AI in Real Estate

Commercial real estate is a business conducted entirely in documents, most of which were scanned by somebody in a hurry. That makes it one of the better fits for extraction — and one where the difference between a demo and a usable system is entirely in how the evaluation is reported.

What the documents are

The diligence pile for a single asset contains, at minimum: the leases with their amendments and any letters of intent that survived into them, the rent roll, the operating statements, the title commitment with its schedule of exceptions, the survey, the appraisal, service contracts, and estoppel certificates. They arrive in three formats and the mix matters more than it should — born-digital PDF for recent documents, scans for anything over a decade old, and spreadsheets for the rent roll, which is the worst of the three because it will have merged cells, a header row that is actually rows four through six, and totals interleaved with data.

A lease is long — a hundred pages plus exhibits is unremarkable — and the answer to a given question is in one paragraph of it, modified by a two-page amendment executed six years later. The retrieval unit is the property and tenancy, not the file, for exactly the reason set out in contract review.

The abstraction, field by field

Lease abstraction means producing a structured summary of a lease: the thirty to fifty fields an asset manager, an underwriter or an accountant needs without reading the document. A representative set:

• Parties and premises — landlord, tenant, guarantor, suite, rentable area, and whether the area is stated or measured.
• Term — commencement, rent commencement (often not the same date), expiry, and any early-access period.
• Economics — base rent schedule by period, escalations, free rent, tenant improvement allowance, security deposit and its burn-down.
• Recoveries — the operating expense method, base year, caps on controllable expenses, exclusions, gross-up provisions.
• Options — renewal, termination, expansion, contraction, right of first refusal, each with its notice window and its rent determination mechanic.
• Restrictions — exclusive use, co-tenancy, assignment and subletting, permitted transfers, continuous operation.

Now the part that decides whether the system is deployable: report accuracy per field against a set of leases a human has abstracted. Expiry date and rentable area will be excellent. Base rent schedules will be good. Operating-expense method, caps and exclusions will be materially worse, because they are expressed in dense prose with cross-references. A single headline figure averaged over fifty fields is not just uninformative, it is actively misleading, because the fields that are worst are the fields that carry the money.

Two design consequences follow. Extraction returns a page-anchored span for every field, so an analyst verifies in seconds. And every field can be absent — a lease with no co-tenancy clause is normal, and a system that invents one because the schema demanded a value is worse than a blank. Extraction confidence and schema edge cases cover the general shape.

The clause that is a formula in prose

The hardest field in the list is the escalation, because it is not a fact to be located. It is an algorithm written in English: “commencing on the first anniversary and on each anniversary thereafter, Base Rent shall increase by the greater of three percent or the increase in the Index over the preceding twelve months, provided that no such increase shall exceed five percent.”

Extracting that as a sentence is useless — a cash-flow model needs it as parameters. The right target schema is the structure, not the text: a floor, a reference index with its publication lag, a cap, a compounding basis, and an anniversary convention. A model asked for those specific fields does far better than one asked to “summarise the escalation”, and the extracted structure is then testable in the way nothing else on this page is: apply it to the known rent history and check that it reproduces the amounts actually billed. Where it does not, either the extraction is wrong or the billing is, and both are worth knowing. This is the same extract-then-reconcile discipline as finance, and real estate is one of the few other places where the independent check is sitting right there.

A cost model with your numbers in it

Whether this is worth doing is arithmetic, and it is arithmetic you can do before writing anything. The inputs are all yours:

per lease:
  input_tokens  ≈ pages × tokens_per_page          (a dense legal page runs
                                                    several hundred tokens;
                                                    measure yours, do not guess)
  passes        = 1 for whole-document context
                  or ceil(pages / pages_per_chunk) for a retrieval design
  output_tokens ≈ fields × tokens_per_field        (structured output is small)

  cost = passes × input_tokens × price_in + output_tokens × price_out
       + rescans / retries
       + OCR cost for the scanned share

portfolio:
  leases × cost              vs.    leases × hours_per_manual_abstract × rate

the number that actually decides it:
  hours_per_verified_abstract — how long an analyst takes to CHECK a generated
  abstract with page-anchored citations. If that is not well below
  hours_per_manual_abstract, nothing else in this model matters.

The last line is the one teams skip. A generated abstract that an analyst must verify line by line against the lease has moved the work rather than removed it, and the citation design is what changes that number. Measure it on twenty leases before scaling to two thousand. The general treatment of this kind of arithmetic is in cost per request and unit economics.

The part with real exposure

Everything above is back-office and low-risk. Two adjacent applications are not, and they are the ones a product team reaches for because they are visible: generating listing copy, and scoring prospective tenants.

Residential listing descriptions and tenant selection sit under anti-discrimination rules in most jurisdictions, and the rules typically cover not only refusing someone but also the wording of an advertisement — statements that indicate a preference, or that describe a neighbourhood in terms that function as a proxy for a protected characteristic. A generator asked to make copy warm and appealing will produce “perfect for a young family”, “in a quiet Christian community”, or a paragraph about school catchments, because those are frequent patterns in listing text. That is a compliance problem generated automatically at scale.

The control is a term blocklist and a human review path on residential copy, and — for tenant scoring — the same conclusion as recruitment: a model may extract checkable facts, and stated, applied-consistently criteria may decide, but a learned score over past accept/reject decisions reproduces those decisions. None of this is legal advice, and the specific rules differ by country and often by city.