AI Missed a Title Matter in 40.8% of Files. Read Who Tested It.

DataTrace, which publishes as Data Trace Information Services LLC, ran an evaluation comparing public-record-only AI title search against what it calls a good and proper title search supported by DataTrace title plant data. The company says the broader evaluation covered more than 200 residential title files, of which 184 were fully searchable and evaluable across Arizona, California, Florida and Texas, with the remainder excluded because public site access or other conditions prevented a complete comparison. Of those 184 files, 75 contained at least one missed title matter, which DataTrace reports as an observed 40.8 percent file-level miss rate. Checking that arithmetic against the company’s own figures, 75 divided by 184 computes to 40.76 percent, so the published rate holds. DataTrace also reports that the misses were not evenly spread: involuntary liens accounted for approximately 79.6 percent of the total weighted risk burden observed, with the remaining categories being encumbrances and mortgages, deeds, vesting and tenancy, and mortgage downlines. The company states plainly that the evaluation was not designed as a statistically representative sample of every US jurisdiction, and that this portion of the research measured search completeness rather than underwriting or claims exposure. The findings were released with a July 29, 2026 press release and detailed in company posts dated August 5 and August 12, 2026.

A second, separate review carries the money figures, and the separation matters. DataTrace says it conducted an underwriting review of 50 residential properties, distinct from the 184-file completeness set, to evaluate what observed misses could mean for examination, insurability and claims exposure. From that review it estimates maximum potential liability of $6.1 million and probable liability of $1.86 million, and it states explicitly that these are underwriting estimates grounded in title expertise and professional judgment rather than predicted losses. DataTrace’s research page then carries a much larger number: maximum potential liability from search misses across 4 million annual existing home sales estimated at $489 billion. That figure is an extrapolation, and the arithmetic is worth doing yourself. $6.1 million across 50 properties computes to $122,000 per property, and $122,000 across 4 million sales computes to $488 billion, which lands on the published figure within rounding. Running the same method on the probable number, $1.86 million across 50 properties computes to $37,200 per property, or roughly $148.8 billion across 4 million sales, which is PSV arithmetic on DataTrace’s figures and not a number DataTrace publishes. Probable liability is about 30.5 percent of maximum in the tested set. One more disclosure belongs in the same read: DataTrace sells title plant data and title automation software, and the comparison it ran was between AI on public records and AI on its own product category. That is a vendor-run study, not an independent one, and nothing in the published material claims otherwise.

The tested files are residential and the tested states are four, so the direct finding does not describe commercial diligence. What transfers is the failure mode, and the failure mode is one commercial teams are actively building toward right now. Pointing a model at a county recorder portal to pull liens, mortgages, assignments and the chain of title is an obvious, cheap, and increasingly common first pass in acquisitions and portfolio work. This evaluation is the first public attempt to quantify what that pass leaves behind, and the answer is not that the model returns garbage. The model returns a clean, confident, well-organized summary that is missing something. DataTrace frames the underlying reason in a sentence worth borrowing: public records provide notice, not verification. County records were built to give the world constructive notice that an instrument exists. They were never built to confirm that every relevant instrument has been found, connected to the right parcel, and read against the others. Indexing conventions, name variants, historical formats and the relationship between documents all vary by county, and those variations become part of the operating environment of any model reading them.

If anything the commercial version of this problem is harder than the residential one the study measured. Commercial title runs through entity chains rather than individual names, which multiplies the name-variant problem the study identifies as a driver of misses. It carries instruments residential files rarely do: ground leases, reciprocal easement agreements, assignments of rents and leases, UCC fixture filings, and mechanics liens with their own statutory timing. Involuntary liens, the category carrying roughly 79.6 percent of the weighted risk in the evaluation, are precisely the class of matter that attaches without the owner’s signature and therefore without a counterparty to ask. And the per-file stakes are not $122,000. The honest conclusion is not that AI title work is unsafe; it is that a public-record-only search is a triage layer whose output has a known, now partially measured, incompleteness. Treating it as a finished search is the error. Nobody in commercial real estate should be replacing a title commitment with a model, and the study does not claim anyone is. What it does establish is that the gap between a retrieved document set and a decision-ready one is large enough to measure.

The first workflow is a record-source register, and it is deliberately unglamorous. Every AI-assisted record pull writes one row per property recording: the jurisdiction searched, the exact portal or dataset the documents came from, the date range the source actually covers, the name and entity variants the search ran under, the date pulled, and what the search could not reach. That last field is the one that matters, because the study’s own excluded files were excluded for exactly this reason, public site access or other conditions preventing a complete comparison, and a model asked to summarize what it found will not volunteer what it could not open. The output is not a title opinion and is never labeled as one. It is an exceptions list with a coverage statement attached, and it goes to the person who orders title, not to the investment committee. The approval gate is simple and absolute: no AI-derived record summary is used to make or support a title decision, and no deal document quotes it as evidence of clear title. It exists to tell a human where to look harder and what to ask the title officer about, three weeks earlier than the commitment would have.

The second workflow is a category checklist run as a review gate, built from the study’s own miss taxonomy rather than from a general prompt. The five categories DataTrace names, involuntary liens, encumbrances and mortgages, deeds, vesting and tenancy, and mortgage downlines, become five explicit questions the model must answer separately and cite for, with the record source and instrument number behind each answer, and with an unresolved flag as an allowed and expected output rather than a failure. Weight the review by the study’s own risk concentration: involuntary liens get the closest human read because that is where roughly 79.6 percent of the weighted risk sat. The reviewer is the diligence lead or the attorney who signs the title objection letter, and the reviewer’s job is not to check whether the model wrote well, it is to check the citations. Every claim with no instrument behind it gets struck. Then the whole exercise gets reconciled against the title commitment when it arrives, and the diffs get logged, because after twenty deals that log is the only honest measurement of how the workflow performs on your files, in your states, on your asset classes. That measurement is worth more than any vendor’s published rate, including this one.