Residential cost segregation is going mainstream.
Most of these vendors appeared alongside ChatGPT, and it shows. Ask a powerful language model to generate a cost segregation study, and it hands you something fluent and plausible because that is what it was built to produce. Run the same property through it twice, and you can get two different answers. Plausible variation is fine in marketing copy, but it's a liability on a signed tax return. Researchers at BetterUp Labs and Stanford University call output that looks finished but cannot hold up "
Allow me to be clear. Bonus depreciation is the system working as designed: to generate economic activity. Short-term rentals employ cleaners, handypersons and realtors and support local tourism economies. AI also has a real place in cost segregation. The debate is about placement. AI belongs at the periphery of a study, finding and verifying facts. It does not belong in the calculations. Keeping it at the edge makes studies more accurate and often larger in scale. Placed in the middle, as the source of calculations, they become indefensible.
CPAs who refer providers to clients today have had only two options. Traditional partners running white-labeled, sometimes outdated Excel macros that allocate top-down: The kitchen gets a fixed percentage of the house, every house, every kitchen. Conservative, but at least the math is visible. The new AI firms are often one person, a cloud account and strong SEO, selling a $395 study whose logic no one can reproduce, including the model that wrote it. Both camps share a deeper weakness: They run on unverified narratives. The facts are whatever the client typed into a form, whatever the seller's old listing said, and a handful of receipts.
I learned what self-reported data is worth building verification systems at Airbnb. The Experiences hosts told us they had the permit and license, and that they followed local tax law and insurance requirements. When we verified, only about 40% of those claims held up. The hosts were not lying. People are unreliable narrators of their own situations. For consumer and government, everything runs on verification, and cost segregation has almost none.
Consider what that opening invites. Nothing in a $500 study confirms the client owns what is being depreciated. The field has no open repository of studies and no third-party check between what an owner claims and what gets deducted.
The bar itself is no mystery. The
Start with the property instead of a template, and treat the owner's story and the listing copy as inputs to verify and gather evidence. Public records, permits, floor plans, invoices and photos let you rebuild the asset component by component so allocation reflects measured rooms and identified materials. This is the grunt work AI is suited for: reading permit PDFs, tagging photos, pulling assessor data and flagging where the listing contradicts the owner.
Machine vision can tell glued-down hardwood from removable carpet and a Bosch range from a Whirlpool, and each carries its own classification and recovery period. Then the facts go to a query engine whose rules were written by people, tested against the tax code and locked. The engine walks a fixed set of questions: Is the part of the building structure, is it removable without damage, and what rules govern the placed-in-service year?
No model improvises logic mid-study. Given the same facts tomorrow, it returns the same answer, and a reviewer can see which inputs were considered, which rule fired, and why an asset landed in a five-, 15- or 39-year property. Rules written by people can be examined, tested and challenged by the profession. Logic learned by a model cannot. Where the owner kept a receipt, the receipt is the evidence. Where they did not, the identified component and its documented cost serve as a substitute for a percentage guess. AI finds and verifies. The engine classifies and computes.
This is also where accuracy pays. A two-year-old kitchen remodel with stone counters and tailored cabinetry disappears into a fixed percentage. Itemized, it produces a larger, better-supported first-year deduction. In fairness, on a plain long-term rental, the two methods land close, and a well-documented top-down study can be perfectly defensible. The gap opens on the properties now flooding in: partial-home rentals, converted residences, remodels. A basement rental is a share of a shared building, with rooms that belong to guests and systems that belong to everyone. Calling it a standalone ADU or Accessory Dwelling Unit is easier for a spreadsheet and fatal to an examiner who finds no separate structure. An engine built on the actual parcel data cannot make that mistake.
One objection to my argument is that a deterministic engine fed wrong facts produces the same wrong answer every time. Consistency alone proves nothing. I agree, which is why the source trail matters more than the engine. When an input is wrong, an examiner, a CPA or the client can identify it, point it out and fix it. A black box offers no such handle. You cannot correct what you cannot inspect.
Here is what worries me about the next few years. As cost segregation scales, CPAs will attach their names to platforms they have not verified, and the race to the cheapest study will erode trust that took decades to build. The damage lands on the audited client and on the CPA who signed. So before signing, ask one question of any study, AI or otherwise: Can it show its work? Where did each number come from? Was it verified or simply asserted? What rule placed it, and would the same facts produce the same answer tomorrow?
Rebuilding the asset will not always produce the largest deduction, and it should not try to. The goal is to be accurate, to be explainable, and to give the same answer twice. As this moves deeper into residential real estate, the profession needs better workpapers more than faster reports: studies that start with the property, keep the source trail and let the deduction follow the facts.
