How to Abstract a Commercial Lease With AI

A commercial lease keeps changing after it is signed. A commencement letter fixes the dates the lease could only estimate. An amendment adds a suite; another extends the term and takes away an option.

Billing, the critical-date calendar and the budget run on the lease abstract, the short summary of each lease's rent, dates, space and rights, never on the lease itself. So every time a document is signed, a lease administrator re-reads the file and updates the abstract.

For a landlord with a few hundred tenants, we assume about 60 documents a month. At an assumed 75 minutes each, that is 900 hours a year, about $43,000 of an administrator's time. Miss one amendment and the abstract keeps billing the old terms.

AI agents now do the reading.

To show the work, we built one tenant's lease file at an invented shopping center: the lease signed in 2021, its commencement letter, an amendment from 2024 and a second from 2026 that exists only as a scan.

Every name and figure in it is invented, and we wrote down the right answers before any AI read a page. Here is what the three later documents changed:

We expected the models to trip on those changes. None did. Asked plainly to abstract the lease, Claude Sonnet, Claude Haiku and Gemini Flash each got every one of them right on this file, scan included, and so did every other run we made.

On this file, the reading was not where the abstract could go wrong. Below is the job as AI agents do it on that file, in five moves, then what Qvery builds so it runs on every lease you sign. The moves run as written on your own files.

This abstracts one retail lease. Office and industrial leases take the same reading with their own fields, and your lease accounting is built from the abstract, not here.

Give the AI Agents the Whole File, Scans Included

Upload every document in the file at once: the lease, its exhibits, the commencement letter and each amendment. The lease alone is wrong in each row of the table above, and its own summary page says so in small print:

Asked for the field list below, every run began with the documents it found, oldest first, and what each one changed. That list is the first thing to read, because a document missing from it is a document the abstract never saw:

The second amendment has no text layer: a tool that only reads text sees three blank pages. Every run read it, because these models read the page as an image.

A fax of a fax is another matter, so the prompt asks for every page the model could not read clearly. If one comes back on that list, get a better copy before anyone checks the abstract.

Ask for Every Field With the Page and the Words Behind It

This is the one prompt the job needs. Change the date to today's and the field list to the fields your lease system holds; the rest runs as written.

Attached is the complete file for one tenant's lease: the lease with its exhibits and every later document that changed it. Abstract the lease as it stands on November 1, 2026.

1. First list every document in the file, oldest first, with its date, its title and one line on what it changes. A later signed document overrides an earlier one only for the terms it changes. Work every date from the commencement date the parties confirmed, never from an estimate.

2. Then return one table with these columns: field, value, document, section, page, the words on the page that support the value, and sure or check. The fields: landlord; tenant; guarantor; premises; floor area; tenant's pro rata share; commencement date; rent commencement date; expiration date; minimum rent schedule (each period with its annual and monthly amount); monthly minimum rent on November 1, 2026; next rent change (date and new monthly amount); percentage rent; operating costs (share, and any cap with what it covers); management fee cap; real estate taxes; security deposit; renewal options (how many remain, length, rent); renewal notice window (earliest and latest date); early termination rights; co-tenancy; exclusive use; radius restriction; relocation; assignment and subletting; permitted use; holdover; estoppel deadline.

3. When a value comes from a later document, cite the later document and name the earlier value it replaced. When you worked a date out, show the arithmetic in one line. If the file does not support a field, write "not in file"; never fill it from what leases usually say.

4. Last, list every clause where a person has to decide something or watch for a condition, and every page you could not read clearly.

The plain request, "Please abstract the lease", got the same terms right. What came back was a memo:

Here are five rows of Claude Sonnet's table, copied into a sheet:

People who abstract leases for a living say what the reading costs them, and what they fear from AI:

The line in the prompt that answers that fear is the last sentence of point 3. On this file no run filled a field from what leases usually say.

Check the Values Against the Words, Not the Lease

The check runs down the table row by row: does the quoted text say the value beside it? The page is opened only where the words do not settle it.

Take the rent row. It says $17,158.33, cites the second amendment's Exhibit C-2 on page 3, and quotes "September 1, 2026 to July 31, 2027 ... $17,158.33". Page 3 of the scan:

On this file, all 36 of 36 pages the three runs cited for the terms we tested were the right pages. The Claude runs' quotes were word for word; Gemini twice rebuilt a table row in its own words, the right figure on the right page but nothing you could search for.

A row whose quoted words are not on the cited page is the row to stop at: the run either misread the page or cited the wrong document, and in both cases the value is unproven until someone opens the file. Asking for the words is what makes that visible.

Send the Clauses That Need a Judgment to a Person

Here the runs stopped agreeing. Every run got the terms right; they differed on what a person should look at.

The cap question is the one with money in it:

One run of six noticed that the tenant's renewal deadline falls on a Sunday, in a lease where time is of the essence.

It is the tenant's deadline and your planning date: the day you learn whether the space comes back in a year. Put the Friday before it in the calendar and ask counsel whether a Monday notice would count.

Some flags are not problems, and an abstract that files them as faults teaches your team to ignore its list:

• Blank initials lines at the foot of every page. Every signature block is signed.
• A late remeasurement. It was agreed after the lease's 12-month window, but both parties signed the amendment that fixed 6,050 square feet, so the amendment, not the window, sets the area.
• Routine obligations listed as watch items. One run returned 22 of them (opening hours, late fees, insurance). They belong in the lease system, not in front of a person every month.

Read the File Again the Day an Amendment Is Signed

An abstract is current only until the next document. So we tested the moment that keeps it current: we handed each model the abstract as it stood on March 1, 2026 and the new second amendment, and asked for an update.

All three changed the ten fields the amendment names. All three also changed five it never mentions, because they move with the new expiration date or the new rent.

Those five are the renewal notice window, the landlord's right to move the tenant, the window to end the lease after a fire late in the term, the half rent owed if the center's grocery store closes, and the rent if the tenant stays past the end. Claude Sonnet's update, beside the March abstract:

Two of the three marked those five as worked out rather than read, and asked a person to confirm them. One recorded the extra $6,900 deposit as already paid; the amendment says it is due on signing.

The risk that remains is the amendment that never reaches whoever updates the abstract.

On our file, an abstract that missed the second amendment would bill $13,108.33 a month from September 1 instead of $17,158.33: $4,050 a month, $48,600 a year, the whole rent of the new suite. It would also charge operating costs and taxes at 4.09% instead of 5.30%.

Let the AI agents draft the update; keep the approval. At first they change nothing on their own: they write the update with its sources, and a lease administrator approves it before billing, the calendar or anything a tenant sees changes.

Qvery's AI Agents Read Every Lease Document the Day It Is Signed

These five moves are what Qvery's AI agents run, on every new lease and every document that changes one. The first run on your own files comes two weeks after we can reach the folder where signed documents land and your lease system.

When a signed amendment, letter or new lease reaches the lease folder or the legal inbox, the AI agents read the whole file it belongs to, scans included, and draft the abstract update: every changed value with its document, page and words, and the dates that move with it.

They put the update and its questions in front of your lease administrator: the values to confirm, the clauses that need a decision, the deadlines that fall on a weekend. Once approved, the update goes into your lease system and the new dates into the calendar.

Take the example of a landlord with 60 documents a month, 75 minutes each by hand and 20 with the AI agents. A lease administrator costs $48.07 an hour (the federal median for property and real estate managers, the nearest occupation, plus benefits). The job costs $43,263 a year today, $60.09 a document.

With the AI agents it costs $14,969, $20.79 a document, including about $3,432 a year to run them. That is $28,294 saved and 660 administrator hours a year, and each $10,000 of the build pays for itself in 4.2 months. Put your own numbers in:

The calculator counts hours. Those hours go back to tenants and renewals, the work a growing portfolio otherwise hires another administrator for. The larger number is the amendment nobody passed along: on one tenant in our file, $48,600 a year of rent never billed.

Bring us three lease files with their amendments and the abstracts your system holds for them. We sit with how your administrators abstract today, map the fields your lease system and calendar use, and settle in writing which folders the AI agents may open.

Two weeks from the day the access is in place, the next signed document goes through them. From then on you get a weekly count: documents read, updates approved, questions raised and the hours your administrators got back.