Self-Check: Is your company friendly toward agents?

Author: Zhang Feng

1. Trend: In the AI era, comprehensive business integration is unstoppable

In recent years, the global technology industry has undergone a silent yet profound paradigm shift. If the past decade was the golden age of “Internet+”, then the present and the next decade will be the deterministic future of “AI+”. From the wave of large models sparked by ChatGPT to various industries rushing to deploy intelligent agents, an undeniable fact is emerging: AI is no longer just a supplementary tool, but is becoming the infrastructure for enterprise survival and development.

Behind this trend are three forces driving together.

First is the exponential leap in cost efficiency. In traditional business processes, manual handling of information, resource coordination, decision-making, and other links are not only costly but also limited in efficiency by human physiological limits. AI agents can work 24/7, processing speeds hundreds or thousands of times faster than humans, with error rates decreasing through continuous training. For example, in customer service scenarios, a well-trained AI customer service agent can handle thousands of conversations simultaneously, with single-service costs at just one percent or even less of human costs. When competitors complete the same tasks at one-tenth the cost and ten times the speed, companies that do not adopt AI are essentially running with their legs tied in a race.

Second is the deep unlocking of data value. Enterprises have accumulated vast amounts of business data, but these data often lie dormant in various systems, failing to transform into real assets. One core capability of AI agents is extracting insights from unstructured, messy data to support decision-making. Tasks that previously took analysts a week to complete, such as report analysis, AI can now accomplish in minutes, discovering correlations and trends that are difficult for humans to perceive. This ability turns data from “post-event records” into a core driver of “real-time decision-making.”

Third is the market competition’s forcing effect. Pioneers are leveraging AI to establish new competitive barriers. In retail, AI-driven dynamic pricing and personalized recommendation systems are reshaping consumer experiences; in manufacturing, AI agents optimize scheduling and predictive maintenance to significantly improve overall equipment efficiency; in finance, AI risk control and smart investment advisory are redefining service boundaries. As industry innovators begin reconstructing business processes with AI, latecomers will face not just a choice of “whether to adopt,” but a gap of “how far behind” they are.

This round of AI revolution differs fundamentally from previous technological changes. The internet changed how information is distributed; mobile internet changed how we connect; AI changes how we “think” and “act.” Intelligent agents are no longer passive tools executing commands but autonomous entities capable of understanding goals, planning paths, calling tools, and forming closed loops. This means that enterprise integration with AI cannot stay at the level of “installing software” or “adding systems,” but must delve into the core of business logic, management processes, and organizational structure.

2. Integration: Understanding the four dimensions of “AI integration”

To assess whether a company is prepared for AI integration, first clarify what “integration” entails. It is not a single-point action but a systemic project involving internal and external, multi-layered systems. Specifically, it includes at least four dimensions.

(1) Internal management integration

Internal management integration is the foundational layer of enterprise AI integration, involving embedding AI agents into various internal operational management processes. This includes but is not limited to:

Human Resources Management: Using AI for resume screening, interview scheduling, employee profiling, training recommendations, performance analysis, etc. AI agents can quickly process large amounts of candidate information, identify the best matches, and even assist in evaluating interview records via natural language processing.

Financial Management: Automated approval of expense reimbursements, automatic extraction and entry of invoice information, real-time monitoring of budget execution, intelligent alerts for abnormal transactions. Mature financial AI agents can automatically match purchase orders, identify duplicate reimbursements, and flag non-compliant invoices.

Administrative Management: Intelligent coordination of meeting arrangements, automatic travel plan recommendations, dynamic allocation of office resources. Agents can consider participants’ schedules and real-time traffic data to find the most suitable meeting time and location.

Process Approval: Automating fixed-rule approval workflows, and intelligently routing and handling exceptions. AI agents can learn past decision patterns, automatically approve routine matters, and flag complex issues for manual review.

(2) External business integration

External business integration is the core value layer of enterprise AI integration, involving applying AI agents to customer-facing, supplier, and partner business processes.

Marketing and Customer Acquisition: AI-driven user behavior analysis, personalized content recommendations, ad optimization, potential customer scoring. Agents can analyze user behavior on websites and apps in real-time, predict purchase intent, and push suitable products at optimal moments.

Sales and Conversion: Intelligent sales assistants can provide sales staff with customer profiles, communication suggestions, competitor comparisons, and pricing strategies. More advanced, automated sales agents (like conversational business bots) can handle the entire process from inquiry to order.

Customer Service: Currently the most widespread AI application scenario. AI customer service can handle most common questions, perceive user emotions via sentiment analysis, and seamlessly transfer to human agents when needed. AI outbound calling systems can conduct customer follow-ups, satisfaction surveys, and overdue reminders.

Supply Chain and Procurement: Automating supplier evaluation, demand forecasting, order tracking, and logistics optimization. AI agents can integrate internal and external data, forecast raw material price trends, and assist procurement decisions.

(3) Ecosystem integration

This is a higher-level integration form, where companies participate in larger business ecosystems through intelligent agents, engaging in “machine-to-machine” dialogue and collaboration.

Cross-organizational process automation: In supply chain scenarios, AI agents of buyers and suppliers can automatically complete inquiries, comparisons, orders, confirmations, reconciliation, and payments, all without human intervention.

Industry data sharing and collaboration: In logistics, finance, healthcare, and other industries, multiple participants’ agents can share anonymized data under unified standards and protocols for joint modeling and optimization.

Platform-based ecosystem and agent marketplace: Some large platforms are building “agent stores,” where companies can publish their own agents for others to call, or subscribe to third-party specialized agents. For example, an e-commerce seller agent can coordinate with logistics delivery agents, payment settlement agents, and marketing promotion agents.

Smart contract execution on consortium blockchains: In multi-party blockchain collaborations, AI agents can monitor whether preset conditions are met and automatically trigger smart contract execution, enabling highly trustworthy automation.

(4) Governance and compliance integration

This is an essential safeguard layer, emphasizing that AI capabilities must comply with laws, regulations, industry standards, and social ethics.

Data compliance: AI agents’ data collection, processing, storage, and transmission must meet regulations like the Personal Information Protection Law (PIPL), GDPR, etc. This includes obtaining user authorization, anonymizing or de-identifying data, and supporting user data deletion rights.

Algorithm transparency and explainability: When AI decisions impact user rights (e.g., credit approval, recruitment, insurance pricing), companies need to explain the key basis of decisions. Black-box models face increasing regulatory scrutiny in high-risk scenarios.

Content safety and value alignment: Public-facing AI chatbots must ensure generated content complies with laws, social norms, and does not spread false information, discriminatory speech, or harmful content. This requires multi-layered safety mechanisms from pretraining to fine-tuning and real-time monitoring.

Responsibility boundaries and emergency mechanisms: When AI behavior causes damage, how is responsibility assigned? Governance structures must clarify AI behavior limits, human oversight points, and incident handling procedures.

3. Current situation: Most companies are unprepared for AI integration

Using the four dimensions above as a benchmark, it is an embarrassing reality that most companies’ business systems and management processes are highly hostile or even “enemy-like” environments for AI agents.

(1) Internal management: data swamp and process maze

The biggest obstacle in internal management is data quality. AI agents are not magicians; they rely on high-quality, structured, semantically consistent data. However, many enterprises’ internal data conditions are so poor they can be described as “dreadful”:

• Same fields (e.g., “Customer Name,” “Product Model”) used across departments with completely different naming conventions and coding systems;
• Paper documents coexist with electronic files, with key information locked in scans, PDFs, or voice recordings, requiring complex OCR and speech recognition before AI can “understand” them;
• Missing historical data, input errors, duplicate records are common, and data governance remains just a slogan;
• Business processes lack standardization; the same approval tasks may follow completely different paths across branches and personnel.

When AI agents try to access such internal systems, they face not a smooth information highway but a swamp full of potholes, dead ends, and pitfalls. For example, an expense approval agent that cannot distinguish between “travel expenses” and “transportation costs” across systems cannot automatically approve reimbursements.

(2) External business: closed interfaces and arbitrary processes

External business systems face similar issues. Many companies’ customer interfaces and transaction systems were designed without considering “AI agents as users.”

Lack or poor quality of APIs is the biggest pain point. An AI procurement agent aiming to compare products automatically will struggle if the supplier’s website does not provide open, standardized, and authenticated product information APIs. It may have to resort to “browser simulation” scraping—fragile, inefficient, and often against site terms. Many companies treat “external interfaces” as a technical issue rather than a strategic one, with outdated API documentation, poor authentication, and opaque rate limits, making AI calls difficult.

Human-centered process design is also problematic. Most business interfaces are designed for human visual perception, reaction times, and attention patterns: complex multi-level menus, hover-triggered options, dynamic CAPTCHAs, mandatory pop-up notices… These may be only slightly inconvenient for humans but are insurmountable barriers for AI agents. An order query bot, for example, encountering a three-level menu, a date picker requiring “last 7 days,” and a pop-up survey, will face a fragile script full of conditional branches and exception handling.

(3) Ecosystem participation: siloed and lacking standards

If internal and external system issues can be addressed by enterprise efforts, ecosystem-level challenges are beyond any single company’s control.

The primary problem is heterogeneity in system architectures. Company A’s ERP and Company B’s CRM have completely different data models; order status codes and logistics status codes are not interoperable. When multiple companies’ agents attempt end-to-end collaboration, the first challenge is not complex logic but the basic “translation” problem—requiring extensive customization, defeating the purpose of automation.

Lack or fragmentation of standards and protocols is a deeper issue. Although some industry groups and alliances promote B2B standards (like versions of EDI, RosettaNet, OASIS), these are often outdated, too complex, costly to implement, or limited to specific sectors. A comprehensive, open, universal standard protocol covering “agent discovery, capability negotiation, data exchange, status synchronization, exception handling, settlement” remains absent.

Trust and security mechanisms are another major hurdle. How does one enterprise verify another’s agent identity and authorization? How to ensure sensitive data isn’t leaked or eavesdropped? When an agent behaves maliciously (e.g., hijacked), how to detect and revoke its access? Existing security frameworks (OAuth, API keys, mTLS) help but fall short when interactions shift from “enterprise systems” to “autonomous agents” and from “request-response” to “multi-turn dialogue and autonomous decision-making.”

(4) Governance and compliance: responsibility vacuum and regulatory lag

Compliance issues are equally concerning. Many enterprises rush AI deployment without establishing governance structures that keep pace with technology.

Responsibility for AI decisions is ambiguous. If an HR AI wrongly filters out qualified candidates, is the algorithm engineer responsible? The business unit? Or the AI itself (if legally possible)? Many companies lack clear, operational responsibility rules, leading to accountability gaps.

Lack of ethical review mechanisms. Which AI applications require ethical review? Who forms the review committee? What standards are used? Most companies have no formal procedures. As a result, AI with obvious risks of discrimination or privacy violations may go live without thorough assessment, only to be caught after complaints or public backlash.

Regulatory “transposition” of traditional compliance. Many companies simply transpose existing regulations into AI scenarios without understanding new challenges. For example, data protection laws require explanations for automated decisions; can a deep neural network’s decision boundary be explained in natural language? Companies often give superficial explanations (“based on features X, Y, Z”) without addressing true interpretability.

4. Action: The path of comprehensive importance

In facing these complex challenges, companies cannot passively wait for mature technology or standards but must proactively advance AI integration from strategic, technical, business, and compliance perspectives.

(1) Strategic level: shift from “tool mindset” to “ecosystem mindset”

Top management must recognize that AI is not just a software project but a strategic variable affecting business models and organizational forms. Integration is not just an IT task but a whole-company effort. Clear AI strategic roadmaps should be developed, clarifying which processes will adopt AI, whether capabilities are built or bought, how to balance automation and human involvement, and how to adapt to external ecosystem changes.

More fundamentally, companies should transform their view of AI from an internal efficiency tool to an external ecosystem participant. This involves thinking about “how our agents will collaborate with others” and “what role we want to play in the agent ecosystem.”

(2) Technical level: build “friendly-to-agent” system architecture

Technical teams need to re-examine existing systems, shifting from “human-centered” to “dual human-and-agent-centered” design. Specifically:

Comprehensive API exposure: All core functions should be accessible via well-designed, well-documented, version-controlled APIs optimized for machine-to-machine calls (e.g., batch operations, asynchronous callbacks, rate limiting, retries).

Data readiness: Establish unified data governance, ensuring key entities (customers, products, orders, suppliers) have unique, clear, machine-readable definitions across the enterprise. Invest in cleaning and annotating historical data.

Embedded intelligence: Decouple valuable business logic (approval rules, pricing strategies, risk standards) from hardcoded applications into callable “capability modules.” Agents can then assemble these modules like building blocks rather than re-implementing or bypassing them.

Observability design: Provide comprehensive logging, monitoring, and tracing for AI agent actions. When multiple agents collaborate, traceability of decisions based on data and timing is essential for troubleshooting and responsibility.

(3) Business level: redesign processes and roles

Business units must not leave AI integration solely to IT. Process reengineering is key.

Simplify and standardize workflows: Before AI integration, review whether existing processes are rational. Avoid automating chaos; streamline and standardize to reduce exceptions.

Design human-AI collaboration: Not all steps should be fully automated. Clarify which tasks are AI-only, which are AI-assisted, and which remain manual. Design smooth handoff and escalation mechanisms.

Empower employees: AI will change work content, not eliminate jobs. Provide training to help staff collaborate with AI, understand decision rationales, and handle exceptions when AI fails.

(4) Compliance level: embed governance mechanisms

Compliance cannot be an afterthought but must be integrated into AI system design, development, deployment, and operation.

AI ethics committees: Establish cross-departmental review bodies for high-risk AI applications, including legal, compliance, technical, and external experts.

Impact assessment processes: Conduct formal assessments before deploying AI affecting sensitive rights—credit, employment, health, insurance—to identify fairness, transparency, privacy, and security risks.

Continuous monitoring and auditing: AI behaviors drift over time; implement ongoing performance tracking and periodic audits to ensure compliance.

Transparent disclosure: When appropriate, disclose to users, partners, and regulators the AI scenarios in use, their basic principles, and user rights (e.g., appeal channels, human intervention).

5. Caution: Key considerations for AI integration

During comprehensive deployment, pay special attention to these points:

Avoid “AI solutionism.” AI is powerful but not omnipotent. Not all problems are suitable for AI, and not all processes justify the cost of automation. Maintain cost-benefit awareness. Small-volume, highly complex, error-tolerant scenarios may be better handled manually.

Beware of data bias risks. AI learns from historical data, which may contain biases—racial, gender, regional, etc. Using biased data can reinforce or amplify discrimination. Conduct fairness audits and bias detection before deploying impactful AI systems.

Don’t overlook “exception handling.” No matter how accurate, AI will err. The real test is how it handles exceptions—missing data, timeout issues, protocol mismatches. An AI system designed only for ideal conditions will fail at the first obstacle.

Establish “human-in-the-loop” supervision. For high-risk decisions, combine AI with human oversight—e.g., AI suggests, humans approve; or AI handles routine, humans intervene in anomalies. This balances efficiency with accountability.

Align AI capabilities with organizational capacity. Many companies introduce advanced AI but keep old processes, cultures, and incentive systems. This mismatch reduces effectiveness. Synchronize process reengineering, organizational change, talent development, and incentive design to maximize AI’s impact.

When asked whether their company has integrated AI, a confident answer should not be “Our IT department is researching” or “We have launched three AI projects,” but rather: “We have assessed our data, processes, systems, and organization; they are ready for AI; and we understand this is just the beginning of a long journey.”

AI integration is fundamentally a corporate self-reinvention. Companies that successfully cross this threshold will enjoy efficiency gains and secure a competitive position in an increasingly intelligent ecosystem. Those that fail to make their systems “AI-friendly” will find AI agents hitting walls and ultimately moving to more compatible environments.