If you're a CFO at a utility, IPP, or energy company, you've probably approved at least one AI initiative in the last 24 months. And there's a reasonable chance you can't tell anyone — including your board — exactly how much money it's making you.

That's not a unique problem. It's the defining problem of enterprise AI right now. According to a 2023 IDC study sponsored by Microsoft (The Business Opportunity of AI), the average return on enterprise AI investment is $3.50 for every $1 spent. According to MIT's NANDA initiative (The GenAI Divide, 2025), 95% of enterprise GenAI pilots produce no measurable P&L impact, and a 2024 RAND Corporation study finds over 80% of AI projects fail to deliver business value.

Both of those statistics are true at the same time. The companies capturing the $3.50 return are doing one specific thing differently from the 80–95% that aren't. This article explains what that thing is, why it matters specifically for utility operations, and what a CFO should require before approving the next AI line item.

Why traditional AI fails the CFO test

Most utility AI projects today follow the same pattern. A vendor sells the company a predictive analytics platform, an anomaly detection tool, or a dashboard layer. The company integrates it with SCADA, AMI, or asset management systems. The platform generates insights — flagging equipment likely to fail, identifying customers likely to churn, surfacing operational anomalies.

And then those insights get handed back to a human.

This is the failure point. Not the AI. The handoff.

The 42-hour problem

Harvard Business Review research (Oldroyd & Elkington, audit of 2,241 firms) found that companies responding to a web-generated lead within 60 minutes are 7× more likely to qualify the lead as those waiting even an hour later. The average enterprise response time? 42 hours. The same dynamic governs operational signals — every hour a generated insight sits unactioned, its value decays.

Every one of those 42 hours represents:

• Equipment that fails before maintenance is dispatched
• Customers who churn before retention can act
• Compliance windows that close before remediation begins
• Revenue opportunities that disappear into a competitor's pipeline

The dashboard saw the problem. The team didn't have the bandwidth to act on it. The cost is real, recurring, and rarely measured.

Source: Harvard Business Review — "The Short Life of Online Sales Leads" (Oldroyd & Elkington)

The hidden cost of unplanned events

In utilities and energy specifically, the cost of slow response is staggering. According to International Energy Agency analysis, AI-enabled optimization in power operations could deliver up to $110 billion in annual cost savings by 2035. The same body estimates that unplanned equipment failures cost utility operators an average of $2.4 million per event — and that the warning signals for those failures are typically present in operational data 24 to 72 hours before the failure occurs.

Read that again. The data was already there. Nobody acted on it in time.

What agentic AI actually does differently

"Agentic AI" is a term that's being used loosely in 2026, so let's be precise. Agentic AI systems are software that:

• Observe — continuously monitor operational data across SCADA, AMI, ERP, billing, CRM, and field operations
• Reason — synthesize signals across multiple systems to determine the most valuable next action
• Act — execute that action through existing systems: work orders, dispatch, alerts, customer communication, CRM updates
• Learn — improve their action quality based on outcomes, building a model of the specific business they're operating in

The critical word is act. Predictive analytics platforms observe and reason. They don't act. The action is left to the human, which is where most AI ROI is lost.

A concrete example

A 2GW combined-cycle gas turbine operator deployed a predictive maintenance platform in 2023. The platform detected vibration anomalies on a critical compressor 36 hours before failure. The alert went to a maintenance manager's email queue. The manager was on PTO. The turbine failed. Total cost: $4.1 million in unplanned downtime plus emergency parts.

In an agentic deployment, the same anomaly detection would have:

• Generated the work order automatically inside the existing CMMS
• Routed it to the on-call maintenance engineer (not the manager on PTO)
• Pre-staged the parts in inventory based on the failure pattern
• Logged the entire decision chain for audit and compliance
• Notified operations leadership only if escalation was needed

Same data. Same insight. Different outcome — because the system took the action.

The CFO's framework: five questions before approving any AI initiative

If you're evaluating an AI vendor for utility operations — whether it's a predictive maintenance tool, a customer service AI, or an operational optimization platform — these are the five questions that separate ROI-producing deployments from dashboard collections.

1. "What action does the AI take, and through which system?"

If the answer is "it generates an alert" or "it produces a dashboard" — that's not agentic AI. That's analytics with a chatbot. The vendor needs to be able to name the specific systems where action gets executed: the CMMS, the CRM, the dispatch platform, the customer notification system. "It surfaces an insight" is the wrong answer.

2. "What's the measured time-to-action today, and what will it be after deployment?"

The CFO test isn't "does the AI work?" It's "how many hours of delay are you eliminating between detection and resolution?" If the vendor can't quote a specific time-to-action improvement — measured in hours, not percentages — they don't actually know whether their system creates ROI.

3. "What's the success metric, and is it agreed in writing before the engagement starts?"

Real agentic AI deployments commit to a specific business outcome — recovered revenue, avoided downtime, reduced O&M cost — measured against a baseline, agreed in writing, with a 90-day measurement window. If the vendor can only commit to "AI capabilities" or "deployment milestones," walk away. They don't believe in their own product.

4. "How does the system govern human-in-the-loop on sensitive decisions?"

Agentic AI doesn't mean unsupervised AI. For pricing, contracts, regulatory filings, and decisions with safety implications, the system should escalate to a human — but document the recommended action, the reasoning, and the supporting data so the human can decide in seconds, not hours. If the vendor can't show you the escalation governance model, they don't have one.

5. "What's the data moat we build by deploying you?"

This question matters for long-term ROI. Every action a well-designed agentic system takes should train a model on the specific patterns of your business. Twelve months in, the system should know your assets, your customers, and your operational patterns better than any new competitor could learn in three years. If the vendor can't articulate the compounding intelligence model, you're paying for software — not a strategic asset.

The free audit model: how to test agentic AI without committing capital

One specific aspect of how agentic AI vendors are selling in 2026 is worth highlighting: the free audit model.

Traditional enterprise AI sales cycles run 6 to 12 months — vendor demos, proof-of-concept negotiations, IT integration scoping, security review, procurement, and finally a paid pilot. Cost to the buyer: 200+ internal hours and significant calendar drag before any value is delivered.

The agentic AI vendors capturing the most ground in 2026 are inverting this. The model:

• Two-week diagnostic engagement, fully funded by the vendor
• CEO-level conversation, not vendor sales pitch
• Vendor maps actual revenue leaks or operational waste using the buyer's data
• Output: a dollar figure — specifically how much addressable value AI can capture in the first 90 days
• Buyer decides whether to proceed with no commitment beyond the audit

This works because confident vendors with proven systems can afford to give away the diagnostic. Underqualified vendors can't, because their products produce ambiguous results that don't survive a focused two-week test.

As a CFO, the free audit is the cheapest possible way to evaluate an AI vendor. If the vendor refuses to do one — or charges $50K+ for a "discovery engagement" — you've learned something useful about their conviction.

What this looks like in practice: three deployment patterns

Three patterns are showing measurable ROI in utility operations today:

Pattern 1: Operational signal → automated work order

Agentic system monitors SCADA, AMI, and asset management data continuously. When patterns indicate likely equipment failure, the system creates a work order in the CMMS, routes to the appropriate field crew, pre-stages parts, and logs the entire decision chain. Average time-to-dispatch reduction: 18-24 hours. Typical 12-month value: $1.2M to $4.5M per major asset class for mid-sized utilities.

Pattern 2: Customer churn signal → automated retention outreach

System monitors billing patterns, service issues, customer service interactions, and external indicators (move-out filings, business closure signals). When churn risk crosses threshold, system triggers retention workflow — personalized communication through the CRM, account manager notification with full context, optional rate-plan analysis pre-built. Typical reduction in voluntary churn: 18-30%. For a 500K-customer utility with average annual customer value of $1,200, that's $11M-$18M of preserved revenue annually.

Pattern 3: Regulatory signal → automated compliance documentation

System monitors operational data, customer interactions, and field activity for events with compliance implications (outage durations, response times, environmental thresholds). When events cross reporting thresholds, system pre-builds the compliance filing with full supporting documentation, routes to the compliance officer for review, and flags any escalations. Reduces compliance officer time on routine filings by 60-70%, eliminates late-filing penalties, and dramatically reduces audit response time.

The 2026 timing question

Gartner's Q4 2025 AI Spending Forecast projected the agentic AI subsegment growing from $15B in 2025 to $753B by 2029 — a 50× expansion in four years, at a 118.7% compound annual growth rate.

That's not a soft macro number. It's specifically the agentic-action layer of enterprise AI — the layer that takes work, not the layer that produces dashboards. It's growing roughly 8× faster than the broader enterprise software category.

For a CFO making capital allocation decisions, two implications matter:

First: companies deploying agentic AI today are building a 2-year compounding data advantage. Every action the system takes trains a model on the deploying company's specific operations. A competitor entering the same vertical in 2028 starts from zero. The deploying company has 24 months of compounded operational intelligence. That gap is unrecoverable.

Second: vendor selection now matters more than vendor selection in any prior enterprise software wave. Vertical AI is consolidating quickly — Sierra hit $100M ARR in 7 quarters, Harvey raised $400M+ in legal, Hippocratic AI raised $402M in healthcare. The pattern: in any vertical, one or two vendors capture most of the market within 24-36 months. Choosing the right one early is worth more than negotiating a 10% better price on the wrong one.

What to do this week

Three concrete actions for a CFO evaluating agentic AI for utility operations:

• Audit your current AI spending. List every AI project, predictive analytics platform, and "intelligence" tool currently funded. For each, ask: "What action does this take, through which system, with what time-to-resolution?" If the answer for any line item is "it produces a dashboard," you've identified an AI investment that's unlikely to produce CFO-defensible ROI.
• Request a free audit from one agentic AI vendor in the next 30 days. If the vendor refuses, charges for it, or routes you to a paid POC before the diagnostic — you've learned something useful. If the vendor accepts and delivers a specific dollar figure within two weeks, you have a quantified business case to bring to your operations team.
• Set the success-metric standard for any 2026 AI engagement. Before signing any AI contract, require that the vendor commit to a specific business outcome (recovered revenue, avoided downtime, reduced O&M cost) measured against a documented baseline, with a 90-day measurement window and a contractual exit if the metric isn't hit. Vendors who push back on this standard are telling you something.