Practice Areas
AI for Real Estate Lawyers: Leases, Title and Transaction Volume
AI for real estate lawyers: how to abstract 200 leases with an agent and a QA sample, what the platforms extract, land-register limits and template drafting.
On this page
Abstracting 200 commercial leases with an AI agent takes an afternoon of machine time and about two days of human checking, and the second number decides whether the exercise was worth doing. AI for real estate lawyers is mostly extraction at scale: leases, title documents, encumbrances, planning conditions. It works when the question set is fixed, the scans are clean and the QA sample is designed before the run.
The 200-lease job, step by step
Take a portfolio acquisition with 200 occupational leases in a data room, a buyer who wants a rent roll reconciled against the leases, and a fixed fee. This is the workflow that produces a defensible abstract.
- Fix the question set before touching a document. Thirty fields is typical: parties, demise, term, commencement and expiry, rent, review mechanism and dates, indexation, break options and conditions, alienation, user, repair, service charge cap, guarantor, side letters, registration status. Write each as a question a junior could answer from the lease, with the answer format specified (date as YYYY-MM-DD, money as a number and currency, free text limited to 40 words).
- Run an OCR pass first. Scanned leases and manuscript amendments are the main source of wrong answers. Claude will say "illegible" rather than invent a figure when it is looking at an image it cannot read, but a bad text layer that says the wrong thing is read as text (practitioner comparison). The PDF extraction playbook has the pipeline.
- Pilot on ten leases. Choose the oldest, the longest and the two with side letters. Compare the output against a hand abstract, fix the questions that produced ambiguous answers, and only then run the remaining 190.
- Require a citation per cell. Every answer carries the page and clause it came from. A cell without a source is treated as empty.
- Have the model flag, not decide. Instruct it to mark [UNCERTAIN] where the document is ambiguous, [SIDE LETTER] where a later document may vary the answer, and [NOT FOUND] rather than inferring.
- Run the QA sample (next section) and log every error by field and by cause: scan quality, question wording, model error, genuine ambiguity.
- Reconcile to the rent roll. The model compares its rent and date fields to the seller's schedule and lists every mismatch; a person resolves each one against the lease.
- Deliver the abstract with the error log attached. The buyer's counsel will ask what was checked and how; the log is the answer, and it is also your record if a field turns out wrong a year later.
The same structure is what the platforms implement. Harvey's Agent Builder announcement quotes Ashurst on an agent that "saves hours on producing lease summaries", against a background of 20 million-plus terms extracted across the platform (Harvey, March 2026); Legora's Tabular Review analyses up to 100,000 documents in parallel, which is why Addleshaw Goddard chose it after nearly a year of evaluation; Thomson Reuters' rebuilt CoCounsel Legal runs Tabular Analysis at 10,000 documents against 100 questions (August 2026). The due diligence page covers question-set design in more depth.
Designing the QA sample
The independent evidence is encouraging and incomplete. In the Vals Legal AI Report of 27 February 2025, AI tools beat lawyers on document extraction and Q&A, with Harvey at 94.8% on Q&A against a 70.1% lawyer baseline, and were six to eighty times faster (Vals). Nobody has published a lease-specific accuracy figure since. The arithmetic is what matters: 200 leases times 30 fields is 6,000 cells, and even 98% accuracy leaves 120 wrong cells whose location you do not know. A sample is how you find out which fields they are in.
| Field type | Examples | Check rate | Why |
|---|---|---|---|
| Money and dates | Rent, review dates and mechanism, indexation, break dates and conditions, expiry | 100% of cells | A wrong break date is a lost right; a wrong rent breaks the valuation |
| Structural terms | Alienation, user, repair, service charge cap, guarantor | 10% of leases re-abstracted by hand, weighted to the oldest scans and longest leases | Errors here are usually question-wording errors and cluster by document type |
| Descriptive terms | Demise description, permitted use wording, insurance | 5% sample plus every [UNCERTAIN] flag | Low consequence, high volume |
| Anything flagged | [UNCERTAIN], [SIDE LETTER], [NOT FOUND] | 100% read by a person | The model has told you where it is unsure; believe it |
If the 10% sample finds an error rate on a field above what the engagement letter promised, re-run that field across all 200 with a reworded question, then sample again. The benchmarks guide explains why a tool can score 94.8% on a benchmark and still need this discipline: Harvey's own long-horizon benchmark puts the best model at 13.3% on a strict all-criteria-pass metric as of 18 September 2026, with criteria-level pass rates around 83–88%, because "mostly right" and "entirely right" are different numbers.
What the platforms do and what they cost
| Tool | What it does for a property team | Price signal (September 2026) |
|---|---|---|
| Harvey | Agent Builder workflows for lease summaries and diligence; Vault; Word add-in; used by 84% of firms with 700-plus lawyers per ILTA 2026 | Not published; estimated ~$1,000–1,200 per seat per month, 20-seat minimum (secondary sources) |
| Legora | Tabular Review to 100,000 documents; overnight counterparty redlines via aOS | From about $3,000 per user per year, 10-seat minimum (secondary source) |
| CoCounsel Legal (Thomson Reuters) | Tabular Analysis 10,000 × 100; Practical Law grounding | Published range $104–$639 per user per month |
| Kira (Litera) | Hybrid generative and proprietary contract review since January 2026; the long-standing lease-abstraction engine | Enterprise |
| Spellbook | Word-native drafting and review for smaller teams; ~4,000 customers | Roughly $99–$350 per user per month |
| Claude or ChatGPT enterprise tiers | Fixed question set, files in a Project, spreadsheet output; no lease-specific features | Claude Team $25–$125 per seat; ChatGPT Business $25–$125 per seat |
For a team below twenty lawyers the honest answer is that a well-designed question set on an enterprise chatbot tier, with the pipeline above, gets most of the way, and the platform premium buys integration, audit logs and someone to call. The Harvey and Legora reviews and the pricing guide go through the trade-off.
Title, encumbrances and the land register
The agent that reads leases cannot fetch the title. Land registers are designed for per-title queries by authorised users: the Austrian Grundbuch is reached through licensed intermediaries (Verrechnungsstellen) and charged per extract, § 12 GBO in Germany requires a legitimate interest for inspection, and other registers work on the same assumption. A browser agent logged in with your credentials pulling a hundred extracts is at best a terms-of-service problem and at worst a data-protection one, and after the Ninth Circuit's reasoning in Amazon v Perplexity (4 August 2026) that "It is the user who 'accesses'" a site through an agent, the access is yours.
The workable pattern is human pull, machine read. A paralegal pulls the extracts under the firm's own access; the model then does what it is good at: tabulating every registered charge, easement, restriction and notice across the portfolio, cross-checking the registered proprietor against the lease's landlord and the survey's boundaries, and listing every encumbrance that the lease does not mention. The output is a table with a source per cell and a list of questions for the seller. Manuscript entries and older registers go through the OCR pass like everything else.
Planning, zoning and the deep-research trap
Planning conditions, zoning designations and local development plans live on municipal websites, which makes them a natural target for deep-research agents and a natural source of invented citations. The University of Pennsylvania's April 2026 study found search-augmented models and deep-research agents fabricated 3–13% of the URLs they cited and returned a further 5–18% that did not resolve, with agents hallucinating more because they cite more (arXiv). A fabricated planning condition in a report on title is an indemnity claim.
Retrieve first, analyse second. Download the development plan extract, the planning permission and the conditions schedule, load them into a Project or a Gemini Notebook, and ask the model to list every condition affecting the intended use with the paragraph it came from. The deep research guide compares the agents and sets out which questions they can be trusted with; for a property report, the answer is "the ones where you supplied the source".
Template drafting and the high-volume shop
Drafting from precedent is where the Word-native tools earn their fees. Microsoft's Legal Agent for Word (Frontier preview from 30 April 2026) reviews a document clause by clause against a playbook and writes tracked changes with explanatory comments; it requires the tenant to enable Anthropic as a subprocessor, which is a vendor-diligence point for firm IT (Legal IT Insider). Spellbook does the same for smaller teams, and Claude's own Word integration produces tracked changes on paid tiers. The Word Legal Agent page and the contract drafting page cover setup; for a lease, the playbook is the landlord's or tenant's standard positions on the thirty fields above, which is why the question set from the abstraction job is also the seed of the drafting playbook.
Volume conveyancing is being rebuilt around the same idea. Keith, a seed-funded "AI-first law firm" for residential property sales backed by Backed VC, was reported in April 2026; Crosby, a 30-lawyer New York firm with no timesheets, bills $250–$1,000 per contract and counts the developer Tishman Speyer among its clients, having reviewed 13,000 contracts by March 2026 (Forbes). Whether or not those models last, the per-document price is now a number a property client has seen, and Deloitte's 2026 survey found 92% of in-house teams expect or are negotiating AI-related rate cuts. The billable hour guide sets out how firms are repricing abstraction and diligence phases; a fixed fee per lease with a stated QA rate is the version that works.
Confidentiality in the data room
A portfolio data room is the seller's, and the process letter governs what leaves it. Before an agent reads it, confirm the platform's terms exclude training on client data (Harvey, Legora, the Team, Enterprise and API tiers of the frontier vendors, as of September 2026; consumer tiers do not by default), that any connector such as the Datasite connector in Claude for Legal respects the room's own permissions rather than a master login, and that the process letter does not prohibit automated extraction. Claude Fable 5.1 requires 30-day retention and is not eligible for zero data retention without express authorisation; for a listed seller's room, run the extraction on Opus 5 or Sonnet 5 under a ZDR arrangement, or inside a legal platform. The confidentiality checklist lists the terms tool by tool, and the M&A page covers the wider data-room questions.
A setup that works for a property team
Build one question set per document type (occupational lease, ground lease, title extract, planning permission) and keep it under version control; every improvement from a QA log goes back into the questions. Run the OCR pass on everything before the model sees it. Check money and dates in full, sample the rest, and attach the error log to every deliverable. Keep register access human and model analysis machine. Price abstraction per document with a stated QA rate, and drafting on the playbook. Module 3 of the AI Lab for Lawyers course works through the scan-to-spreadsheet pipeline on a set of sample leases, including the sampling design above.
Frequently asked questions
Can AI abstract commercial leases accurately?
Yes for the structured fields, provided the scans are clean and the question set is precise. Independent testing in February 2025 found AI tools beat lawyers on document extraction and Q&A, and firms such as Ashurst report agent-built lease summaries saving hours. Accuracy is high but not perfect, so the money and date fields are checked in full and the descriptive fields are sampled; a lease with a side letter or a manuscript amendment is always read by a person.
How big should the QA sample be for a lease abstraction?
Check every cell in the fields that carry money or dates, because a wrong break date is a lost right. For the descriptive fields, re-abstract at least 10% of the leases by hand, chosen to include the oldest scans, the longest documents and every lease the model flagged as uncertain. If the sample finds an error rate above what the engagement letter promised, widen the sample or re-run the field across all documents.
Can an AI agent pull title extracts from the land register?
Not lawfully at scale. The Austrian Grundbuch is accessed through licensed intermediaries and charged per extract, § 12 GBO in Germany requires a legitimate interest for inspection, and other registers similarly assume per-title queries by authorised users. The workable pattern is for a person to pull the extracts under the firm's own credentials and for the model to read, tabulate and cross-check them against the leases and the survey.
Which AI tools do real estate teams use for lease review?
Large firms use Harvey and Legora, whose Tabular Review handles up to 100,000 documents, alongside Thomson Reuters' Tabular Analysis, Kira and Luminance. Smaller teams get most of the same result from an enterprise tier of Claude or ChatGPT with a fixed question set and a spreadsheet output, or from Word-native tools such as Spellbook for drafting. The tool matters less than the question set and the QA sample.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Playbooks
Extract Data from PDFs and Scans with AI: Contracts, Invoices and Court Files
How lawyers extract data from PDFs and scans with AI: OCR first, a fixed schema, a full lease-extraction prompt, batch runs in Cowork or n8n, and sampling.
11 min read
Use Cases
AI Due Diligence: From Twenty Associate Hours to Two
AI due diligence now runs as a tabular agent over the data room: 10,000 documents against 100 questions, every cell cited. Numbers, a question matrix, the junior gap.
10 min read
Tools
Legora Review: The Agentic OS Challenging Harvey
Legora reviewed as of September 2026. The $5.6bn Stockholm challenger to Harvey: aOS, Tabular Review, Wexler litigation facts, pricing, clients to Salesforce.
7 min read
Tools
Spellbook Review: Contract Drafting Inside Word for Smaller Teams
Spellbook review for lawyers: the Word add-in at about $180 a month, 4,000 customers, grounded frontier models and an autonomous agent in early access.
6 min read
Practice Areas
AI for M&A Lawyers: Diligence, Drafting and Deal Management
AI for M&A lawyers now compresses diligence by 15–75%, drafts disclosure schedules and runs closing checklists. A deal-phase map of agent tasks, gates and fees.
9 min read
Guides
Deep Research Tools for Lawyers: Claude, ChatGPT, Gemini and Perplexity Compared
Deep research AI for lawyers compared: what Claude, ChatGPT, Gemini and Perplexity research modes do, where they fail on law, and a verification table.
10 min read