A homeowner in Washtenaw County, Michigan, got an energy audit last year. Her auditor spent three hours crawling through the attic, running a blower door test, scanning walls with an infrared camera, and marking up a floor plan. Eleven recommendations came back ranked by payback period, from air sealing at $800 (pays for itself in fourteen months) to a ground-source heat pump at $28,000 (twenty-two-year payback). Total cost of the audit: $475, and she did the air sealing and stopped. Everything else sits in a kitchen drawer.
That drawer is the retrofit problem in miniature. Not whether the technology exists, and not whether the savings are real. Whether anyone can tell a homeowner what to do first, in what order, and why, at a price that doesn't consume the savings before they start.
A team at Michigan State University just ran the most rigorous test yet of whether AI can fill that role, and the answer splits cleanly in two: the diagnosis is excellent, and the prescription is wrong.
What MSU Actually Tested
Researchers led by Lei Shu, Armin Yeganeh, Sinem Mollaoglu, Jiayu Zhou, and Dong Zhao queried six leading AI models with building-specific information drawn from ResStock 2024.2, the National Renewable Energy Laboratory's dataset of U.S. residential building prototypes covering forty-nine states. Each model received descriptions of actual homes, including construction type, insulation levels, window specifications, HVAC systems, and climate zone, and was asked to do two things: identify which retrofit would produce the maximum CO2 reduction, and identify which would deliver the shortest discounted payback period.
Six models were tested: ChatGPT o3, DeepSeek R1, Grok 3, Gemini 2.0, Llama 3.2, and Claude 3.7.
All six performed well on the first task. Ask an AI chatbot what's wrong with a 1,400-square-foot ranch in climate zone 5A with R-13 wall insulation and a fifteen-year-old gas furnace, and it will correctly tell you the furnace is the biggest carbon problem. Reliable and consistent across models. It reads the building description, identifies the component with the largest emissions footprint, and recommends replacing it.
Now ask the same model which retrofit pays back fastest.
Different question entirely.
The models struggled. Not because they couldn't identify legitimate retrofits. They could. Every model returned measures that would, in fact, save energy and reduce emissions. But when the question shifted from "what reduces CO2 the most?" to "what pays for itself the fastest?", the models started recommending expensive deep retrofits ahead of cheap quick wins. They confused the biggest carbon reduction with the best investment, and those are not the same thing at all, because a $25,000 heat pump that saves $1,200 per year in heating costs is a worse first investment than $900 in attic air sealing that saves $650 per year, even though the heat pump saves considerably more carbon over its full lifetime.
As the researchers put it directly: the AI models provide effective retrofit options but had difficulty determining the best one.
The Ordering Problem
Why does the order matter? Because homeowners don't retrofit all at once, and ACEEE surveyed 1,500 U.S. homeowners and found that 65 percent could invest $1,000 in energy upgrades. Not $15,000. Not $40,000. A thousand dollars, which buys exactly one small retrofit done well. If that one retrofit is the wrong one, not wrong in the sense that it doesn't work, but wrong in the sense that something cheaper would have paid back faster and freed up capital for the next upgrade sooner, the homeowner has spent money correctly but in the wrong sequence, and that sequencing error compounds over years.
Run the math on a homeowner with $12,000 to spend across three retrofits over five years. If she does them in optimal payback order (air sealing first at $900 with a fourteen-month payback, then attic insulation at $2,800 with a four-year payback, then a heat pump water heater at $3,500 with a six-year payback), the cumulative savings by year five total roughly $4,900. If she does them in the order a general-purpose AI recommends, prioritizing by carbon impact instead (heat pump water heater first, then insulation, then air sealing), cumulative savings by year five drop to approximately $2,700. Same upgrades. Same total cost, different order, and a $2,200 difference in realized savings in just five years that widens every year after because the early payback accelerates spending on the next measure.
That is not a rounding error; that is a family's grocery budget for four months.
A Model That Gets the Order Right
The same MSU group built something that does what the chatbots cannot. Shu and colleagues fine-tuned a large language model using Low-Rank Adaptation (LoRA) on physics-based energy simulations and techno-economic calculations derived from those 536,416 ResStock building prototypes, covering nine major retrofit categories. None of that training data was scraped from the internet or assembled from manufacturer spec sheets. It came from calibrated energy models that simulate how real buildings in real climates respond to specific interventions, with cost data layered on top to calculate discounted payback periods for every combination of building, climate, and retrofit.
Results are striking. In 98.9 percent of cases, the fine-tuned model identifies the optimal retrofit for CO2 reduction within its top three recommendations. For shortest discounted payback period: 93.3 percent. Fine-tuning produced an order-of-magnitude reduction in CO2 prediction error compared to the general-purpose models, and multi-fold reductions in energy use and retrofit cost prediction errors. It performs reliably even when the homeowner's description is incomplete, missing a detail about window type or uncertain about existing insulation thickness, because the training data includes enough building archetypes to interpolate from partial information.
Crucially, the fine-tuned model was evaluated against physics-grounded baselines, not against other AI models. Its benchmark is not "better than ChatGPT." It's "agrees with the thermodynamic simulation." A different and harder standard. Passing at 93 percent for payback and 99 percent for carbon means the model has internalized something the general-purpose chatbots have not: the relationship between energy physics, construction costs, and local utility rates that determines whether a $3,000 investment pays back in three years or thirteen.
What the Chatbots Are Missing
General-purpose AI models are trained on text. They know that insulation reduces heat loss because they've read thousands of articles about insulation reducing heat loss. They know heat pumps are more efficient than gas furnaces because the internet says so, repeatedly. What they don't have is a calibrated physical model of how a specific 1,800-square-foot colonial in climate zone 4A with double-pane low-e windows and an uninsulated crawl space responds to a specific intervention at current local energy prices.
Consider the difference: knowing that exercise is good for you is not the same as knowing that, given your specific blood pressure, weight, and medication regimen, a 30-minute daily walk will reduce your systolic pressure by 8 mmHg within six weeks while HIIT training will spike it dangerously for the first month. Both the general advice and the specific advice are "correct." Only one is safe to act on without supervision.
Energy retrofits have the same structure as medical advice. A heat pump is almost always the correct long-term answer for decarbonizing a home's heating system. But recommending a $12,000 cold-climate heat pump to a homeowner whose ductwork leaks 35 percent of its airflow into an unconditioned attic is recommending a high-performance engine for a car with no tires. Fix the ducts first. A chatbot doesn't know that, because it doesn't model duct losses as a function of attic temperature differential; it just knows that heat pumps are good.
The Audit Gap
Professional energy audits cost between $250 and $650 for a Level 2 assessment, which is what most homeowners need: a detailed walkthrough with blower door testing, infrared scanning, and a prioritized recommendation list. Nationally, that averages about $437, according to Angi's 2026 data. That is not expensive for the value delivered, and for any individual homeowner considering a $10,000-plus retrofit, paying $437 for a professional opinion is straightforward arithmetic.
But the federal tax credit that used to reimburse up to $150 of that cost expired on December 31, 2025, eliminated in the One Big Beautiful Bill Act. And the bottleneck was never really the money; it was the experts. Certified energy auditors earn an average of $21.50 per hour, according to PayScale's 2026 data. At that wage, the pipeline of new auditors entering the field is not keeping pace with the demand created by IRA rebate programs, state weatherization pushes, and a residential energy-efficient technology market that, according to IEA estimates, exceeded $200 billion globally in 2026.
This is the gap that AI could fill. Not as a replacement for the auditor, but as a first-pass screening tool that tells you whether you need one at all. A homeowner who can describe her house in three sentences and get a ranked list of retrofits with approximate payback periods, accurate to within 7 percent of what a physics simulation would produce, doesn't need to pay $437 to learn that her biggest quick win is air sealing. She needs the auditor for the $15,000 decision, not the $900 one.
What Does Not Exist Yet
MSU's fine-tuned model is a research prototype, not a consumer product. You cannot download it, and no company has licensed it. Its paper (arXiv:2602.20181) describes the architecture and the results but does not release the model weights. The training pipeline uses ResStock data that is publicly available from NREL, and the fine-tuning approach (LoRA on an existing LLM backbone) is well-understood, which means a startup or a utility company or a government agency could replicate this work. Nobody has, and that absence matters because right now, millions of homeowners are asking ChatGPT and its competitors what to do about their energy bills. Those models give answers that sound authoritative, cite real technologies, and recommend measures that genuinely work. They just can't tell you which one to do first, and for a homeowner with limited capital, that's the only question that actually matters.
What You Should Do Right Now
If you're considering a retrofit costing more than $3,000, pay for the audit. That $437 buys you a prioritized list based on your actual house, not on what the internet thinks a house like yours probably looks like. That list saves you from the sequencing error. If the audit shows air sealing as the top priority, do it before you buy the heat pump, even if the heat pump is the bigger carbon win, because the air sealing pays back in fourteen months and frees up the savings to help fund the heat pump two years sooner than you'd otherwise afford it.
If you're considering a retrofit under $1,000, the chatbot is probably fine. Air sealing, LED lighting, smart thermostat installation: these are measures where the general AI recommendation aligns with the optimal recommendation because the cost is low enough that payback order barely matters. Where chatbots fall short is ranking, not identification, and ranking only hurts you when multiple expensive options compete for limited capital.
If a consumer product based on the MSU approach reaches the market, it could collapse the cost of that initial screening from $437 to near zero, and the 93 percent accuracy on payback ranking means it would agree with a professional auditor's priority list in nineteen out of twenty cases. That changes the math for the 35 percent of homeowners whom ACEEE found cannot invest even $1,000, because it tells them which single, cheapest, fastest-payback measure to pursue first, and it does so without a home visit, a blower door, or an invoice.
Diagnostic technology for houses at scale exists, and prioritization technology that ranks those diagnoses correctly exists as a research prototype. What doesn't exist is a product, and someone should build it.
Limitations
MSU's general-purpose evaluation tested six LLMs using 2024-era models and ResStock 2024.2 data. Model capabilities improve quarterly; results could differ with newer versions. Their fine-tuned model was evaluated against its own training data source, and independent validation on field-measured retrofit outcomes has not been published. Payback calculations assume stable energy prices and current utility rate structures; in states with time-of-use pricing or rapid rate changes, realized paybacks will diverge from predictions. Our $2,200 sequencing cost estimate uses illustrative retrofit costs and savings rates and should be treated as directional, not as a universal figure. Actual homeowner outcomes depend on local energy prices, contractor availability, incentive programs, and building-specific conditions that no model, general or fine-tuned, captures without site-specific input.