AI Lease Abstraction: Why Claude Can't Manage Your CRE Portfolio

General-purpose AI is a remarkable lease reader. It still can't manage your portfolio, and we have the data to show why.

It's a fair question for any CRE team in 2026: if you can drop a lease into Claude and get a clean summary back in seconds, why pay for a lease intelligence platform? We wanted a real answer, not a hunch. So we ran a head-to-head test on a real portfolio and looked hard at what a general AI actually delivers.

We should say up front: we're fans. Claude reads leases well, and it reads a single clause as well as anything on the market. But reading a lease and managing a portfolio of them are two different jobs. The test made the gap between them very concrete.

We tested AI lease abstraction on 119 commercial leases

We took 119 office and retail leases across four buildings: 683 documents and 12,623 pages in all. Then we asked a general-purpose AI to abstract them the way a property manager would, with an ordinary prompt and a folder of PDFs per tenant.

To give it a fair shot, we handed it things most teams wouldn't have on day one:

• A curated folder of documents for each tenant, with summaries and stray uploads already removed
• A 50-field abstract template with plain-English definitions of every term
• The instruction to answer "as things stand under the latest document"

Every answer was checked against values that experienced lease reviewers had already approved, and against the page it cited. In other words, the general AI started with a head start a buyer would have to build themselves. Without that scaffolding, "upload your leases to an AI" isn't a product. It's a project.

Where AI lease abstraction falls short: rent roll and critical date fields

On fields that just need careful reading, like holdover, late fees, permitted use and signage, the general AI did well. The trouble started on fields where the right answer depends on how lease administrators record the value, not just what the clause says.

Share of answers that were correct and traceable to the cited page, across the same 119 leases.

In many of these misses, the AI read the clause correctly. It just didn't know the form the answer needs to take. A rent bump, for example, is recorded as a flag on the rent schedule, not as a sentence. Rent commencement is recorded only when it differs from the lease's commencement date.

These aren't obscure fields. They're the inputs your rent roll and critical-date calendar are computed from. And the conventions behind them live in years of reviewer decisions across more than 275,000 documents, not in a prompt.

AI hallucinations in lease abstracts: rights and deposits that don't exist

The general AI gave 1,187 answers for terms reviewers had not recorded on those leases at all. On one building, it reported 43 early-termination rights where reviewers had found 29. A few examples of what it produced:

• An early-termination right marked "expired, not exercised," a status the documents never establish
• A security deposit "reduced from $33,000 to $12,500," where the cited text only says the existing deposit stays in place
• A renewal option marked "0 remaining, exercised," where the lease shows one option and nothing about its exercise
• A temporary 2020 rent reduction reported as the lease's rent abatement, with amounts the cited text doesn't contain

Every one of these reads like plausible lease administration. That's the danger. In a spreadsheet, a phantom termination right looks exactly like a real one, and it can quietly flow into a valuation, a lender package or a renewal decision.

Tracking lease amendments: why every value needs a source

The leases in our test averaged 5.7 documents each, and one ran to 39. Fifty of the 119 had five or more, and 202 of the 683 documents were extension amendments. That's normal for a mature portfolio.

We asked both systems a simple follow-up for each document: which terms does this one set? The general AI scored 0.57 on that measure. Prophia scored 0.88.

That gap matters the moment someone asks "says who?" A rent roll audit, a tenant dispute or a lender's review needs the history of a term: its original value, every amendment that changed it, and the words behind each change. The general AI can tell you where things stand today. It often can't tell you how they got there. In Prophia, every current value links to the document and highlighted clause that set it.

Lease review and audit: what happens after the AI is done

Of the 119 leases, the general AI produced zero without at least one error. That isn't a knock on Claude. It's the reality of AI extraction today: no model's raw output, ours included, should go straight into your system of record unchecked.

So the question that decides whether AI helps you isn't "how well does the model read?" It's "what happens to an output that's mostly right?"

With a general AI, the answer is a spreadsheet. A reviewer gets a file name and page number, goes back to the PDF and hunts for the clause. A correction is a cell edit with no record of what it replaced.

Prophia is built around that review step:

• Every value is a box on the page. Click a term and the document opens to the highlighted words. Checking a value takes one click.
• Every tag has a review state. Values start as "needs review" and are approved or corrected, with the original prediction kept so corrections are visible.
• A full audit trail. Who changed what, when, and what the value was before.
• QA flags point at the problem. For example, a rent schedule that ends before expiration, or a right an amendment struck out.
• Review is prioritized. Lease scores, open flags and tasks show which leases need attention first.

That layer of AI plus human review is what turns a fast first pass into data you can actually act on.

What lease administration software does that a spreadsheet can't

Even a perfect abstract is a snapshot. Managing a portfolio means the data has to do work, and keep doing it as leases change. Here's what a general AI's spreadsheet can't do on its own:

• Rent roll. Computed directly from reviewed rent schedules, abatements and suites, so a corrected value updates everything downstream.
• Stacking plan with future dates. See any building as it will look on a future day, with each suite's first-offer, first-refusal, termination and renewal rights linked back to the lease.
• Critical dates. Reminders for expirations and option deadlines, so nothing slips.
• Yardi audit and integrations. Prophia's rent roll compared suite by suite against your accounting system, with mismatches flagged. Lease data syncs to Yardi and reads units from MRI.
• Search and chat with receipts. Every answer links to the page and highlighted clause it came from.
• Re-tagging when the next amendment lands. New documents are reconciled against the lease's existing data, struck rights are marked, and unchanged terms carry forward. With a general AI, a new amendment means running the building again and reconciling two spreadsheets by hand.
• Access for your own AI tools. A REST API and MCP server let tools like Claude read your reviewed, trusted data instead of re-reading the PDFs.

The bottom line on AI in commercial real estate

General-purpose AI is a capable reader, and for a one-off look at a handful of leases it can be a reasonable tool. But your portfolio needs more than reading. It needs a definition of what to record and how, a history behind every value, a review process you can trust, and data that feeds your rent roll, critical dates and accounting system, then stays current as amendments arrive.

That's the job Prophia was built for. AI does the first read; Prophia turns it into a lease record your team can rely on.

See it on your own leases. Book a demo and we'll show you what your portfolio looks like when every value comes with its source.