I attended the “AI for Energy and Utilities” track at Dreamforce 2025 and wanted to share insights on how Agentforce is being deployed in the energy sector. This is a fascinating intersection of AI, critical infrastructure, and sustainability.
Why Energy Sector AI Is Unique
The energy sector has constraints and opportunities that differ fundamentally from other industries:
Critical Infrastructure:
- Power grid failures affect millions
- Real-time decisions (milliseconds matter)
- Safety-critical operations (lives at stake)
- 24/7/365 reliability requirements
Regulatory Complexity:
- NERC CIP (North American Electric Reliability Corporation Critical Infrastructure Protection)
- FERC regulations (Federal Energy Regulatory Commission)
- State utility commissions (50+ different regulatory bodies)
- Environmental regulations (EPA, state-level)
Physical-Digital Convergence:
- AI agents control physical infrastructure (transformers, switches, generators)
- Operational Technology (OT) + Information Technology (IT) integration
- Industrial control systems (ICS) security
- Real-time sensor data (millions of IoT devices)
Sustainability Imperative:
- Net-zero commitments (2030-2050 targets)
- Renewable energy integration (solar, wind variability)
- Grid decentralization (distributed energy resources)
- Carbon accounting and reporting
Dreamforce Energy Sector Use Cases
1. Predictive Maintenance for Grid Infrastructure (PG&E)
What it does:
- Analyzes data from 150,000+ transformers, substations, transmission lines
- Predicts equipment failures 2-6 weeks in advance
- Prioritizes maintenance crews (highest-risk equipment first)
- Reduces wildfire risk (California Public Safety Power Shutoff optimization)
Technical architecture:
IoT Sensors (temperature, vibration, oil quality)
↓
Edge Computing (local anomaly detection)
↓
Data 360 (unified sensor + maintenance history)
↓
Agentforce Predictive Maintenance Agent
↓
Work Order System (Salesforce Field Service)
↓
Dispatch crews to highest-risk equipment
Results (18 months):
- Unplanned outages: 3,400/year → 1,950/year (-43%)
- Wildfire-related incidents: 87% reduction
- Maintenance cost savings: $42M/year (optimized crew dispatch)
- Customer satisfaction (uptime): 3.8 → 4.3 out of 5
Key insight: AI agents bridge the gap between sensor data and physical action. This is operational technology (OT) meeting AI.
2. Renewable Energy Forecasting and Grid Balancing (NextEra Energy)
What it does:
- Forecasts solar/wind generation 24-72 hours ahead
- Balances variable renewable supply with demand
- Optimizes battery storage charging/discharging
- Coordinates with fossil fuel “peaker” plants
The grid balancing challenge:
Supply must EXACTLY match demand every second
- Too much supply → frequency spikes, equipment damage
- Too little supply → brownouts, blackouts
Traditional grid (predictable):
Coal/gas plants ramp up/down slowly
Renewable grid (variable):
Solar: drops 80% in 5 minutes (cloud cover)
Wind: fluctuates with weather patterns
AI agents must predict and balance in real-time.
NextEra’s multi-agent system:
Agent 1: Weather Forecasting Agent
- Ingests NOAA, European weather models
- Predicts cloud cover, wind speeds (hyperlocal)
- Updates every 15 minutes
Agent 2: Demand Forecasting Agent
- Analyzes historical usage patterns
- Accounts for weather (heat waves spike AC usage)
- Detects anomalies (major events, holidays)
Agent 3: Grid Optimization Agent
- Receives forecasts from Agents 1 & 2
- Decides: charge batteries, discharge batteries, or start peaker plants
- Sends control signals to distributed energy resources (DER)
Agent 4: Market Trading Agent
- Buys/sells electricity in wholesale markets
- Optimizes for price + carbon footprint
- Coordinates with grid operator (ISO/RTO)
Results:
- Renewable energy curtailment: 12% → 3% (less waste)
- Battery utilization: +38% (better ROI on storage)
- Wholesale market savings: $18M/year (buy low, sell high)
- Carbon emissions: -22% (optimized fossil fuel usage)
Key insight: Multi-agent orchestration is critical. Grid balancing requires 4+ agents working together in real-time.
3. Customer Energy Efficiency Agent (Duke Energy)
What it does:
- Analyzes customer energy usage patterns
- Recommends personalized efficiency improvements
- Calculates ROI on home upgrades (solar panels, insulation, HVAC)
- Coordinates rebates and incentives
Conversational UI example:
Customer: "Why is my electric bill so high this month?"
Energy Efficiency Agent:
"Your August bill was $287, up 34% from July ($214).
Analysis of your usage:
- Air conditioning: 62% of bill (+$48 vs July)
- Water heater: 18% of bill
- Appliances: 20% of bill
Your AC is running 40% more than similar homes in your area.
This could indicate:
- Dirty air filter (replace every 3 months)
- Poor insulation (check attic, windows)
- Older AC unit (16+ years old, inefficient)
Recommendations:
1. Replace air filter → Save $12/month (free)
2. Attic insulation upgrade → Save $38/month ($1,200 upfront, 32-month payback)
3. New high-efficiency AC → Save $67/month ($6,500 upfront, 97-month payback)
Duke Energy rebates available:
- Insulation: $400 rebate (reduces payback to 21 months)
- AC replacement: $1,200 rebate (reduces payback to 79 months)
[Schedule Energy Audit] [Apply for Rebates] [Talk to Specialist]"
Results:
- Customer engagement: 14% → 31% (participation in efficiency programs)
- Average energy savings: 18% per participating household
- Customer satisfaction: +22 points (customers appreciate personalized advice)
- Demand reduction: 340 MW (equivalent to avoiding one new power plant)
Key insight: AI agents enable mass personalization. Each customer gets tailored advice, not generic efficiency tips.
4. Electric Vehicle (EV) Managed Charging Agent (Southern Company)
What it does:
- Coordinates charging for 45,000+ EVs on the grid
- Optimizes charging times (off-peak, when renewable energy is abundant)
- Prevents grid overload (too many EVs charging simultaneously)
- Enables Vehicle-to-Grid (V2G) for grid stability
The EV challenge:
One EV charging = equivalent to 3 homes worth of electricity
45,000 EVs charging simultaneously (6pm, everyone comes home):
= 135,000 homes worth of demand
= Grid overload, brownouts
Solution: Stagger charging, incentivize off-peak
Managed charging workflow:
EV owner plugs in at 6pm
↓
Agent asks: "When do you need your car charged?"
Owner: "8am tomorrow"
↓
Agent calculates:
- Battery level: 30% (needs 14 hours to fully charge)
- Available time: 14 hours (6pm → 8am)
- Grid conditions: Peak 6-9pm, off-peak 11pm-6am
- Renewable availability: High wind generation 1-5am
↓
Agent schedules:
- Delay charging until 11pm
- Charge during off-peak + high renewable hours
- Finish by 7am (1 hour buffer)
↓
Owner saves money (off-peak rates)
Grid avoids overload
More renewable energy utilized
Results:
- Peak demand reduction: 18% (spreading EV charging)
- Customer savings: $42/month average (off-peak rates)
- Renewable energy utilization: +12% (charging during high wind/solar)
- Grid infrastructure avoided: $280M (no new substations needed)
Future: Vehicle-to-Grid (V2G)
EVs can discharge back to grid during peak demand:
Heat wave → grid stressed → brownout risk
↓
Agent asks EV owners: "Can we use 20% of your battery to stabilize the grid?
We'll pay you $15 and recharge overnight for free."
↓
500 EVs discharge 10 kWh each = 5 MW of grid support
↓
Crisis averted, EV owners earn money
Key insight: AI agents coordinate distributed resources (thousands of EVs) for grid-level benefits.
NERC CIP Compliance for Energy Sector AI
Energy sector has stringent cybersecurity requirements under NERC CIP (Critical Infrastructure Protection).
Key NERC CIP requirements:
CIP-005: Electronic Security Perimeter
- AI agents accessing grid control systems must go through secure perimeter
- Network segmentation (OT network isolated from IT network)
- No direct internet access for critical systems
CIP-007: System Security Management
- All agent actions logged (audit trail)
- Malware protection on agent infrastructure
- Security patching within 35 days
CIP-010: Configuration Change Management
- Any agent configuration change requires approval
- Baseline configuration documented
- Testing required before production deployment
CIP-013: Supply Chain Risk Management
- Salesforce as vendor must meet supply chain security requirements
- Third-party LLMs (OpenAI, etc.) require security review
- Vendor risk assessment updated annually
Penalties for non-compliance:
- $1M per day per violation
- FERC enforcement actions
- Potential loss of grid operating authority
This is why energy sector AI deployments are slow and expensive.
Real-Time AI at Scale: The Technical Challenge
Energy grid operates at millisecond timescales. AI agents must be real-time.
Latency requirements:
Traditional enterprise AI:
- Response time: 1-5 seconds (acceptable)
- Batch processing: hourly/daily updates (common)
Energy grid AI:
- Response time: 10-100 milliseconds (required)
- Real-time processing: continuous streaming data
PG&E’s architecture for real-time AI:
Edge Layer (substations, transformers):
- Local anomaly detection (microsecond response)
- Send alerts to cloud only when anomaly detected
- No round-trip to cloud for routine operations
Cloud Layer (Agentforce):
- Predictive modeling (2-week forecasts)
- Work order generation and crew dispatch
- Historical analysis and pattern recognition
Hybrid approach:
- Edge handles real-time safety (circuit breakers, relays)
- Cloud handles prediction and optimization
- Edge + Cloud communicate via secure VPN
Why not 100% cloud?
- Latency (round-trip to cloud takes 50-200ms)
- Reliability (grid must operate even if cloud connection lost)
- Bandwidth (millions of sensors generate terabytes/day)
Edge AI + Cloud AI = hybrid architecture for energy.
Sustainability Reporting and Carbon Accounting Agents
Energy companies have net-zero commitments and need to track carbon emissions.
Duke Energy’s Carbon Accounting Agent:
What it does:
- Tracks emissions from all generation sources (coal, gas, nuclear, renewables)
- Calculates Scope 1, 2, 3 emissions (GHG Protocol standard)
- Forecasts emissions trajectory toward 2050 net-zero goal
- Recommends operational changes to reduce carbon
Example analysis:
Current emissions: 42.3 million metric tons CO2e/year
Breakdown:
- Coal plants: 65% of emissions (being retired by 2035)
- Natural gas: 28% of emissions (peaker plants)
- Fugitive methane: 7% of emissions (pipeline leaks)
2050 Net-Zero Gap: -42.3 million metric tons
Pathway to net-zero:
- Coal retirement: -27.5 million tons (by 2035)
- Gas plant efficiency: -3.2 million tons (by 2030)
- Methane leak reduction: -2.1 million tons (by 2028)
- Renewable expansion: -6.8 million tons avoided (by 2040)
- Carbon offsets: -2.7 million tons (forestry, CCS)
Agent recommendation: Accelerate coal retirement by 2 years → saves $180M in carbon costs
Regulatory drivers:
- SEC climate disclosure rules (Scope 1, 2, 3 reporting)
- State renewable portfolio standards (RPS)
- Carbon pricing (some states/regions)
- ESG investor pressure
Key insight: AI agents turn sustainability from compliance exercise to operational optimization.
Cost-Benefit Analysis for Energy Sector AI
PG&E’s ROI (24 months, predictive maintenance agents):
Investment:
- Agentforce licenses (NERC CIP compliant): $1.8M/year (300 field engineers + operations center)
- IoT sensor deployment: $45M (one-time, 150K devices)
- Edge computing infrastructure: $12M (one-time)
- Implementation and integration: $8M (one-time)
- Ongoing operations: $2.4M/year
- Year 1 total: $69.2M
- Year 2+ total: $4.2M/year
Returns:
- Avoided outage costs: $58M/year (lost revenue + penalties)
- Maintenance optimization: $42M/year (crew efficiency)
- Wildfire liability reduction: $120M/year (insurance + legal costs)
- Asset life extension: $18M/year (equipment lasts longer)
- Total annual benefit: $238M/year
ROI: 244% in Year 2
Payback: 3.5 months (insanely fast for infrastructure)
Why such high ROI?
- Wildfire liability is massive in California (PG&E paid $13.5B in wildfire settlements 2017-2020)
- Outage penalties can be $10M+ per incident
- Grid infrastructure is expensive (transformers cost $500K-$2M each)
For PG&E, AI agents are existential (prevent bankruptcy from wildfire liability).
Integration with Operational Technology (OT) Systems
Energy AI doesn’t just integrate with IT systems (Salesforce, databases) - it integrates with physical control systems.
Common OT systems in energy:
SCADA (Supervisory Control and Data Acquisition):
- Real-time monitoring of grid equipment
- Remote control of switches, breakers, transformers
- Historical data logging
DMS (Distribution Management System):
- Models the electrical distribution network
- Power flow analysis
- Outage management
EMS (Energy Management System):
- Transmission-level control
- Economic dispatch (which generators to run)
- Frequency regulation
DERMS (Distributed Energy Resource Management System):
- Coordinates solar, wind, batteries, EVs
- Virtual power plant aggregation
Agentforce integration pattern:
Agentforce Agent
↓
Salesforce Platform (IT environment)
↓
MuleSoft Integration Layer (DMZ)
↓ (one-way data flow for security)
OT Historian (read-only access to SCADA data)
↓ (separate write-only channel with approval)
DMS / SCADA (control commands)
↓
Physical equipment (transformers, switches)
Critical security principle:
- Agents can read OT data freely (monitoring)
- Agents can write control commands ONLY with human approval (safety)
- Air-gap between IT and OT networks (prevent cyber attacks)
NextEra Energy learned this the hard way:
Early deployment had agent directly controlling battery discharge. A software bug caused agent to discharge all batteries simultaneously → grid frequency spike → emergency shutdown.
Now: Human-in-the-loop for all control actions.
Workforce Transition: Field Engineers + AI Agents
Energy sector has aging workforce (average age 55+) and knowledge retention challenges.
Duke Energy’s approach:
- Train field engineers to work with AI agents
- Agent provides recommendations, engineer makes decisions
- Capture expert knowledge in agent training data
Example: Substation maintenance
Veteran engineer (30 years experience):
- Hears transformer humming → knows it's oil circulation issue
- Smells ozone → knows it's corona discharge
- Feels vibration → knows bearing is failing
AI agent:
- Analyzes sensor data (temperature, vibration, oil quality)
- Predicts failure probability
- Recommends maintenance action
Human + AI:
- Agent provides data-driven diagnosis
- Engineer validates with physical inspection
- Engineer's expertise improves agent over time
Change management:
- 67% of field engineers were initially skeptical of AI
- After 6 months working with agents: 82% say it improves their work
- Key: Position AI as tool, not replacement
My Recommendation for Energy & Utilities
If you’re in energy sector considering Agentforce:
-
Start with non-control use cases
- Predictive maintenance (monitoring, not control)
- Customer service (billing, efficiency advice)
- Carbon reporting (data analysis)
-
Don’t touch OT control until you have:
- NERC CIP compliance validated
- Cybersecurity maturity model (NIST Cybersecurity Framework)
- Human-in-the-loop approval workflows
- Comprehensive testing (tabletop exercises, simulations)
-
Budget for OT integration complexity
- 2-3x higher than standard enterprise Salesforce
- Specialized OT security expertise required
- Edge computing infrastructure
-
Plan for 18-24 month deployment
- NERC CIP compliance takes time
- OT integration is complex
- Workforce training and change management
-
Focus on safety-critical ROI
- Wildfire prevention (if applicable)
- Outage reduction
- Equipment failure prevention
- These have massive financial impact
The opportunity is enormous (energy sector wastes $100B+/year on inefficiency), but safety and reliability are non-negotiable.
Questions for the Community
-
For other energy/utility folks: How are you approaching OT/IT convergence for AI agents?
-
For Priya (security): How would you design air-gap between IT agents and OT control systems?
-
For Carlos (finance): How do you model ROI when benefits include “avoided wildfire liability” (probabilistic)?
-
For critical infrastructure industries (water, transportation, manufacturing): Similar challenges with physical-digital AI?
I’m speaking at DistribuTECH 2026 on “AI Agents for Grid Modernization” - happy to share more detailed case studies offline.