Positive AI for Utilities: Ethics, Grid, Procurement

Utilities don’t have a “should we use AI?” problem anymore. They have a “what kind of AI are we willing to live with?” problem.

The public mood around AI has turned sour for understandable reasons: deepfakes, misinformation, concentrated corporate power, messy labor practices in data work, and the very real energy footprint of large-scale computing. That same skepticism shows up inside technical communities too—many researchers now see AI as a net negative unless it’s carefully steered.

Here’s the stance I’ll defend: energy and utilities are one of the few places where a positive vision for AI can be concrete, testable, and genuinely pro-social—because the sector is measured in reliability metrics, safety incidents, outage minutes, emissions intensity, and audited procurement spend. If you want AI to help society, put it to work where outcomes are measurable and accountability is non-negotiable.

This post sits in our AI in Supply Chain & Procurement series, so we’ll connect that bigger “positive vision” idea to the day-to-day reality of utility procurement: sourcing transformers and switchgear, managing vendor risk, forecasting demand for spares, and keeping crews stocked during storm season.

Why a “positive vision for AI” matters in utilities

A positive vision isn’t a marketing slogan. It’s a set of design constraints and success criteria—who benefits, who bears the risk, and how you prove it.

Utilities are critical infrastructure. When AI goes wrong here, it doesn’t just create bad recommendations—it can:

• Mis-prioritize asset inspections and increase failure risk
• Trigger poor switching decisions during restoration
• Distort load forecasts and raise balancing costs
• Create procurement blind spots that delay lead-time items

The upside is equally tangible. Done responsibly, AI can reduce outages, reduce costs, and reduce emissions—while improving planning and procurement resilience.

A useful test: If your AI project can’t be tied to reliability, safety, affordability, or decarbonization metrics, it’s probably not a “positive vision” project. It’s just experimentation.

The credibility gap: why scientists (and operators) are skeptical

Many researchers and practitioners are wary because AI has been paired with the wrong incentives:

• Speed over safety: shipping models before failure modes are understood
• Scale over stewardship: training bigger systems without clear justification
• Centralization over resilience: over-reliance on a few vendors and opaque tooling
• Cost shifting: hiding labor and environmental impacts in the supply chain

Utilities feel these tensions directly. They’re asked to “innovate with AI” while also meeting strict reliability and cybersecurity standards—and while facing long lead times for critical equipment.

So the question becomes: How do you pursue AI that improves the grid without importing the worst habits of the broader AI ecosystem?

What “AI for the public good” looks like in energy systems

A positive vision becomes practical when it’s tied to operational outcomes and governance. Below are areas where I consistently see AI provide legitimate value—especially when paired with disciplined procurement and risk controls.

Grid optimization that operators can audit

The goal isn’t to replace dispatchers or grid operators. The goal is decision support that’s explainable enough to trust under pressure.

Examples that tend to work in production:

• Load and renewable generation forecasting that incorporates weather ensembles and local conditions
• Volt/VAR optimization recommendations constrained by equipment limits
• Outage prediction and restoration triage using vegetation, asset health, and storm path data

The practical requirement: every model output needs a traceable reason and a safe fallback. In control rooms, “the model said so” is not a reason.

Predictive maintenance with procurement baked in

Most predictive maintenance programs stall at the same point: they predict failures but don’t change outcomes because parts, contractors, and outage windows aren’t aligned.

A positive vision for AI in maintenance includes procurement from day one:

• Forecast failure risk for components (e.g., transformers, breakers, reclosers)
• Convert risk into time-phased material requirements for spares
• Trigger sourcing actions when lead times threaten reliability targets
• Prioritize work orders using cost-of-failure and customer impact

Predictive maintenance without supply chain integration is just better reporting.

AI for supply chain & procurement resilience (where it really pays)

Utilities have been living through equipment constraints and volatile pricing for years. AI can help, but only if it’s aimed at the right problems.

High-value procurement use cases:

• Demand forecasting for spare parts using outage history, planned work, and seasonal patterns
• Supplier risk scoring that blends delivery performance, quality events, financial health indicators, and geopolitical exposure
• Spend analytics and contract compliance to find leakage and standardize specifications
• Should-cost modeling for categories like wire & cable, poles, switchgear, and fleet maintenance
• Scenario planning to answer “what if a top supplier slips by 12 weeks?”

This is also where ethical AI gets real: if an algorithm deprioritizes a community for restoration-related materials, you need governance that catches it.

Four actions utilities can take to steer AI toward better outcomes

The RSS article’s core idea is that scientists shouldn’t give up; they should actively steer AI toward beneficial outcomes. In utilities, that steering translates cleanly into four actions.

1) Build and enforce ethical norms (not just principles)

Ethics statements don’t prevent incidents. Operational norms do.

For utilities and energy procurement teams, that means defining requirements like:

• Data provenance: which systems feed the model, who owns them, and how they’re audited
• Model accountability: named owners for performance, drift, and incident response
• Human override: clear authority for operators/buyers to reject outputs
• Equity and service impact checks: especially for outage response, disconnection risk, and ratepayer impacts

A practical tool I like: an “AI go/no-go checklist” embedded in the sourcing process, so vendors must demonstrate controls before contracts are signed.

2) Resist harmful uses by documenting them early

Utilities are tempted to use generative AI wherever there’s text: call center scripts, employee monitoring, customer risk profiling, procurement negotiations. Some of these are fine. Some are reputational landmines.

Resisting harmful uses doesn’t mean banning AI. It means:

• Running red-team exercises on customer-facing and operator-facing workflows
• Testing for failure modes: hallucinations, bias, prompt injection, data leakage
• Defining “never do” zones (for example, automated disconnection decisions without strong safeguards)

If you can’t write down the plausible misuse cases, you’re not ready to deploy.

3) Use AI responsibly where it measurably improves lives

The fastest way to earn trust is to ship AI that helps real people.

In energy, that often looks like:

• Reducing truck rolls through better diagnostics
• Shortening outage duration by improving triage and crew routing
• Cutting procurement cycle time for low-risk buys while tightening scrutiny for critical equipment
• Improving interconnection processing for distributed energy resources

Responsible AI is not “perfect AI.” It’s AI paired with guardrails, transparency, and monitoring.

4) Renovate institutions: procurement is where governance becomes real

Most AI governance collapses at the vendor boundary. The model might be “ethical,” but the contract isn’t.

Utilities can renovate their institutions by modernizing procurement and supplier governance:

• Add AI-specific clauses: training data disclosures, incident reporting SLAs, audit rights
• Require model cards or equivalent documentation for high-impact systems
• Define cybersecurity and data handling requirements aligned to critical infrastructure standards
• Set performance-based acceptance criteria tied to reliability, safety, and cost

If procurement doesn’t ask the questions, nobody will.

Practical blueprint: turning “positive AI” into a procurement-ready program

Here’s a procurement-centric approach I’ve seen work because it respects how utilities actually operate.

Start with one high-friction workflow

Pick a process where delays or errors have clear costs:

• Long-lead equipment expediting
• Spare parts stocking for storm season
• Contractor qualification and safety documentation
• Invoice exception handling

You’re looking for something with enough volume to matter, but not so critical that experimentation creates unacceptable risk.

Define success metrics before you train anything

Utilities are great at metrics—use that strength. Examples:

• Reduce stockouts for critical spares by X%
• Improve on-time-in-full delivery by X points
• Cut PR/PO cycle time by Y days
• Reduce emergency buys and premium freight by $Z

Make sure metrics aren’t only cost-based. Reliability and safety deserve equal billing.

Use a “constrained AI” mindset

In critical infrastructure, the best models are often the least flashy:

• Constrain recommendations to approved suppliers and specs
• Require citations to contracts, standards, or prior performance
• Limit actions to “suggest” rather than “execute” for high-impact steps

Generative AI can help draft supplier communications or summarize contracts, but procurement decisions should remain accountable and reviewable.

Build the monitoring loop

A positive vision includes long-term care:

• Drift monitoring (supplier performance changes, new materials, new failure patterns)
• Feedback capture from buyers and field teams
• Post-incident reviews when AI contributes to a bad call

If you can’t monitor it, you can’t responsibly run it.

Common questions (and straight answers)

Is generative AI actually useful in utility procurement?

Yes—for summarization, classification, document workflows, and guided sourcing support. No—as an autonomous negotiator or an automatic decision-maker for high-stakes buys.

Won’t AI increase energy demand and contradict utility climate goals?

It can, which is why efficiency and right-sizing belong in the business case. Favor smaller models, targeted automation, and clear ROI per compute hour. If the AI workload is large, treat it like any other load: plan it, optimize it, and account for it.

What’s the fastest way to reduce AI risk with vendors?

Put it in the contract. Audit rights, incident reporting, data handling, and performance acceptance criteria matter more than vendor slide decks.

Where this fits in the AI in Supply Chain & Procurement series

Most companies get distracted by shiny AI demos. Utilities don’t have that luxury. Your supply chain is part of grid reliability, and AI decisions in procurement can either reinforce resilience or quietly undermine it.

A positive vision for AI in energy and utilities is simple: use AI where it measurably improves reliability, affordability, and decarbonization—and govern it like the critical system it is. If that sounds strict, good. Critical infrastructure should be strict.

If you’re planning AI initiatives for 2026 budgets right now, start by mapping two things: the procurement workflows that most threaten reliability, and the AI controls you’d need to sleep at night after deployment. What would it take for your team to say, “Yes, this makes the grid better,” and mean it?