AI Searched 200 Title Files. The Misses Clustered in One Place.

On July 29, 2026, DataTrace, which describes itself as the nation’s largest provider of property and ownership data and title automation solutions, released a study titled “AI Title Search Tested in the Real World: What Accuracy, Risk, and Readiness Really Look Like.” The design is a head to head. Public-record-only AI search sits on one side, title search supported by DataTrace plant data on the other, run across 200 residential title files. The company reports that the public-record-only AI search missed at least one meaningful matter in 40.8 percent of searchable files. It reports that the most significant gaps occurred in high-risk categories such as involuntary liens, which had an issue fail rate of more than 36 percent. And it reports a third figure that most coverage skipped: of the 200 files evaluated, AI was unable to search 16, or 8 percent, because they lacked the title plant data or comparable normalized datasets needed to complete the search. Read those together and the denominator does real work. The 40.8 percent is a share of searchable files, which means the 16 unsearchable files sit outside that percentage rather than inside it.

The release also carries an underwriting extrapolation, and it deserves the label the company itself puts on it. Using what DataTrace describes as a simple illustrative extrapolation across annual existing-home sales, the analysis estimates approximately $489 billion in maximum potential liability and $148 billion in probable liability associated with missed title matters. That is a modeled figure applied to a national residential transaction count, not a measured loss, and it carries no commercial application. The conceptual distinction underneath the study is the durable part. DataTrace separates search completeness from insurability, and argues that while AI may successfully retrieve many public records, producing an insurable title decision requires connecting related documents, validating ownership history, identifying missing information, and applying underwriting judgment. Annette Cotton, the company’s chief data officer, framed the position this way: “The future isn’t AI versus title professionals. It’s AI powered by trusted title data and guided by experienced title experts.” One disclosure belongs at the top of this brief rather than the bottom. DataTrace sells the title plant data that served as the benchmark, delivering normalized datasets across more than 1,850 U.S. jurisdictions and a document library of 9 billion recorded document images. This is a vendor study whose result favors the vendor’s product. That does not make it wrong. It makes it a company claim, not a PSV finding, and it should be read as one.

The obvious objection is that this is a residential test and commercial real estate is a different business. Take the objection seriously and it cuts the other way. Every structural condition that produced the misses is more pronounced on a commercial file, not less. Chains of title run longer. Ownership sits inside layered entities, so the name being searched is a single-purpose LLC formed last quarter rather than a person who has held the property for a decade. Encumbrances arrive as reciprocal easement agreements, ground leases, air rights, subordination and non-disturbance agreements, mechanics liens filed by subcontractors nobody at the buyer ever met, and use restrictions recorded against a parcel that has been split and replatted twice. The 40.8 percent is not a number to transplant onto a commercial deal, and PSV is not transplanting it. It is evidence on a narrower question: whether a public-records sweep constitutes a complete search. On that question the commercial answer is worse than the residential one.

The category where the study found the deepest failure is the one that should register hardest. Involuntary liens are, by definition, the encumbrances no party to the transaction volunteers: tax liens, judgment liens, mechanics liens, municipal and utility charges, code enforcement penalties. They are filed by strangers to the deal, often in a different index, sometimes in a different office, and occasionally in a jurisdiction that publishes nothing usefully machine-readable. An independent record supports the same shape of finding from the other direction. The American Land Title Association’s March 2026 Critical Issues Study, conducted with ndp | analytics across more than 449 responses, states plainly that fraud and forgery are risks not detectable through public records searches, and that they represent the largest segment of claims, with an average residential claim cost of $143,000 and refinance-related claims averaging $207,000. The same survey reports that 82 percent of purchase transactions require careful review and analysis of 11 or more documents, that 21 percent involve more than 50 documents, that professionals rely on at least 9 different document sources in half of all transactions, and that 27 percent must still obtain documents in person. That last figure is the whole problem in one line. A retrieval system cannot retrieve what is not online.

The workflow is not “have AI do the title search.” It is a coverage ledger that makes the gap visible before anyone relies on it. The inputs are the preliminary title report or commitment, the underlying recorded instruments, the survey, and the jurisdiction list for every parcel in the deal. The assistant reads the commitment and returns one row per exception and requirement: what the item is, which recorded instrument it comes from, whether that instrument is actually in the file, whether it has been read, and what it does to the buyer. Alongside it sits a second table listing every index the search touched and every index it did not, by county and by record type, with the untouched ones flagged. The output is not a title opinion and is never described as one. It is a map of what has been examined and what has not. The reviewer is the transaction attorney or the title officer. The approval gate is simple and hard: no exception moves to cleared on the strength of a model summary, only on a reviewed document.

The second workflow targets the study’s worst category directly. For every entity in the ownership chain and every parcel identifier, maintain a standing list of the places involuntary liens actually get filed in that jurisdiction, and record for each one whether it was searched, by whom, and on what date. Where an index is not available online, the row says so and names the office that has to be visited or called. That is the commercial translation of ALTA’s 27 percent finding, and it converts an invisible gap into an assigned task with a name attached. The assistant drafts the jurisdiction checklist, tracks completion, and flags any parcel where a search has gone stale relative to the closing date. It does not clear anything and it does not decide anything. Neither of these workflows makes the search faster in the way a demo promises, and PSV is not claiming they do. They make incompleteness legible, which is a different benefit and the one that actually bears on risk.