Why closings are still manual, and what typed data changes

Watch a law clerk close a residential purchase and you will see a strange thing: the same dozen facts, entered over and over. Buyer names. Seller names. Property address. Purchase price. Deposit. Closing date. Mortgage principal. They arrive once, in the Agreement of Purchase and Sale and the lender's instruction letter, and then a person retypes them into requisitions, lender forms, statements of adjustments, trust ledgers, cover letters, and half a dozen portals that do not talk to each other.

Every retype is a chance to transpose a digit in a principal amount or misspell a name that must match title exactly. So firms re-check everything, which means the same facts get re-read as often as they get re-keyed. The file is not complicated. It is just repeated.

Documents in, typed fields out

The unlock is embarrassingly specific: turn documents into typed, validated data, once, at intake. In Parcl, an uploaded agreement or instruction letter goes through OCR and an extraction model that does not return prose; it returns a schema. Names as parties, dates as dates, money as money. Validation runs at the boundary: closing dates are checked as real dates, addresses verified against Canada Post, amounts cross-checked between documents that mention them. A human confirms the extraction on one screen, and that is the last time anyone types those facts.

Generate, do not retype

Once the matter is typed data, the downstream documents stop being typing jobs and become renderings. The statement of adjustments is arithmetic over known fields. The lender package is a set of bank-specific forms filled from the same source of truth. Payout letters, trust ledger entries, the statement of account: generated, consistent with each other by construction, and regenerated in seconds when a closing date moves.

• •One source of truth per matter. Every form pulls from the same validated fields, so documents cannot disagree with each other.
• •Integrations instead of portals. Title searches, title insurance orders, and e-signing happen from inside the matter, with results filed back automatically.
• •An immutable audit trail. Every extraction, edit, and generated document is recorded, which is exactly what a regulated profession needs from automation.

The clerk's job stops being data entry with legal characteristics, and becomes review with time to think.

None of this required inventing new AI. It required taking extraction seriously as an engineering problem, with schemas, validation, and audit trails, instead of treating it as a demo. Closings stayed manual for decades because the data kept arriving as paper-shaped blobs. Type the blob once, and the rest of the file mostly writes itself.