A couple in Somerville, Massachusetts closed on a 1,920-square-foot colonial in 2022, negotiated $12,000 off the asking price for a failing water heater, and moved in without ever learning that their home hemorrhages roughly $1,400 per year in preventable energy waste through uninsulated rim joists and single-pane basement windows. They are not unusual, and they are not unlucky. According to a Resources for the Future survey of 1,784 homeowners, only 4 percent had undergone a recent energy audit. An RMI analysis puts the formal assessment rate even lower: fewer than 3 percent of all US single-family homes have ever received a DOE Home Energy Score or RESNET HERS rating.
That is 127 million homes flying blind on energy performance, and nobody ever told them they were losing money.
Four distinct AI approaches now claim they can fill that gap without sending anyone to your front door. I compared them to the in-person gold standard to figure out which ones are worth your time and which are expensive-sounding ways to tell you nothing useful.
What Each Method Actually Measures
Public-data algorithms (ClearlyEnergy, UtilityScore) pull property records, tax assessor data, local climate norms, and utility rate schedules. They estimate what a home should consume based on its age, square footage, and construction type, with nationwide coverage at zero cost to the homeowner. RMI tested these against DOE Home Energy Score assessments on roughly 8,000 homes in 27 states and found them "accurate enough" for directional guidance, though the study did not publish a single summary accuracy metric, which tells you something about the spread.
Smart thermostat data mining uses Nest or Ecobee temperature logs, local weather history, and utility bills to reverse-engineer your home's thermal envelope. A peer-reviewed study in MDPI Energies validated this approach on homes with WiFi thermostats and found it could predict attic R-values with 94 percent R-squared accuracy and furnace efficiency with 95.4 percent R-squared. Those numbers are remarkable, but the sample was small and the method requires hardware that fewer than half of US households own.
Street View computer vision skips the interior entirely. Researchers at Notre Dame, Maryland, and Utah trained an AI to analyze passive design features from Google Street View images across 1,402 census tracts in Chicago, covering nearly 300,000 households. The model predicted energy costs with 74 percent accuracy when combined with demographic data, which is impressive for a system that never sets foot inside a building and completely inadequate for a homeowner trying to decide whether to replace their furnace or insulate their attic first. It maps neighborhoods. It does not diagnose houses.
In-person DOE Home Energy Score (HES) remains the benchmark. A qualified assessor spends two to four hours in your home with a blower door, an infrared camera, and an energy modeling tool, all for $150 to $400 depending on the market. Accuracy: the highest available, because the assessor sees what algorithms guess at. Availability: essentially nonexistent at population scale, because there are roughly 2,400 active HES assessors in a country with 130 million single-family homes.
A Comparison Nobody Has Published
| Method | Input Required | Cost | Best Accuracy Reported | Homes Reachable |
|---|---|---|---|---|
| Public-data algorithm | Address only | Free | Directional (no single metric published) | ~130 million |
| Smart thermostat mining | Thermostat data + bills | Free | 94% R² on R-values | ~60 million |
| Street View CV | Exterior photo | Free | 74% energy cost prediction | Any photographed home |
| In-person HES | On-site inspection | $150 to $400 | Gold standard | ~3.9 million assessed to date |
Run the math on total savings identified, not per-home accuracy. If a free algorithm reaches 100 million homes at 70 percent accuracy, it flags roughly 70 million homes with directionally correct upgrade recommendations. If an in-person audit reaches 3.9 million homes at 95 percent accuracy, it flags 3.7 million. Nineteen to one. Imprecise, widespread awareness beats precise, rare expertise on aggregate energy reduction, which is the metric that actually matters for the climate.
What the AI Cannot See
Carbon monoxide from cracked heat exchangers, backdrafting combustion appliances venting exhaust into living spaces, mold colonies growing behind vapor barriers in bathrooms, gas leaks at furnace connections. Algorithms cannot smell gas. These are safety hazards that kill people, and no model estimating your R-value from tax records and climate averages will detect a cracked heat exchanger or a blocked flue pipe or a patch of black mold behind a bathroom wall that has been slowly poisoning a family for three years while their energy dashboard showed a B-plus. This is the strongest case against remote audits: they create a false sense of completeness that discourages the in-person visit capable of catching life-threatening problems that have nothing to do with energy efficiency and everything to do with whether the occupants wake up tomorrow morning.
Then there is the action gap, which may be the more damning problem. A Resources for the Future study found that even among homeowners who received professional in-person audits, only 30 percent implemented the recommended upgrades. A vaguer recommendation from an algorithm, delivered as a web dashboard rather than a face-to-face conversation, may produce even less action. The audit is not the bottleneck, and it never was; motivation is, and no amount of algorithmic sophistication can fix the gap between knowing your attic needs insulation and actually calling a contractor to install it.
What You Should Actually Do
Start with the free tools. ClearlyEnergy generates a report from your address in under a minute. If your utility offers smart meter analysis (PG&E's HomeIntel claims $350 per year in identified savings), use it. These cost nothing, take five minutes, and will not tell you everything, but they will tell you whether your home is roughly average, better than average, or dramatically worse than average for its size, age, and climate zone. That alone is more information than 97 percent of homeowners currently have.
If the free tools flag your home as significantly underperforming, book the in-person audit and spend the $150 to $400, because that is when the investment pays for itself several times over. The DOE maintains a Home Energy Score assessor directory. Many utilities subsidize or fully cover the cost. The in-person audit catches what the algorithm misses, produces a prioritized upgrade list with dollar estimates, and provides the face-to-face accountability that correlates with actually doing something about the results.
If you are buying a home, run the address through ClearlyEnergy before the inspection contingency expires. A home with estimated annual energy costs 40 percent above comparable properties is telling you something about the envelope, the HVAC equipment, or both, and that information should factor into your offer price alongside the cosmetic issues your inspector flagged.
Limitations of This Analysis
RMI's accuracy assessment compared algorithm outputs to DOE Home Energy Score, which is itself a modeled estimate, not a measurement of actual utility consumption. The "accurate enough" finding refers to asset-based performance comparisons, not operational billing accuracy. Smart thermostat study R-squared values came from a limited sample with cooperative homeowners, which likely skews toward better-maintained homes. Street View accuracy was validated only in Chicago. The 2,400 active HES assessor count is a 2023 DOE figure. PG&E's $350 savings claim is self-reported by the program operator with no independent verification. No study has tracked whether AI audit recommendations lead to retrofits at higher or lower rates than in-person audit recommendations. The 97 percent unaudited figure is derived from RMI's less-than-3-percent assessment rate and rounded; the true number depends on how you count informal contractor assessments that are not registered in any database.