where AI actually pays off

a practical filter for finding AI work worth funding

Most AI portfolios fail because teams rank by excitement instead of yield. Digital twins win the executive vote and rarely deliver. Companies spend millions building perfect systems before anyone sees value, on business cases that promise better decisions or improved agility and fall apart the moment budgets tighten.

The approach here works the other way around. Start with the workflow, find the bottleneck, and only then ask whether AI is the right tool.

start with bottlenecks, before platforms

The first question is rarely “Copilot, a custom model, or something else.” It is “where does the actual business friction sit.” Map the workflows and the data first. The tool choice falls out of that, once you know what is slow, manual, or invisible.

separate the problem type

Some candidates are genuine AI problems. Others are process cleanup, a data foundation that needs building, a reporting gap, or a governance issue wearing an AI costume. Naming the real problem type up front saves the budget that would otherwise vanish into the wrong fix.

fund narrow, measurable pilots

The strongest starting point is one or two bounded use cases where value can be tested quickly and safely. Broad “AI strategy” decks burn budget. A narrow pilot with a clear metric compounds.

keep human judgment explicit

AI can organize, summarize, flag, compare, and draft. The important business decisions stay owned by the right people. Designing that line at the start keeps the system trustworthy and keeps accountability where it belongs.

six questions for every candidate

For each candidate area:

• What workflow or decision is slower, more manual, or less visible than it should be?
• What is the business value if it improves?
• What data exists today, and who owns it?
• Where does human judgment need to stay?
• What is the smallest useful test?
• What result would make us continue, adapt, or stop?

Those six questions filter “AI for AI’s sake” out fast and leave the candidates where better workflow, data, or decision support creates measurable value.

the discipline for the ones that pass

A candidate that clears the filter earns a real business case, scored on eight dimensions that force honesty before any capital moves:

• Specific initiative. “Predictive maintenance for critical pumps and conveyors” is an initiative. “AI for operations” is too vague to fund. If you cannot write the plan, you do not understand it well enough to pay for it.
• Phased plan. Pilot, integrate, scale, pattern.
• Decision gate. The date you need evidence to kill, pivot, or scale. Every initiative gets one. No zombie projects.
• Computable ROI. Benefit in plain math where every variable is measurable: avoided downtime hours times profit per hour, plus labor saved, minus program cost, every term pulled from real data. A digital twin equation that reads improved decision quality is unfundable.
• Data friction. Model and inputs surface what is missing. A forecast that needs 18 months of history when you have 6 just doubled its timeline.
• ROM cost and benefit. Is this a $500K test or a $5M one, and what is steady-state value at scale.
• Scope discipline. Name what stays in phase 1 and what gets deferred. Every project starts too broad.

Run the rubric and the ranking sorts itself. Predictive maintenance with computable downtime savings and a six-month gate beats a digital twin with vague ROI and an eighteen-month proof. Scope discipline is what moved one digital-twin case from eighth priority to fourth: narrow it to a single high-energy unit, link it to process controls, prove value in twelve months.

Some “wins” die on contact with reality. Autonomous blocking for OT cybersecurity sounded impressive until the reality check: autonomous action can shut down a plant. That feature got killed and replaced with monitoring plus a human approval step.

where to start in an energy portfolio

Applied to an energy operator, three areas tend to carry the most value with the least risk:

• Screening inbound opportunities against known technical and commercial guardrails, so weak deals get filtered before they consume diligence time.
• Identifying and explaining underperforming generation assets, turning scattered operating data into a ranked, explainable view of what is dragging and why.
• Rolling up portfolio operating data into a comparable corporate view, so the same metric means the same thing across every asset.

Map those bottlenecks, test the data that actually exists, then decide whether the right next step is AI assistance, reporting, process redesign, or data cleanup. The answer is often one of the last three, and that is the point.

bottom line

Rigorous beats strategic. Computable beats transformational. Phase 1 proof beats enterprise vision. Start with the bottleneck, separate the problem type, fund the narrow test, and keep judgment where it belongs. That is how AI work gets funded and delivers, instead of drifting into zombie status or dying at the first budget review.

The longer argument for why this gap stays open: easy alpha.