IBM Manages AI Agents as 'Digital Workers'

The AI Agent Lifecycle Management (ALM) consists of eight stages, similar to human resource management. In the initial phase, items for agent introduction are selected and opportunities are identified within the organization, followed by an evaluation of value and risks. Next, the skills the agent will perform are designed, and a prototype is created.

This is followed by pilot operation and deployment phases to test the agent in a real environment, including operational learning, improvement, promotion, and decommissioning stages to manage the agent's overall lifecycle.

The agent operation phase comprehensively covers concepts such as continuous learning and improvement, promotion, and decommissioning. If an agent is performing as well as initially evaluated, it is maintained. If demand creates bottlenecks, a decision may be made to split agents to expand processing capacity.

Conversely, underutilized agents should be decisively decommissioned, and flexible management, such as adding skills for new feature expansion, is necessary. During this process, security and risk management must also be performed in parallel to ensure stable agent operations.

Before agents are deployed for work, their performance and safety must be thoroughly verified through rigorous testing. First, the accuracy and safety of work results must be ensured, and potential hacking concerns or prompt manipulation possibilities must be checked in advance. Additionally, verifying the reliability of data-driven facts generated by agents is essential.

Along with this, token cost efficiency must be analyzed, and user convenience in using the agent must be evaluated from various angles to minimize all risk factors before enterprise-wide deployment.

Before deploying agents across the company or opening them up for customer service, a more stringent process must be followed. This process requires using a comprehensive checklist to review the agent's quality, data protection level, hacking security vulnerabilities, compliance, and ethical aspects.

Change management and release management for agents are more efficient when handled by a dedicated organization. This organization collects and incorporates user feedback to continuously improve agent performance, and is also responsible for optimizing and managing operating costs.

Furthermore, it must consolidate redundant agents and provide clear guidelines for potential risk factors. Specifically, establishing criteria for agents to be decommissioned and carrying out the decommissioning process are also important tasks for the dedicated organization.

Managed Connectivity Point (MCP) technology is a crucial tool for linking agents with a company's legacy systems. However, as the scope of connection to MCP servers expands, system risks are bound to increase proportionally.

Therefore, the dedicated organization must systematically manage MCP connection risk levels by categorizing them into four tiers: Green (low), Yellow (medium), Orange (caution), and Red (high).

MCPs are very useful tools for enterprises, but problems can arise if they are not managed properly. Therefore, clear management policies for MCPs must be established, and they should be classified and managed by tiers such as Red, Orange, and Blue. It is also important to integrate all MCPs through a central catalog and continuously update them.

Enterprises are encouraged to introduce software-based kill switches rather than physical devices, establishing a safety mechanism that can immediately cut off all MCP connections in emergency situations.

For example, if an emergency situation occurs where SAP's ERP system is integrated with MCP and the ERP system is hacked, the company must immediately activate the kill switch to cut off all MCP connections. This is an essential safety mechanism to prevent further damage to the system and to disconnect all connections until it can be safely restored.

While in the past, individuals handled tasks and generated performance, going forward, work will change to a collaborative form between people and agents. Accordingly, performance indicators will also shift from simply task throughput to efficiency-focused metrics such as reduced task completion rates or reduced rework rates.

In this transformation, problem definition, accurate judgment, and outcome verification capabilities will emerge as even more critical core competencies for human employees.

Future key performance indicators (KPIs) will focus not on how much expertise an individual has accumulated, but on how widely they have provided and shared the skills they created as input for agents to learn. This will be a new criterion that values skill sharing and utilization.

New KPIs will be classified into six types—outcome, quality, risk, leverage, learning, and judgment—to measure the collaborative performance of agents and humans from multiple perspectives.

If past managers merely directed and performed tasks as Task Owners or Task Performers, their role will now be redefined as Work Orchestrators. This means that managers will shift towards utilizing various agents and skills to perform more creative or high-value tasks.

Managers in this new role will focus on effectively responding to exceptions within the system and designing optimal collaboration systems between humans and AI.

The first step in an organizational roadmap for AI agent adoption is to accurately understand the concept of 'Agentic Work'. Agents should be recognized not merely as tools to assist humans, but as technologies or machines that perform tasks independently under specific conditions.

Subsequently, a strategic approach led by the HR organization is needed to design the agent portfolio, ultimately leading to platform selection and technical implementation stages.

The fourth step in building an AI agent is deciding on the agent architecture. Once this stage is complete, most employees will transition to skill engineering to set the agent's persona.

Concurrently, an operating system, such as MCP (Managed Connectivity Point) integration and Agent OPS (Operations), must be established to create an environment where agents can function efficiently.

The time it takes to adopt AI agents varies by company, and the adoption strategy must be adjusted based on the organization's technological maturity. If a company already has a Live Coding environment at level 3-4, it is recommended to move directly to level 5 rather than focusing further on prompt engineering or learning new AI tools.

That is, it is effective to quickly transition to the stage of creating one's work as skills and embedding those skills into agents for utilization.

IBM calls its AI agents 'Digital Workers' and manages them in the same way as human employees. Digital Workers are assigned employee numbers and email accounts, and the IBM CEO emphasized that in 2024, HR organizations must manage Digital Workers in the same way they manage human workers.

IBM has established and applied a 4-stage lifecycle management system for Digital Workers: recruitment and onboarding, qualification verification, placement and operation, and retirement.

To officially recognize the capabilities of Digital Workers, IBM has implemented a policy of issuing certifications to AI agents in collaboration with Pearson Credly. This goes beyond merely indicating that an agent possesses certain skills; it officially certifies that the agent is qualified for those skills.

These digital badges enhance credibility by allowing the agent's skill level and qualification status to be publicized in social environments.

IBM's SVP emphasized the importance of efficient resource management by using the phrase 'starve out' unused agents. 'Starving' here means that if an agent's utilization is low, the token supply will be cut off first. Since tokens are the cost required for an agent to perform tasks, cutting them off means the agent can no longer operate.

This process proceeds in the order of certification revocation, service account removal, and logging backup before disposal, preventing unnecessary resource waste and maximizing the efficiency of the agent ecosystem.

The governance and organizational structure of AI agents are most commonly managed through collaboration between HR and IT organizations. This stems from the shared experience that it is much more efficient and quicker to implement when the existing human resource management system is adapted and applied to agent management.

HR is responsible for policies and lifecycle management related to agent 'human resources,' and IT supports technical implementation and operational infrastructure, thereby building an integrated and systematic agent management system.