Why most AI agent pilots on solar sites never reach production

The pitch deck usually arrives before the pilot does. It promises an autonomous agent that forecasts output, flags a failing string before the string fails, and bids the battery into the evening peak while the operations team sleeps. Some asset managers have taken to pasting those decks into an AI detector free of charge before passing them upstairs, which tells you how much trust the category has earned so far.

The skepticism has numbers behind it. RAND found that more than 80% of AI projects never reach meaningful production, roughly twice the failure rate of ordinary IT work. MIT's NANDA study, published last year, traced the funnel more precisely: about 60% of firms evaluate enterprise-grade AI tools, 20% run a pilot, and 5% reach production with measurable impact. Solar carries every one of the usual reasons for that drop-off, plus a few of its own.

The data was never built for an agent

A typical portfolio is the sediment of a decade of acquisitions. Three inverter vendors, two SCADA platforms, a tracker OEM with a documented API and another whose data comes out of a Modbus register somebody mapped by hand in 2019. The pilot never sees any of this. It runs on one well-instrumented site with a historian someone cleaned up for the demo, and it performs beautifully.

Production data behaves differently. It arrives late, with gaps, in formats that were never designed to be read by anything except a human on a night shift. Gartner expects 60% of AI projects without production-ready data to be abandoned through 2026, and renewables sit at the wrong end of that curve. Microsoft's own assessment of the sector points to the same thing: geographically dispersed portfolios running on multiple vendors and legacy systems have made data integration hard to scale, which is why adoption lagged other industries in the first place.

The economics are thinner than the slide implies

Vendors quote real numbers. Forecast accuracy improves by up to 10%. Yield recovery lands in the range of 1% to 3%. Maintenance visits drop, which matters when a single truck roll to a remote site runs into the thousands. BloombergNEF has projected that AI-optimized renewable portfolios capture 12% to 18% more revenue than manually managed assets.

Now put that against a 60 MW plant with a lean O&M contract. Two percent of annual production is a genuine gain, but the integration work, the license, and the engineer who has to babysit the thing for a year all land in the same budget line. Agentic AI spending across the energy sector currently sits near $897 million a year, a rounding error next to what the industry spends on modules. The value case is not fake. It is just small enough that a finance committee can defer it for another quarter, and it frequently does.

Nobody wants to sign the curtailment order

This is the part the deck skips. An agent that trims output at the wrong moment, or misses a fault that a technician would have caught on a walkthrough, creates a liability question with no obvious owner. Is it the vendor, the O&M contractor, the asset manager who approved the pilot, or the lender whose covenant assumes a certain availability?

Governance gets bolted on at the end, when it needs to be there first. Pilots tend to start with a model and finish with a lawyer, and the lawyer usually wins. What survives is a narrower system: an agent that opens a ticket instead of closing a breaker, or drafts a bid that a trader still has to approve.

What the ones that survive have in common

They are unglamorous. They start with a baseline, so somebody can prove afterwards that the pilot did something. They target one workflow with a known cost, usually fault triage or day-ahead forecasting, rather than a portfolio-wide operating brain. They keep a human in the loop on anything that touches the grid.

They also tend to be bought rather than built. MIT's data shows externally partnered deployments succeeding at roughly twice the rate of internal builds, and mid-market firms scaling in about 90 days against nine months for large enterprises. Smaller operators are quietly better at this than the majors, because they cannot afford a hero project.

The window for figuring it out is narrowing. US developers are on track to bring 43.4 GW of new utility-scale solar online in 2026, a 60% jump over last year and more than half of all new generating capacity. Data center demand is climbing from 31 GW in 2025 toward 66 GW in 2027. That is a lot of new plant per available operations engineer, and the arithmetic points one way. The agents are coming to solar because the headcount is not.

Sources

• MIT NANDA, "The GenAI Divide: State of AI in Business 2025"
• RAND Corporation, "The Root Causes of Failure for Artificial Intelligence Projects" (2024)
• Gartner forecast on AI-ready data and project abandonment through 2026
• Microsoft Cloud Blog, agentic AI in renewable energy operations (April 2026)
• BloombergNEF projections on AI-optimized renewable portfolios
• US EIA / pv magazine USA on 2026 utility-scale solar additions and data center load