• Utility technology leaders recognize the value of agentic AI; the real challenge is determining where autonomy is appropriate given constraints in regulatory, safety, and accountability.
• AI agents introduce new cyber, compliance, and governance risks that CIOs must proactively manage and extend beyond just the automation itself.
• Agents are treated as binary (e.g. fully manual or automatic) rather than a scale based on criteria such as risk, which limits opportunities.

Many utilities face common obstacles when implementing agentic AI due to the novelty of the technology. Research has shown the most common obstacles revolve around:

• Data readiness (especially given legacy applications and fragmented data systems).
• Ethical challenges such as concerns of data security and privacy.
• A lack of regulation, creating hesitation in agent deployment without clear regulatory guardrails.

1. Anchor agentic AI opportunities to utility business reference architecture.
2. Define the autonomy vs. risk model for your organization to provide selection guardrails.
3. Define agentic AI use case selection criteria (e.g. business impact and implementation feasibility).
4. Identify candidate use cases and their characteristics.
5. Prioritize use cases for implementation roadmap.

Data Processing

Extract, classify, validate, and structure data from documents or other unstructured inputs.
Examples:

• AMI Data Normalization Agent: Ingest and correct smart meter data, flag anomalies for review
• Water Quality Aggregator: Collect and standardize water quality and prepare for regulatory reporting

Decision Support

Gather information from multiple systems, perform analysis, and generate insights.
Examples:

• Water Demand Planner: Recommend optimal pumping schedules based on consumption patterns and weather and population data
• Gas Trading Optimizer: Recommend optimal procurement volumes based on market prices, storage levels, and demand

Monitoring & Alerting

Continuously scan inputs and detect issues, trigger alerts, and follow-up tasks.
Examples:

• Gas Leak Risk Detector: Detect potential leaks to trigger response by aggregating operational and customer data
• Transformer Health: Predictive maintenance by monitoring transformer operations to alert teams of anomalies

Triage & Orchestration

Classify incoming requests, understand intent, and route work to the correct workflow or system.
Examples:

• Outage Response Coordinator: Identify fault location and triggers crew dispatch and customer notifications
• Water Main Break Orchestrator: Upon pressure drop, isolate affected zone and notify customer and schedule repair crew

Scripted bots that mimic human actions in structured, repetitive workflows, no learning, and relies on predefined rules.

Adds AI to handle unstructured inputs and simple decisioning. Largely workflow-driven, but can classify, extract, and route.

Interpret goals, reason within a bounded scope, and execute multi-step tasks via tools/APIs.

Multiple Agents working together to solve more complex problems through coordination and task decomposition. Agents help do the work.

AI systems that manage and optimize an entire process within a specific domain, integrating data, analytics, and execution. Agents own the work.

Ecosystem of agents operating across business functions to autonomously execute and optimize processes. Agents run the business.

• Cost reduction through task automation
• Improved speed and accuracy

• Improves process beyond simple rules
• Handles variability in inputs

• Handles ambiguity and complex queries
• Reduces need for rigid workflows

• Enables modular and scalable automation
• Handles multi-step workflows

• End-to-end optimization
• Continuous improvement and learning

• Enterprise-wide optimization and coordination

• Breaks when systems change
• Cannot handle unstructured data or exceptions

• Relies on predefined workflows
• Requires training data and tuning

• Dependence on data quality and integrations
• Limited to narrow domains/use cases

• Orchestration complexity
• Higher compute & integration requirements

• Strong integration and data foundations
• Regulatory explainability and requirements

• High complexity, governance, risk, and compliance requirements
• Need human oversight

Automating billing data transfer between CIS and ERP systems

Extracting customer email requests and routing to appropriate teams

Customer agent resolving billing inquiries end-to-end (issue handling and correction)

Outage response agents predict outage scope and triggers corrective response

Predictive maintenance system monitors equipment health and schedules maintenance

Storm response integrates weather data, grid ops, crew logistics to manage restoration efforts

• Introduce general agentic AI concepts (e.g. automation level, capability patterns, utility centric trends)
• Identify agentic AI use cases aligned to business reference architecture

• Review or design risk vs. autonomy model suitable for the organization
• Define appropriate risk appetite and guardrails
• Determine appropriate level of agent autonomy

• Review the use case evaluation tool
• Define the business impact and feasibility criteria
• Evaluate the identified use cases

• Review the results from the evaluation tool
• Define a phased implementation plan for selected use cases suitable for pilots

Gain a shared understanding of agentic AI and where it could create value in the organization. This ensures agentic AI are designed to drive business value and uncovers opportunities that matter most for the business.

Autonomy of AI agents is right-sized based on their risk profiles to ensure they are designed safely. Guardrails are put in place to help justify the level of autonomy versus risk exposure.

Quantify the business impact and feasibility of agentic AI use cases, allowing the organization to identify a curated list of agents that reflect real operational constraints.

Agentic AI use cases are no longer ideas, but initiatives the organization can prioritize to start piloting and experimenting having completed the due diligence required.

• Initiate Agentic AI Strategy: IT is typically the initiator of many AI initiatives; the methodology outlined in this blueprint helps start conversations around agentic AI.
• Improve AI Governance and Risk Management: Agentic AI introduces many risks, especially within the context of utilities. Ensuring guardrails are designed upfront will improve pilot successes.
• Understand Gaps in Technology Foundations: An evaluation criteria is feasibility; IT will be able to identify the requirements to implementing AI agents compared to the actual realities (e.g. data quality, integration), creating an opportunity to address foundational gaps.
• Provide a Framework to Justify AI Agents: The methodology and tools within the blueprint produces a structured analysis linking risk, value, and readiness to justify AI agents.

• Align AI Agents to Business Value: Organizations can clearly link AI agents to value drivers that support the utility business (e.g. generation, retail, and customer services).
• Define a Strategic Agentic AI Roadmap: Scaling agentic AI requires an incremental approach to move up the maturity level (e.g. single vs. multi-agents). Create a roadmap to realize the organizations agentic ambitions.
• Prioritize Investments: The methodology is a way to conduct due diligence into the value and feasibility of agentic AI use cases and supports the justification for the business to invest dollars in pilot initiatives.
• Unified Vision of Agentic AI: Create a common understanding of agentic AI concepts and where the organization wishes to go, democratizing ideation of use cases.

Call #1: Review the business capability map and value streams.

Call #2: Identify domains where agentic AI may create value.

Call #3: Determine your organization’s autonomy and risk tolerance.

Call #4: Apply autonomy and risk model to identified use cases.

Call #5: Define use case evaluation criteria.

Call #6: Evaluate use case value and feasibility.

Call #7: Prioritize use case based on risk, value, and readiness.

Call #8: Develop a sequenced implementation roadmap.

Agentic AI Foundations and Business Alignment

Calibrate Use Case Autonomy Level Depending on Risk

Evaluate Use Case Value and Feasibility

Prioritize Initiatives and Build Roadmap

1.1 Introduce research, framework, and purpose of workshop

1.2 Discuss business context and agentic AI strategy

1.3 Review the business reference architecture and identify agentic AI opportunities

2.1 Brainstorm list of agentic AI use cases across business capabilities

2.2 Design the autonomy vs. risk model

2.3 Apply autonomy vs. risk model to use cases to select appropriate autonomy level

3.1 Define the value and feasibility metric for use case evaluation

3.2 Evaluate the identified use cases

3.2 Determine high-level solutioning of each use case (e.g. level of autonomy, agent patterns)

4.1 Plot use cases on 2x2 decision matrix

4.2 Define a phased implementation roadmap for selected use cases; identify ownership and next steps for pilots

1. Alignment of agentic AI concepts
2. Agentic AI opportunities mapped to business reference architecture

1. Autonomy vs. risk model
2. Defined use case selection criteria
3. Shortlist of agentic AI opportunities

1. Defined use cases with design criteria (e.g. automation level, operational role)

1. Holistic view of use case viability
2. Prioritized roadmap for implementation

• Agentic AI failure impacts the ability for the utility to provide core operational services (e.g. grid stability, water supply).
• The higher the risk, the higher the potential impact of failure or disruption.

• Potential for AI agent to cause physical harm to people or environment or violate safety compliance.
• AI agents must comply with safety standards and regulatory reporting requirements (e.g. NERC, EPA).

• Handling of customer-sensitive, confidential, or operational data (e.g. billing details, personal info).
• Improper use or exposure of this data can lead to privacy violations, regulatory penalties, or loss of public trust.

• Measure of the ease of reversing actions executed by agents and if there are downstream effects (e.g. opening a switch vs. equipment damage).
• The less reversible the action, the lower the acceptable level of autonomy.

• Assesses the potential economic consequences of AI agent’s decisions.
• Includes both direct costs (equipment damages, fines) and indirect costs (customer churn, reputational damage).

With the expansion and growth of renewable power-generating assets, safety audits continue to grow as an insurmountable task. A single wind farm could have as many as 75 separate turbines, a dramatic increase compared to a traditional single-turbine power plant. This led to a significant increase in the volume of internal safety audits, which were time-consuming, taking up to two weeks to complete. This process was a major drain on company resources and was performed by employees who were not dedicated auditors, making it inefficient.

AES partnered with Anthropic and Google Cloud to develop a multi-agent system to dissect the audit process into simpler tasks. The three agents created were:

• Document processing agents
• Task breakdown agents
• Report generation agents

The system automated the entire safety audit process, leveraging Anthropic and Google’s enterprise grade security, speed, and accuracy for handling sensitive audit data.

• Time Reduction: Reduced the process from two weeks to ~1 hour; a 99% decrease.
  
• Reduce Audit Costs: 99% reduction in total audit costs.
  
• Audit Accuracy Improvements: Agents produced more accurate reports than the manual process by 10% to 20%.
  
• Resource Reallocation: Enabled AES to reallocate field resources to support other business needs closer to their core responsibilities.