How to abstract a commercial lease with AI you can trust

To abstract a commercial lease with AI and trust the result, give the model the whole document stack in order, your own abstract template, and an instruction to cite a page and quote the lease for every field it fills in. Then have a person check every date, dollar, and option against the cited page before anyone relies on the abstract.

The rest of this article is the detail behind those two steps. It is written for the leasing broker, tenant rep analyst, or property manager who abstracts leases by hand today and has pasted a lease into a chat tool with mixed results.

What is a lease abstract for, and who relies on it?

An abstract is the short version of the lease that people use in place of the lease. A tenant rep team uses it to track option notice dates and to know what the client is paying as a renewal approaches. A property manager uses it to bill rent steps and reconcile operating expenses. A capital markets analyst uses it to build the rent roll behind a valuation.

None of those people re-read the lease, which is the point of the abstract and also the risk. A wrong notice date in the abstract becomes a missed option, and a wrong escalation becomes a wrong invoice.

Why does pasting a lease into a chat window give uneven results?

The usual experiment is one PDF, one prompt that says "abstract this lease," and no template. The model gets the base lease without the three amendments, picks its own fields, and returns a tidy summary with no way to check any line of it. Some of it is right, and the reader cannot tell which part.

Accuracy percentages on vendor pages do not help with this, because a per-field percentage is the wrong measure for a document people rely on as a whole. As a purely arithmetic illustration, an abstract with 60 fields where each field is independently 98 percent likely to be right has about a 30 percent chance of containing no error at all. The useful questions are where the errors tend to sit and how quickly a reviewer can find them.

We build lease abstraction workflows for brokerage teams, and the errors we watch for sit in a short list of places. The method below is designed around that list.

Assemble the full document stack first

Before the model sees anything, gather every document that makes up the deal, in this order:

1. The base lease
2. Every amendment, in date order
3. The commencement letter or commencement date memorandum
4. All exhibits, including the work letter, the rules, and any rent schedule
5. Side letters, estoppels, or assignments, if they exist

Name the files so the order is obvious, and tell the model the order in the prompt. If an amendment is missing, the abstract will state the superseded term with complete confidence, and nothing in the output will reveal the gap. Ask the model to list every document it received and to flag any amendment that is referenced in the stack and not included in it.

Give the model your template and your definitions

A generic field list produces a generic abstract. Give the model the template your team already uses, with a definition beside each field. "Commencement date" should say whether you mean lease commencement or rent commencement. "Base rent" should say whether you want annual, monthly, or per square foot, and for which period. "Term" should say whether free rent months count.

These definitions hold much of a team's abstracting judgment, and they often live in the heads of the people who do the work. Writing them down once is one of the most valuable steps in the whole build.

Require a page citation and quoted language for every field

Each field in the abstract should carry three things: the value, the document and page it came from, and the lease language quoted word for word. With that in place, the reviewer checks a source instead of re-reading the lease.

Two more instructions belong here. The first is to have the model write "not found" when the lease is silent. Anthropic's guidance on reducing hallucinations recommends explicitly giving the model permission to say it does not know, extracting word-for-word quotes first on long documents (it gives more than 20,000 tokens as the threshold), and having it retract any claim it cannot support with a quote. The same page says these techniques reduce hallucinations without eliminating them, and that critical information should always be validated.

The second is to treat the quote as something to verify. A quote typed out by a model in a chat window can itself be wrong, so the reviewer opens the cited page. Teams building on the API have a stronger option: Anthropic's citations feature returns the cited text along with page numbers for PDFs, and its documentation says those citations are guaranteed to contain valid pointers to the provided documents.

How should the abstract handle amendments?

Ask for two outputs. The abstract itself states the current term for each field. A separate conflicts list shows every place an amendment changed an earlier term, with the original language, the new language, the amendment that changed it, and the page for each.

The conflicts list is the part a senior person should read in full. It turns the hardest reasoning in the abstract into something visible, and it catches the case where a second amendment restates the rent schedule and the third quietly changes the expiration date.

Where do AI abstracts usually go wrong?

Area	What goes wrong	What the reviewer checks
Amendments	The base lease term is reported after an amendment replaced it	The conflicts list, against each amendment
Escalations	A cap, a floor, or a cumulative versus non-cumulative limit is dropped	The quoted escalation clause, in full
Options	The notice window is misread, or conditions on the option are left out	Notice dates, counted from the quoted language
Operating expenses	Exclusions, caps, and base year terms are summarized loosely	The exclusions list and any cap language
Cross-references	A term defined in an exhibit is filled in from the body of the lease	The exhibit page the definition lives on
Square footage	Rentable and usable square footage are confused	The figure, its label, and the measurement standard cited

On the last row, record which measurement standard the lease cites and leave the number as the lease states it. BOMA International publishes floor measurement standards and interpretations covering office, retail, and industrial buildings, among other property types, and states that it does not certify, approve, or endorse any individual, firm, device, or software for measuring floor areas. An abstract should report what the lease says about area and leave any remeasurement to the people qualified to do it.

Which fields does a person check against the page every time?

Tier the review so the reviewer's attention goes where an error costs the most.

• Check against the cited page every time: every date, every dollar amount, every option and its notice window, the escalation mechanics, and the full conflicts list.
• Read the quote and move on: use clauses, assignment and subletting, maintenance responsibilities, insurance requirements.
• Skim: parties, addresses, and other descriptive fields.

The reviewer signs the abstract when the first tier is done. We cover how to design that step in more depth in how to design the human review step in an AI workflow.

Scanned leases, long leases, and file limits

Older leases are often scans, and a full stack with exhibits can run long, so check both before you design the workflow around a single upload.

Anthropic's PDF documentation for the Claude API lists a maximum request size of 32 MB (which it notes varies by platform) and a maximum of 600 pages per request, or 100 pages when the request's context window is under 1M tokens. It requires a standard PDF with no password or encryption. It also notes that dense PDFs can fill the context window before reaching the page limit and suggests splitting the document into sections. Those are API limits, and the limits in the app your team uses may differ.

The same documentation says each page is processed as both text and an image, which is how the model can work with a page that has no text layer. It recommends clear, legible text and pages rotated upright. The citations documentation adds that scans with no extractable text cannot be cited. In practice, run OCR on scanned leases so they have a text layer, fix sideways pages, and put abstracts from poor scans through a heavier review.

If you have to split a long stack, split at document boundaries and keep the definitions section and the relevant exhibits with whatever part refers to them.

Getting the abstract into your workbook or lease system

An abstract saves time only if it lands where the team already works, in the same format every time. That means the same field order, the same date format, the same units, and the same wording for "not found." If the output goes into an Excel abstract template or is keyed into a lease administration system, have the model fill that exact layout so nobody reformats it by hand. The same approach applies to other deal documents, which we cover in getting deal data into Excel and your CRM with AI.

Keep the leases inside the firm's own AI workspace and off personal accounts. What a vendor does with uploaded documents depends on the plan and its terms, so confirm those with your counsel or compliance lead before client leases go in. Our guide to secure AI workflows for confidential client data covers the setup.

Packaging the method as a reusable skill

Once the template, the definitions, the citation rule, the conflicts list, and the review tiers are settled, package them so every abstract follows them without anyone retyping a prompt. In Claude that package is a skill. Anthropic's skill authoring guidance describes a template pattern for outputs that must follow an exact structure, and a validation loop in which the output is checked and fixed before the work proceeds. Both fit an abstract well: the template fixes the format, and the check confirms that every field has a value or "not found," a page, and a quote.

Before the team relies on the skill, run it on leases you have already abstracted by hand and compare the two field by field. That comparison gives you an accuracy picture for your own leases, and it tells you which definitions to tighten. We describe the other skills a brokerage team tends to build in Claude skills for commercial real estate, and if you would like help building this one against your own template, you can get in touch.

Common questions

How accurate is AI lease abstraction?

There is no single honest number, because accuracy depends on the lease, the scan quality, the number of amendments, and how the fields are defined. Accuracy percentages on vendor pages are vendor claims. The more useful test is to run the method on leases your team has already abstracted by hand and compare field by field, paying most attention to dates, dollars, and options.

Can AI read scanned leases?

Often, yes. Anthropic's documentation says Claude processes each PDF page as both text and an image and recommends clear, legible text, and it also says that scans with no extractable text cannot be cited by the API citations feature. Run OCR on scanned leases first, and review those abstracts more heavily.

Does AI handle lease amendments correctly?

Only if it receives every amendment, in order, and is told to report what each one changed. Ask for a conflicts list that shows the original term, the amended term, the amendment that changed it, and the page for each. A person should read that list in full.

Is it safe to upload a client's lease to an AI tool?

That depends on the plan your firm is on, the vendor's data terms, and any confidentiality agreement covering the lease. Keep leases inside the firm's own AI workspace instead of personal accounts, and confirm the terms with your counsel or compliance lead before client documents go in.

Do I still need a person to review an AI lease abstract?

Yes. Anthropic's own guidance says its techniques for reducing errors do not eliminate them and that critical information should always be validated. A person should check every date, dollar amount, and option against the cited page, and can skim the descriptive fields.