Why ChatGPT fails at title curative work
County record fragmentation, paywalled plants, handwritten scans, and iterative investigation are exactly where a chat assistant breaks. Here is the architecture that does the job instead.
Every title team we talk to has tried it: paste the title commitment into a chat assistant and ask what is wrong with the file. The answer reads confident, cites nothing, and cannot be handed to an underwriter. That is not a prompt problem. Title curative is a different shape of work than the one chat assistants were built for, and the difference matters for anyone buying AI for a title desk.
The work is iterative, not one-shot
A curative file is a loop: search, read, cross-reference, chase, verify, repeat. You pull the mortgage, find the assignment chain dead-ends, query the servicer directory, discover a name variation, re-search under the variation, and only then know which party can sign the satisfaction. Each step decides the next one. A single-pass question answerer has no loop, no memory of the file's evolving state, and no tools to touch the systems where the answers live.
The data is hostile on purpose
- More than 3,100 county record systems, each with its own portal, index conventions, and document formats. There is no federal layer and no common API.
- The deep history sits in title plants behind commercial paywalls, not on the open web an assistant can reach.
- A large share of instruments are scans of handwriting, stamps, and microfilm. Reading them is an OCR and vision problem before it is a language problem.
- The identifiers that matter, instrument numbers, book and page, parcel IDs, are inconsistent even inside one county's history.
No citation, no deliverable
A curative conclusion is only worth what it points to. A defect row that reads 'unreleased mortgage, 2011' is useless to a title officer unless it links to the recorded instrument and page. A chat assistant produces prose from weights, not provenance from records. When it states a fact about a specific property, nothing in the output tells you whether it read the instrument, misread it, or reconstructed a plausible one from training data. In this industry a plausible fabrication is worse than an error, because it gets copied into a commitment.
The architecture that does work
ClearFile is built as a harness around a model, not as a model with a chat window. The parts that matter:
- An agent loop that plans the investigation and decides the next query from the last result, running until every defect has a resolution path or an honest dead end.
- Real tools: county portal queries, plant access where licensed, OCR and vision over scans, and servicer lookup directories.
- A citation layer binding every claim to a document, page, and record URL, with one hard rule: a claim with no source is dropped, not guessed.
- A review UI where a licensed human verifies rows link by link and signs the package.
- Correction capture: when the reviewer fixes a row, the correction becomes a paired training example, so the next file in that county starts smarter.
What the machine is actually for
Not judgment. Speed and coverage. The machine reads a pile of scans in an hour, cross-references three counties without fatigue, and never forgets a follow-up date. The reviewer spends their time on the ten rows that need a licensed decision instead of the ninety that just need chasing. That division is the entire product thesis: evidence, not answers.
If a system will not show you the page it read, it is asking you to co-sign its imagination.
A test you can run this week
Take one stuck file. Give the chat assistant the commitment and give ClearFile the same file. Compare three things: defects found, sources attached per defect, and hours of human time left in each package. That comparison is the whole evaluation, and it is why we price per file and give the first ten away.
Common questions
Can RAG over county documents do curative work?
Retrieval over documents answers questions about documents you already hold. Curative work needs actions against systems you do not hold: searching recorder portals, resolving holders through servicer and MERS directories, and tracking submissions over weeks. It is an agent loop with tools and a citation layer, which is a different system than a retrieval index.
Does the AI decide anything about insurability?
No. ClearFile produces work product for licensed review. The title officer or attorney reads the cited package and decides. The system's rule is to abstain rather than fill gaps, precisely so the reviewer is reviewing evidence instead of editing fiction.