A system can read every letter of a name correctly and still put the wrong person in your chain. That is the part of title AI that disappears between the extraction demo and the finished report. “James R. Miller” is legible, correctly classified as a grantor, and faithfully copied into the index. Whether he is the James R. Miller who acquired this parcel twelve years earlier is another question. The spelling does not answer it.

The identity posts earlier this month concerned the person participating in today’s closing. This problem lives in the historical record. Your examiner may never meet any of these people, and the software has to help distinguish them using the evidence the instruments actually contain.

Reading and Matching Are Different Jobs

October’s extraction post followed the path from characters to fields to usable records. Matching is where those records acquire relationships. The system proposes that two appearances of a name refer to the same person, that a document affects this parcel, or that a release belongs with a particular lien.

Each proposal needs a basis. The same name is one piece of evidence. The legal description, dates, capacity, related instruments, and other relevant details can strengthen or contradict it. Some records provide enough support for a straightforward connection. Others leave a question that software should preserve for examination.

This is why a vendor’s name-extraction accuracy tells you very little about the reliability of the chain it builds. Reading the label on a document and deciding where that document belongs are separate accomplishments.

Two Errors Pull in Opposite Directions

Imagine two unrelated people named James Miller appearing in the same county’s records. Combine them and the search can attach one person’s instruments to the other. Now imagine one owner whose name appears with a middle initial on one deed, without it on a mortgage, and under a former surname elsewhere. Split those appearances too aggressively and relevant instruments may fall out of view.

The hypothetical examples point to competing errors. Overmatching invents connections. Undermatching loses them. A product can improve one headline number by moving the threshold toward the other error, which is why “we catch more matches” is not a complete description of an improvement.

For an examiner, the useful question is whether the system makes the evidence easier to evaluate. A larger pile of unexplained candidates may increase recall while consuming every minute the automation was supposed to save.

Keep the Original Name

Normalization makes records easier to find. It can standardize capitalization, spacing, and familiar variations. It can also erase a distinction if the normalized value replaces the original text.

Keep both. The examiner should be able to see the name as recorded, the form used for searching, and the proposed connection. A suffix, a middle initial, or a representative capacity may be the detail that explains why an apparent match is wrong. A trust with a similar name is not interchangeable with the individual trustee, and a person appearing in different capacities should not lose those capacities when the system groups results.

I want software to reduce the time it takes to inspect that evidence. If it hides the original behind a polished identity profile, it has made the examiner dependent on a conclusion that may be harder to challenge than the paper ever was.

Make Uncertainty Visible

A useful matching screen explains why a candidate was proposed and where the evidence conflicts. It links back to the relevant source rather than asking the examiner to trust a similarity score. The person reviewing the file needs to distinguish a spelling variation supported by the chain from a same-name record with no demonstrated connection to the property.

There should also be a way to leave a connection unresolved. Forcing every candidate into “same person” or “different person” creates false certainty when the evidence supports neither conclusion yet. Record the reviewer’s decision and reasoning, and preserve the ability to revisit it when another instrument changes the picture.

One file-specific decision should not silently become a countywide rule. An examiner accepting “J. R. Miller” on this chain has not established that every future appearance of those initials belongs to the same person.

Test the Relationships

Build a small evaluation set around known matching problems. Include same-name strangers and documented name variations. Ask the vendor to show both missed connections and incorrect ones, with the sources and the reviewer’s path to correcting them. A hundred correctly transcribed names are not a substitute for that test.

This is part of the conversation worth having about Autopilot and Scribe: how the extracted record becomes evidence an examiner can use, and how uncertainty reaches the person responsible for resolving it. Bring the awkward names to the demo. The clean ones have already had their turn.

Discuss record matching across Electra’s title workflow →