise efficiency and competitiveness but lacks in-depth discussion of potential social impacts such as employment replacement, privacy erosion, and digital divide. As Agents undertake increasingly more work tasks, these ethical and social issues will become increasingly prominent, and enterprises need to bear corresponding social responsibilities while pursuing efficiency.
McKinsey's report provides a systematic analytical framework and action guide for Agentic AI scaled deployment. Its core contribution lies in identifying the important insight of data infrastructure as a bottleneck for Agent scaling, and constructing a complete framework from architectural principles to implementation pathways around this insight.
From a use case scenario perspective, the report's covered application scenarios span major enterprise functional domains including knowledge management, marketing, customer service, finance, human resources, and supply chain, as well as cross-functional decision support and risk management scenarios. The report's survey data reveals that marketing, knowledge, and IT are the most active functional domains in AI adoption, providing reference for enterprise Agentification priority decisions.
From an implementation strategy perspective, the report's four-step pathway possesses strong operability. Recommending enterprises start with high-value workflows for pilots rather than spreading efforts comprehensively is a pragmatic suggestion helping control risk and accumulate experience. Reusability is the key principle throughout the entire implementation process—from reusable data interfaces to reusable Agent capability modules, these reusable assets will accelerate the progression from pilot to scale.
From a governance model perspective, the federated governance model proposed in the report attempts to balance domain autonomy with enterprise-wide accountability. Although this model faces challenges at the execution level, its core idea—business domains own day-to-day governance, central teams are responsible for platforms and oversight—provides a valuable reference framework for Agent governance in complex organizations.
Looking ahead, the development of Agentic AI will profoundly transform enterprise operational models and organizational forms. In this transformation, technical capability, data infrastructure, and organizational capability building are all indispensable. Enterprises that can balance technology evolution with organizational resilience will be more likely to gain sustained competitive advantage in the Agent era. Data is no longer merely a byproduct of operations but has evolved into the cornerstone of enterprise core competitiveness in the Agent era.