You can summarize a 90-page commercial lease with AI in about 10 minutes, and most of those minutes are not the AI’s — they are yours, spent preparing the file and checking the answer. The model does the reading in seconds; the discipline that makes the result safe to act on is the two minutes of prep before and the five minutes of verification after. This is a walkthrough of that exact clock: what to do, what to ask, and where a wrong value will cost you real money if you skip the check. Follow it and a lease review that used to eat an afternoon becomes a coffee-break task — without turning your firm into someone who trusts a machine’s guess.
The 10-minute clock
The whole reason this works in 10 minutes is that you spend the time where it pays. A summary produced in fifteen seconds and trusted blindly is a liability; the same summary produced in fifteen seconds and checked in five minutes is a genuine time saving over the two hours a careful manual read takes.
Here is the budget for a single lease, start to finish:
| Minutes | Step | What you are doing |
|---|---|---|
| 0-2 | Prep and gate | Confirm the file is one document, check it has real text, decide it is safe to upload |
| 2-4 | Prompt | Paste a field-anchored prompt that names exactly what to pull |
| 4-6 | Read | The model reads the lease and returns a structured summary |
| 6-10 | Verify | Spot-check the fields where an error is expensive, against the source |
Notice that the two human-heavy steps — prep and verify — take seven of the ten minutes. That ratio is the point. The people who get burned by this are the ones who spend zero minutes on either end and treat the middle as the whole job.
Minutes 0-2: prep the file and set the confidentiality gate
Before anything else, answer two questions about the file in front of you: is it one clean document, and is it safe to upload?
Check that the text is real. Open the PDF and try to select a paragraph with your cursor. If the text highlights, the file is digitally native and the model reads it directly. If nothing selects, you are holding a scan — a picture of a page — and the model has to recognize the characters first, which is where a faxed amendment turns a 5% cap into a 3% one. A poor scan is the single biggest predictor of a bad summary, so if the copy is a low-resolution fax, re-scan the clean original at 300 DPI before you start. Our companion piece on how AI actually reads a lease traces why the scan-versus-native distinction governs everything downstream.
Set the confidentiality gate. A commercial lease carries tenant financials, negotiated terms, and sometimes personal guaranty details — data you cannot casually hand to a consumer chatbot. The rule for a small firm is simple: use a paid business or team tier where the provider contractually excludes your inputs from model training, and turn off any “improve the model” setting on the account. Free consumer tiers are the ones to avoid for real deal documents. If a specific lease is unusually sensitive, redact the party names before uploading — the model can still extract rent, dates, and options from an anonymized copy. Two minutes spent here is what keeps a productivity habit from becoming a disclosure problem.
Minutes 2-4: the prompt that does the work
The difference between a useless summary and a useful one is the prompt. “Summarize this lease” gives you a book report. What you want is a structured abstract with named fields, a forced “not found” for anything missing, and a page citation for every value so you can verify it in seconds.
Paste something like this, then attach the lease:
You are abstracting a US commercial lease for a broker. Extract the
following fields into a table. For each field give the value, the page
number, and the exact quoted clause it came from. If a field is not
present in the document, write "NOT FOUND" - never guess.
Parties: landlord, tenant, guarantor
Premises: address, suite, rentable square feet
Lease type: gross / modified gross / NN / NNN / absolute
Term: commencement date, expiration date, total months
Base rent: the rent for EVERY year or step of the term, not just year one
Escalations: fixed % or CPI, and the exact schedule
Free rent / abatement: any months of abated or reduced rent
Operating expenses / CAM: pass-throughs, base year, and any CAM cap
(state whether the cap is on controllable expenses only, and whether
it cumulates or resets each year)
Renewal options: number, length, notice window, rent mechanism
Termination / kick-out / go-dark rights: conditions and notice
Critical dates: renewal notice deadlines, expiration, option dates
Security deposit / letter of credit: amount and conditions
Assignment / subletting: consent standard
Estoppel / SNDA obligations
After the table, list every AMENDMENT you found and note which fields
above each amendment changed. Flag any field where the original lease
and an amendment disagree.Two things make this prompt earn its keep. It forces the model to say “NOT FOUND” instead of inventing a plausible value for a term the lease never contained, and it demands a page-and-quote citation, which turns your verification pass from a re-read into a two-second glance. The field list itself mirrors what a professional abstract captures — if you want the ground-level version of why these particular terms matter, our explainer on lease abstraction for CRE professionals covers the reasoning behind each one.
Minutes 4-6: let the model read
With a clean file and a sharp prompt, the reading itself is the fast part. A current model reads all 90 pages in one pass and returns the table, usually inside a minute or two.
Tool choice matters here mainly for length. A 90-page lease plus amendments is a large document, and the models differ in how much they can hold in a single view — Claude is built to take very long documents in one pass, which suits a full stack, while ChatGPT, Gemini, and Microsoft Copilot all handle a single lease comfortably. Any current general assistant will do the job for one document; the version number on the model matters far less than the quality of your file and your prompt. Purpose-built proptech tools wrap the same underlying capability in a CRE-specific schema and portfolio views, a trade-off our document intelligence playbook weighs in full.
When the table comes back, resist the urge to copy it into your file and move on. What you have is a well-formatted first draft, not a verified abstract. The next four minutes are what separate the two.
Minutes 6-10: verify by cost of error
You defeat the risk of a confident-but-wrong value without re-reading all 90 pages. You spot-check the handful of fields where an error is expensive, and you trust the rest on the model’s word.
Rank the fields by what a mistake costs, not by how hard they were to extract. A misread square footage, a wrong renewal-notice deadline, or a base rent captured as monthly when it was annual can cost a firm five or six figures; a misspelled guarantor middle name cannot. So verify these every time, using the page citations the prompt forced the model to provide:
- Economic terms: base rent for each year, escalations, free rent, and the CAM cap structure. Open the cited page and confirm the number.
- Critical dates: commencement, expiration, and especially the renewal-notice window — a blown notice date is one of the costliest errors in the business.
- Options and rights: renewal count and term, termination and kick-out conditions. Confirm the option actually exists and reads as the model summarized it.
- Amendment conflicts: any field the model flagged as changed by an amendment. This is the highest-value check of all, for reasons the next section makes plain.
Because the model handed you a page number and a quote for each value, confirming one takes a couple of seconds, not a re-read. Let the low-stakes fields — contact spellings, exhibit references — ride on the model’s word. Done this way, verification costs minutes per lease, and it is the step that makes the whole approach safe for a firm with no in-house legal team.
The amendment trap: why 90 pages is the hard case
A 90-page lease is rarely 90 pages of a single document. It is almost always an original lease plus a first amendment, a second amendment, an assignment, and an estoppel certificate — stapled into one PDF over the years. That structure is the defining hazard of the long file, and it is where naive summaries fail most expensively.
The trap is simple. The original lease sets base rent at $32.00 per square foot. A second amendment three years later resets it to $38.50. A summary that reports $32.00 is not a typo — it is a confidently wrong number pulled from a superseded page, and nothing on the surface flags it as stale. This is exactly why the prompt above ends by demanding a list of amendments and any field where the original and an amendment disagree.
When you verify, treat every amendment as a correction layer sitting on top of the base lease, and make sure the summary reflects the most recent version of each term. If the model reports a rent, a term length, or an option that an amendment quietly changed, that is the error that survives into your rent roll and resurfaces at the worst possible moment. The 90-page file is not harder because it is long; it is harder because length usually means layers.
What AI gets right, and what it gets wrong
Setting honest expectations beats any headline accuracy figure. Current document AI has a clear competence gradient, and knowing where your lease sits tells you how hard to check.
On standard, plainly stated terms in a clean file — base rent, straightforward percentage escalations, commencement and expiration dates, security deposit, defined renewal options — independent testing across leading commercial real estate document AI tools puts extraction accuracy in the low-to-mid 90s. On complex provisions such as options, exclusives, and co-tenancy, reported accuracy falls to roughly the high 70s to mid 80s, and on reconciling amendments it can drop further, into the 60s to 70s depending on the tool and the document quality. Those figures, drawn from vendor and third-party benchmarks, are field-level averages dominated by easy header fields, so accuracy on your hardest clauses is usually lower than the top-line number suggests.
What the model reads poorly is predictable: heavily negotiated or unusual clauses, complex expense-recovery structures, and anything a bad scan physically destroyed. What it cannot do at all is render legal judgment — whether a co-tenancy clause is actually triggered, or whether a default has occurred, stays human work. A tool that pretends otherwise is the one to walk away from. Used within those limits, though, the machine handles the enormous, easy majority of every lease in seconds, and your scarce attention goes only to the hard tail — the operating pattern behind our manifesto on how small CRE firms out-operate the giants.
When 10 minutes is the wrong goal
Summarizing one lease in a chat window is a brilliant use of AI. Abstracting your entire portfolio that way is not, and knowing the difference saves you from a bad decision.
The copy-paste-into-a-chatbot workflow is perfect for a deal in progress, a lease you are underwriting this week, or a term you need to check before a call. It falls apart at volume. When you are staring at 400 leases for an acquisition or a portfolio migration, the ten-minutes-per-lease math becomes days of manual driving, the outputs live in disconnected chat threads instead of a queryable database, and there is no systematic way to enforce the verification step. That is the point where a firm should weigh a purpose-built tool or a custom pipeline against the manual approach — a build-versus-buy question that turns on volume, not enthusiasm. For a single lease, though, the 10-minute clock in this article is the whole answer.
FAQ
Can ChatGPT summarize a 90-page commercial lease?
Yes. A current general assistant such as ChatGPT or Claude can read a full 90-page lease and return a structured summary of its key terms, usually in a minute or two of processing. The quality depends far more on your inputs than on the tool: a digitally native PDF and a prompt that names the exact fields to extract will produce a useful abstract, while a poor scan and a vague “summarize this” prompt will not. Claude is often preferred for very long documents because it is built to hold large files in a single pass, but any current model handles one lease. Treat the result as a first draft to verify, not a finished abstract.
How accurate is AI at summarizing a lease?
Accuracy depends on the field and the document, so a single number is misleading. On standard, plainly stated terms in a clean file — base rent, escalations, dates, security deposit — extraction across leading tools typically lands in the low-to-mid 90s. On complex provisions like options and co-tenancy it falls to the high 70s or mid 80s, and reconciling amendments can drop into the 60s to 70s. Published figures are also averages weighted toward easy header fields, so your hardest clauses score lower than the headline. Verify economic terms, critical dates, and options against the source every time.
Is it safe to upload a lease to ChatGPT or Claude?
It is safe if you use the right tier and settings, and risky if you do not. Use a paid business or team plan where the provider contractually excludes your inputs from model training, and switch off any “improve the model” option on the account. Avoid free consumer tiers for real deal documents. For an unusually sensitive lease, redact party and guarantor names before uploading — the model can still extract rent, dates, and options from an anonymized copy. The two-minute confidentiality check is what keeps this from becoming a disclosure problem.
Which AI is best for reading a long lease?
For a full 90-page stack, the practical differentiator is how much document the model can read in one pass, and Claude is designed to take very long files in a single view, which suits a lease plus amendments. ChatGPT, Gemini, and Microsoft Copilot all read a single lease comfortably as well. For one document, any current assistant works; the model’s version number matters far less than the cleanliness of your PDF and the precision of your prompt. If you are processing leases at volume rather than one at a time, a purpose-built proptech tool is usually the better fit.
What should I ask the AI to pull from a lease?
Ask for a structured table of named fields, not a prose summary. Name the fields explicitly: parties, premises and square footage, lease type, term dates, base rent for every year, escalations, free rent, CAM and any cap, renewal options with notice windows, termination rights, critical dates, security deposit, assignment consent, and estoppel obligations. Require the model to give a page number and a quoted clause for each value, and to write “NOT FOUND” for anything absent rather than guessing. Finish by asking it to list every amendment and flag any field where an amendment changed the original.
Does the AI read scanned lease PDFs?
It can, but scans are where quality drops. A crisp, high-resolution scan of printed text goes through optical character recognition with strong accuracy and behaves almost like a native file. A low-resolution scan, fax, or phone photo is where characters get misread — a “$32.00” becomes “$3,200,” or a “5%” cap becomes “3%,” silently. Before summarizing, try to select the text: if you cannot, re-scan the clean original at 300 DPI. Handwritten marginal notes and signature-line changes are something current tools cannot reliably capture, so read those by hand.
What lease terms does AI get wrong most often?
The most common errors are misread characters from bad scans, superseded values pulled from a page an amendment later changed, and hallucinated terms the lease never stated. What these share is that they produce a plausible, completed-looking answer rather than an obvious error, so they slip through unreviewed. Complex or heavily negotiated clauses — recovery structures, co-tenancy, unusual options — are the hardest for the model and deserve a clause-by-clause check. This is why the verification pass targets economic terms, dates, options, and amendment conflicts specifically.
Do I still need a lawyer or human review?
Yes, but far less of it, and only in the right places. The model handles the mechanical extraction; a human verifies the fields where an error is expensive and applies judgment the machine cannot. Legal questions — whether a clause is enforceable, whether a default has occurred, whether a co-tenancy right is triggered — remain a lawyer’s work, not a summary’s. For routine lease review, a broker or analyst spot-checking the cited economic terms and dates is the review the workflow needs. The AI removes the tedious reading, not the professional accountability.
How is an AI summary different from a real lease abstract?
A summary is a readable recap; a lease abstract is a structured, verified dataset your firm can rely on and query. The 10-minute workflow here produces the raw material of an abstract — the fields in a table — but it becomes a true abstract only after the verification pass confirms the expensive terms and reconciles the amendments. The distinction matters because an unverified summary carries the model’s silent errors into whatever you do next. If you want the full definition of what a professional abstract captures and why, our lease abstraction explainer lays it out.
Can I summarize my whole portfolio this way?
Not efficiently. The chat-window workflow is ideal for one lease at a time — a deal you are underwriting, a term you need before a call — but it does not scale to hundreds of leases. At volume, the per-lease minutes become days, the outputs scatter across disconnected threads instead of a queryable database, and there is no systematic way to enforce verification. A portfolio of that size is where a firm should weigh a purpose-built tool or a custom pipeline, a decision our document intelligence playbook works through in detail.
Key takeaways
- Ten minutes is real, but seven of them are yours — two minutes of file prep and confidentiality gating up front, five minutes of verification after. The model’s reading is the fast, easy part.
- The prompt is the whole game. Name the exact fields, force a “NOT FOUND” for anything missing, and demand a page-and-quote citation for every value so verification takes seconds.
- Set the confidentiality gate before you upload: a paid business tier with training turned off, and redacted party names for unusually sensitive leases.
- The 90-page file is hard because it is layered — original lease plus amendments — so the amendment reconciliation check is the most valuable verification you do.
- Verify by cost of error: confirm economic terms, critical dates, and options against the cited source every time, and let the low-stakes fields ride.
Wondering whether your firm’s lease pile is a clean 10-minute job or the kind of volume that needs a real pipeline — and where your review time should actually go? That is exactly what a short working session settles against your own documents. Book your free AI-readiness assessment →
Dirk Jan van Veen, PhD
