Are AI cost segregation studies audit-defensible?

The split in the market

The category has quietly divided in two. Knowing which half you are buying from is the whole decision.

1. 01

   The cheap tier appeared with the chatbots

   A crop of $495 desktop and AI 'studies' promises a 40-page PDF in an hour. They apply template percentages to a purchase price with no site inspection, then format the output to look like an engineering report.

2. 02

   Licensed engineers are refusing to validate them

   Owners are bringing AI-generated studies to real engineering firms and asking them to sign off. The firms decline, because signing means standing behind an inspection that never happened.

3. 03

   The examiner does not care that AI was used

   The IRS question is not 'was a model involved,' it is 'does the work meet the standard.' A study that cannot produce the photos, the engineer of record, and the reconciliation fails that test no matter how it was written.

What an examiner actually asks for

Defensibility is not a matter of how the PDF looks. It is whether the file can produce these four things on request, and whether we can too.

• An engineer of record. A named, credentialed preparer whose reputation is attached to the classifications. Unlevered's component library, takeoff methodology, and systems models were designed with a civil engineering partner, and every study is engineering-reviewed.
• Evidence of a real inspection. Dated photographs and documents tied to the actual property, not stock renderings. Unlevered rebuilds the property from its own evidence into a 3D component model, and every value is labelled by how it was established: measured, public record, derived, or a governed default.
• Reconciliation to cost records. Component values that trace back to closing statements, receipts, and public records rather than a percentage of the price. The Workpaper retraces the calculation line by line to those sources.
• A methodology that can be replayed. The Cost Segregation Audit Techniques Guide expects a defensible, documented method. Unlevered's engine is deterministic: the same inputs produce byte-identical output, enforced in continuous integration by replaying reference studies and comparing hashes.

The right role for software, and for AI

We are a modern platform and we are proud of the technology. The point is where the technology is allowed to act.

• Software decides the math. A pure, deterministic engine computes every classification, basis allocation, and depreciation figure. It has no model in it, and it produces the same result every time from the same inputs.
• AI only reads and proposes. Transcribing a dimension, spotting a room in a photo, pulling a line off a receipt. Every proposal is validated by software, and material ones need human review before they count.
• A person stands behind it. The study is engineering-reviewed and signed. Your CPA reviews it, applies independent judgment, and files. Nobody is asked to trust a black box.

Unlevered prepares and signs the engineering study. Your CPA remains the sole preparer of your return. These statements describe how the platform works and are not tax advice; your advisor confirms the figures before you rely on them.

Common questions

Are AI cost segregation studies legit?

An AI-generated or template cost segregation PDF is not the same thing as an engineering-based study, and the difference is what an examiner tests. The IRS looks for an engineer of record who stands behind the work, dated photographic evidence of a real inspection, component classifications reconciled to actual cost records, and a documented methodology. A language model cannot produce those, and it will give two different answers to the same property on two different runs. An engineered study is reproducible, cited, and signed. That is what holds up.

Is a cheap DIY or $495 cost segregation study worth it?

It depends on what you are buying. A desktop study that applies template percentages with no inspection typically allocates less to shorter-life property than an engineering-based study, and more importantly it cannot produce the inspection evidence and engineer of record an examiner asks for. The low price looks like a bargain until an Information Document Request asks for documents that do not exist. For any property where the deduction is meaningful, the incremental cost of a real engineered study is trivial next to what is at stake.

What makes a cost segregation study audit-defensible?

Four things: an engineer of record who stands behind the classifications, dated photographic and documentary evidence of a real inspection, component values reconciled to actual cost records, and a documented, repeatable methodology calibrated to the IRS Cost Segregation Audit Techniques Guide. Unlevered adds a fifth: the whole study is sealed under a content hash and reproducible byte for byte, so the file can be reopened and re-derived years later without asking for anything it cannot produce.

Does Unlevered use AI?

Yes, in a bounded role, and never to decide the numbers. Software, not a model, performs all geometry and tax math. AI is used only to transcribe and propose: reading printed dimensions off a floor plan, identifying rooms in photographs, extracting line items from a document. Every proposal is validated by deterministic software, and material AI-proposed facts require human review before they take effect. The engine that computes classifications, basis, and the schedule contains no AI at all. If AI is handling allocation on a human-verified component list, that is defensible; if AI is making the classification decisions, it is not.

Will an AI-generated study cause a problem in an audit?

It can, because it invites questions it cannot answer. An examiner will ask about the physical inspection, the engineer of record, and the source records. When those documents do not exist, the examination gets difficult and the deductions are at risk. The problem is not the tool; it is that the work product cannot be defended. An engineered study is built to be defended from its own file.