AI for Real Estate Lawyers: Lease Abstraction, Title Review and the Length Ceiling

Imagine an abstract that reads “Renewal Options: two 5-year options, Section 3.4”. It looks right. Section 3.4 of the lease is about signage, and the tenant has no renewal option at all. Nobody typed that row; a model did, because a lease that long usually has one.

That is the specific risk of AI for real estate lawyers, and the clearest statement of it comes from a vendor. Bryckel, which sells lease-abstraction software, published a page asking whether ChatGPT can abstract a lease. Its answer: yes for “summarising individual clauses, extracting rent schedules, identifying key dates, translating complex legal language into plain English, drafting preliminary abstraction tables”, but “Generative AI models sometimes generate information that appears plausible but is not actually present in the document”, renewal terms included. What follows is a method built around that admission.

Why AI for real estate lawyers works, and why that cuts both ways

Real estate work is high-volume, repetitive and document-bound: the same twenty fields from every lease, the same Schedule B exceptions on every commitment. That is the profile on which the Better Call GPT study found a GPT-4-class model matching junior lawyers at determining whether an issue exists (F-score 0.871 against 0.860) while being weaker at locating it in the text. Extraction with references plays to the strength and guards the weakness.

The firms have noticed: Harvey reports that Ashurst built lease-summary workflows on its Agent Builder, and Legora bought Cadastral, an agentic real-estate platform, in June 2026. Small-firm evidence is quieter: Pierce, at a small Missouri business and real estate firm, told Clio’s 2025 Legal Trends Report that research which took “an hour or two” of keyword guessing now takes “literally five minutes and my first two hours of research are done”.

The flip side: a document-bound practice is exactly where a plausible wrong answer hides best.

The length ceiling: 40 to 120 pages and silent truncation

Bryckel’s page counts are the starting point: retail leases run 40 to 80 pages, office leases 60 to 120, ground leases 100 or more, before amendments. An engineering team that built its own abstraction tool learned the lesson the hard way: “because GPT can’t process texts that contain more than 8 thousand words, the text has to be further split into semantic parts” (Ascendix). That limit belongs to an older model, and current context windows are far larger, but the underlying problem moved rather than disappeared. Anthropic’s own documentation says that “more context isn’t automatically better. As token count grows, accuracy and recall degrade, a phenomenon known as context rot”, and Thomson Reuters’ prompting guide warns that a model “is more likely to forget or fail to consider information contained in the middle of a prompt”.

The rule: chunk by article, extract one article at a time, and put the question after the text. The long-contracts guide covers the mechanics and the context window explainer the reason.

The abstraction table: Field / Extracted Value / Clause Reference

Bryckel’s method: assign a role (“commercial real estate analyst”, “lease administrator” or “real estate attorney”), define the task, the fields, the output structure and how to handle missing information, and require a table with columns Field / Extracted Value / Clause Reference. Its example rows are the idea in miniature.

The clause reference is what makes the table safe: a paralegal checks “Section 3.4” in thirty seconds, and without it the only way to verify the renewal row is to re-read the lease. Two more columns make invented renewals visible.

Act as a commercial lease administrator working for the [tenant / landlord]. From the lease article pasted below only, extract into a table with exactly these columns: Field | Extracted value | Clause reference (section and page) | Confidence (High / Medium / Low) | Note.
Fields: parties; premises; commencement and expiry dates; renewal options (number, length, notice window, method of exercise); base rent and escalations; security deposit; permitted use; assignment and subletting; break rights; repair standard; insurance; operating-expense cap.
If a field is not in the pasted article, write NOT FOUND. Never infer a term from what leases usually say. After the table, list every amendment, side letter or estoppel the pasted article refers to.

<lease_article>
[paste one article]
</lease_article>

Run it per article, then merge. Bryckel also notes non-determinism, “the same prompt applied to the same document may produce slightly different outputs across multiple runs”, so run the money and date fields twice.

Prompts for the rest of the file: rent schedules and critical dates

The tasks Bryckel rates as suitable for a general model are narrower than a full abstract, and they are the ones to automate first: a rent schedule is arithmetic on stated figures, a key-dates list is extraction, a plain-English client summary is rewriting. None needs the model to know “market”.

From the rent and term articles pasted below only, produce (1) a rent schedule: Period | Annual base rent | Monthly instalment | Escalation mechanism (quoted) | Clause reference; and (2) a critical-dates table: Event | Date or trigger | Notice required by (show the calculation) | Consequence of missing it | Clause reference. Where a date depends on a fact not in the text, such as the delivery date, write DEPENDS ON [fact].

<lease_extract>
[paste]
</lease_extract>

Re-check the arithmetic in a spreadsheet: Thomson Reuters lists “math, counting, and sorting” among model weaknesses, and “90 days before the anniversary” is a calculation, not a lookup.

Amendments, side letters and estoppels need a human

Bryckel is blunt that amendment hierarchies, side letters and estoppels need judgement, for a structural reason: an abstract of the original lease is wrong the moment a second amendment changes the rent-review mechanism. The current position is a chain, not a document. A model can reconstruct the chain if you give it every link in date order and ask it to show its work.

Here are the original lease and every amendment, side letter and estoppel in date order, each labelled with its date: <doc_1>...</doc_1> <doc_2>...</doc_2>. For term, rent and reviews, renewal options, permitted use, assignment and break rights, state the current position with the chain that produced it (original clause -> amended by doc X clause Y -> ...). List any amendment that changes a clause already superseded, any document lacking signatures or dates, and any conflict between the estoppel and the lease as amended. Do not summarise anything not in these documents.

Then do what the model cannot: decide whether the unsigned side letter binds anyone, and whether the estoppel’s rent figure is the tenant’s mistake or your client’s problem.

Title commitments and survey exceptions

Vendor content claims a model can map a commitment’s exceptions against the survey and draft an objection letter in 15 minutes rather than 90. I could not verify that. What I would use is the cross-referencing. Give the model the Schedule B exceptions and the survey’s legend, and ask for a table: Exception number | Instrument and recording reference | Where it appears on the survey (or NOT SHOWN) | Whether it touches the improvements or the access | Question for the examiner. That is a list, not an opinion; title work stays human.

The same discipline applies to purchase agreements. Asking for “every non-standard contingency, seller-warranty limit, closing-cost allocation and HOA provision, quoted, with the clause number” is a review against a checklist you own; asking “what must a residential seller disclose in [state]” is open legal research, where models invent statutes. The playbook review guide explains why the first works and the second fails.

Confidential operating provisions and PII in leases

Leases are not public documents. Bryckel lists “tenant sales reporting, confidential operating provisions, personally identifiable information” among their contents. The UK Upper Tribunal’s finding in UKUT 81 (IAC) that putting client documents into an open AI tool “is to place this information on the internet in the public domain” was about Home Office letters, but the reasoning does not care what the document is.

For question-and-answer over a single lease, Gemini Notebook on a Workspace account is the grounded option: it answers only from uploaded sources, with clickable citations. The NotebookLM guide covers set-up and account caveats.

What the specialist tools claim, and what is unverified

Real estate attracts numbers. An affiliate page claims 20 to 30 lease abstracts an hour against two or three by hand, and a practice going from 8 to 28 closings a month per lawyer. A comparison site claims ChatGPT and Claude both hit 90 to 95 percent on standard terms; specialist tools are quoted at 90 to 97 percent field accuracy. None publishes a methodology, and I treat all of them as marketing until someone does.

The verified evidence is more modest. Ashurst’s firm-wide trial, Vox PopulAI, found large time savings on some drafting tasks and “frequent hallucinations across all GenAI tools trialled”. Both are true at once; hence the clause-reference column.

Where to go next: the practice-area hub and the other practice-area guides apply the same extraction discipline elsewhere, the tax guide covers the arithmetic problem at its harshest, and the prompt library has the lease and diligence recipes. The document-extraction exercise in AI Lab for Lawyers uses a commercial lease, so you leave with the abstraction routine working on your own tools.