A signed letter of intent on a 14-tenant retail center gives a buyer roughly 30 days to learn everything the seller already knows, and the data room lands with just under 400 documents in it. Leases and their amendments, estoppels, a rent roll, service contracts, title, an environmental report, and a folder of exhibits that turn out to hold the real terms. What follows is what five disciplined days look like when a lean team runs the review with AI carrying the reading and a person owning the decision. It is a method, not a miracle, and the honest parts are the ones the vendor demos skip.
The deal that forces the method
Due-diligence windows in commercial real estate commonly run 30 to 60 days, and complex deals push past 90. The bottleneck inside that window is almost always the same: reading the leases. Manual lease abstraction takes roughly four to six hours per lease when done properly, so a 14-tenant center with layered amendments is a week of one person’s undivided attention before anyone looks at title or environmental. A 100-lease acquisition at four hours each is about 400 analyst-hours, close to ten working weeks of one reviewer against a clock measured in days.
AI changes the shape of that problem without removing the risk. Reported deployments bring per-lease extraction down to 20 to 40 minutes, and one cut a two-hour review to about 17 minutes. Run that in parallel across a data room and a review that would have taken six to eight weeks compresses toward days. The catch is that speed on the easy 90% means nothing if the tool reads a co-tenancy clause wrong with full confidence, so the method has to be built around that catch, not around the speed.
The reading and structuring layer underneath all of this is covered in our guide to turning lease stacks into structured data; this piece points that layer at a live deal on a deadline.
Day 1: triage before you extract a single field
The first day produces no abstracts. It produces a map. Before any document goes near an extraction prompt, someone who knows deals sorts the 400 files into four piles: money terms (leases, amendments, guaranties, the rent roll), legal risk (title, environmental, litigation), context (service contracts, insurance, tax bills), and noise (duplicates, blank scans, flyers that ended up in the room).
This triage is the highest-value hour of the week, and no tool does it for you. A person who has closed deals knows that the estoppel folder is where a landlord’s version of a lease gets contradicted by the tenant’s, and that a “second amendment” with no first amendment in the room is a flag, not a typo. AI can count and de-duplicate the pile, but deciding which pile a document belongs in is human judgment, and getting it wrong on day one poisons everything downstream.
Day one also sets the guardrails. Decide now which fields you will trust AI to read straight through (tenant name, suite, square footage, lease start and end, base rent) and which you will read yourself no matter what the model says (co-tenancy, exclusives, percentage rent, renewal and termination options, CAM caps, landlord obligations). That split is not a temporary crutch you automate away later; it is the correct division of labor for a document type where a single misread exclusive can cost a tenant relationship. Which documents an AI can and cannot read cleanly is worth internalizing before you start, which we lay out in our field guide to the CRE documents AI can and cannot read.
Days 2 and 3: bulk extraction and the rent roll
With the piles set, the money-terms stack goes through extraction in batches. In a business-tier general model with a saved prompt that returns your standard abstract fields, or a purpose-built product if volume justifies one, a reviewer feeds documents in and gets structured output back. The target for these two days is a first-pass abstract for every lease and a draft rent roll assembled from those abstracts, not copied from the seller’s spreadsheet.
That distinction matters more than any feature. The seller’s rent roll is a claim; the rent roll you build from the underlying leases is evidence. When the two disagree, and they will, you have found something the diligence exists to find. A base rent that steps up in the lease but sits flat on the seller’s roll, a tenant shown as current whose lease expired eight months ago on a holdover, a recovery method that the roll rounds off and the lease spells out. AI makes building the independent rent roll fast enough to do inside the deal clock, the single biggest change it brings to a small team’s diligence.
Two rules keep days two and three from producing confident garbage. First, demand grounding: every extracted number should link back to the clause and page it came from, so verification is a click rather than a hunt. Second, treat every headline accuracy figure as a claim about the easy fields on clean documents. Vendors commonly cite accuracy above 95%, which holds for parties, dates, and base rent on a native-text lease but falls on scanned amendments and unusual clauses. What those percentages quietly exclude is worth understanding before you rely on one, which we break down in our look at what a 95% document-AI accuracy claim leaves out.
Day 4: the exceptions only a person can settle
By day four the machine has done what it is good at, and the job becomes the opposite of bulk processing. A reviewer works the exception queue: every field flagged as low confidence, every clause on the day-one “read it yourself” list, and every disagreement between the built rent roll and the seller’s.
This is where the reading gets slow on purpose. Co-tenancy provisions, percentage rent with breakpoints, CAM caps with base-year mechanics, conditional landlord contributions, and cancellation options are the clauses where extraction is least reliable and exposure is highest. A model will summarize a co-tenancy clause fluently and still miss that the trigger is tied to a named anchor that just filed for bankruptcy. The exhibits are the other trap: the operative rent schedule or work letter often lives in an exhibit the model treated as an appendix, and a diligence that skims the exhibits has read the wrong document.
A specific failure mode has to be checked deliberately rather than trusted away. Document AI does not fail by leaving a field blank; it fails by filling one in with a plausible value that is not in the source: a renewal option the lease does not grant, a security deposit rounded to a number that appears nowhere. Catching that means reading the model’s answer against the clause, not instead of it, and the concrete checks for these invented values are worth building into the review, which we cover in our piece on document AI’s hidden hallucination problem.
Day 5: reconciliation and the go or renegotiate call
The last day turns a stack of verified abstracts into a decision. Reconciliation means three things: the independent rent roll ties to the leases behind it, the exceptions are resolved or listed as open risks, and the findings that change the underwriting are pulled to the top of a one-page memo the principal can act on.
The deliverable is never the abstracts. It is the short list: the three tenants whose renewals fall inside the hold period, the co-tenancy clause that lets the second-largest tenant drop to half rent if the anchor goes dark, the two leases where the seller’s stated recovery income overstates what the CAM language allows. Those findings are what a buyer renegotiates on or walks from, and producing them in five days instead of five weeks is the difference between diligence that informs the deal and diligence that rubber-stamps a price already fixed.
None of this replaces counsel on the clauses that need a lawyer. The method gets the routine 90% off a human’s plate fast enough that scarce hours land where they matter: the exceptions, the exhibits, and the reconciliation. That reallocation is the real edge a small firm holds over a larger competitor drowning in process, an argument we make in full in the small-firm AI playbook.
Where AI carried the work, and where it did not
Split the week by who did what:
| Task | Carried by AI | Owned by a person |
|---|---|---|
| Sorting and de-duplicating the data room | Counting, flagging duplicates | Deciding which pile each file belongs in |
| First-pass abstracts of standard fields | Bulk extraction with grounding | Spot-checking against source |
| Building the independent rent roll | Assembling from abstracts | Confirming it ties to the leases |
| Money clauses and options | Drafting a summary | Reading every one against the clause |
| Exhibits and work letters | Surfacing them | Judging whether they are operative |
| The go / renegotiate memo | Nothing | Everything |
Read down the right column and the pattern is plain: AI removed the volume, not the judgment. A firm that expects the tool to produce the decision has misunderstood the trade; a firm that expects a fast, reliable first draft of everything except the money clauses has understood it exactly.
When to reach for a purpose-built tool
The five-day method runs on a general model with a saved prompt for many small firms, which is frequently the honest answer under modest volume. The threshold where a purpose-built product earns its price depends on how often you do this and how much structure you need afterward.
Under a handful of deals a year, a thin workflow in ChatGPT, Claude, or Gemini on a business tier plus the discipline in this article is usually enough. When diligence is your core loop and you abstract dozens to hundreds of leases across recurring acquisitions, a dedicated platform earns its keep through a persistent data store, audit trails, and side-by-side source review. Prophia markets AI-accelerated abstraction and a due-diligence binder on roughly a two-week turnaround; Kira, now part of Litera, markets deep clause extraction for legal-grade M&A review. Verify each vendor’s current capability against its own documentation before you buy, because proptech features shift every quarter.
Both cost less than one blown acquisition, which is the number the diligence exists to protect: a market-rate training workshop that builds the fluency to run the general-model method runs roughly $2,000 to $15,000, and a scoped custom automation for a firm doing this at volume runs roughly $25,000 to $150,000.
Frequently asked questions
What is AI-assisted due diligence in commercial real estate?
AI-assisted due diligence uses document AI to read the leases, amendments, and financials in a data room, extract structured data, and surface exceptions, so a reviewer spends their hours on judgment instead of bulk reading. The AI decides nothing: it produces first-pass abstracts and an independent rent roll in days, and a person verifies the money clauses and writes the findings that change the underwriting.
How long does AI-assisted due diligence take?
For a data room of a few hundred documents, a disciplined AI-assisted review can produce a decision-ready package in about five working days, against the four to six weeks a fully manual review would take. On larger portfolios, AI-accelerated review compresses a six-to-eight-week window toward three to four weeks by running extraction in parallel. The routine fields drive the speedup; the money clauses and exhibits still take human time.
Can AI abstract a full data room without human review?
No. AI produces a reliable first pass on standard fields, but a human has to verify the clauses where money and legal risk concentrate, because that is where accuracy drops. Co-tenancy, percentage rent, CAM caps, renewal and termination options, and terms buried in exhibits are most likely to be misread with full confidence. Keep a human in the loop on those and treat the tool as a fast first reader.
Which documents does AI struggle with in CRE due diligence?
Scanned amendments, faxed estoppels, hand-annotated riders, rotated pages from old ground leases, and exhibits that hold the operative terms are where naive extraction breaks down. Clean, native-text leases read well; the messy documents that fill a real data room read worse, and the model rarely tells you it struggled. Assume anything scanned, annotated, or non-standard needs a person to confirm the values against the source.
Should a small CRE firm use a general AI model or a purpose-built tool like Kira or Prophia?
For a firm doing a handful of deals a year, a business-tier general model with a saved abstraction prompt plus disciplined human review is usually enough and far cheaper. A purpose-built platform like Prophia or Kira earns its price when diligence is your core loop and you abstract dozens to hundreds of leases across recurring acquisitions, where the persistent data store and source-linked review save real time. Count your deal frequency before buying a portfolio-scale subscription.
How accurate is AI lease abstraction for due diligence?
Vendors commonly report accuracy above 95% on standard fields such as parties, dates, and base rent, and that holds on clean documents. Accuracy falls on non-standard clauses and poor scans, where the financial risk sits, so headline figures are claims about the easy cases. Treat the tool as accurate enough to draft the routine abstract and never accurate enough to skip review of the clauses that decide money.
Is it safe to upload a data room to an AI tool?
Only on the right plan. The safe pattern is a business or enterprise tier where your inputs are not used to train the model by default, backed by SOC 2 Type II or equivalent certification. ChatGPT and Claude both offer tiers that contractually exclude your data from training and hold SOC 2 Type II; ask any proptech vendor for the equivalent. Read the data-processing terms of the specific plan, not the marketing page.
What does AI-assisted diligence cost for a small firm?
Running the method on a general model costs little beyond a business-tier subscription plus the time to build fluency, which a market-rate training workshop delivers for roughly $2,000 to $15,000. A purpose-built platform is a subscription scaled to volume, and a custom automation for a firm doing diligence at scale runs roughly $25,000 to $150,000. Weigh any of these against the cost of one missed clause on one deal.
What is the first thing to do with a data room before running AI on it?
Triage it by hand. Before any document goes through extraction, someone who knows deals sorts the files into money terms, legal risk, context, and noise, and decides which fields the AI can read straight through and which a person reads no matter what. That first hour of human judgment sets up everything downstream, and skipping it, so the model reads a duplicate or misfiles an estoppel as a lease, is the most common way an AI-assisted review goes wrong on day one.
Where to start
Before you point AI at a live deal, get an honest read on two things: whether your team is fluent enough to judge what the tool produces, and which of your real documents will defeat it. A free AI-readiness assessment gives you that read, a short working session that maps your document mix, deal cadence, and workflows, and returns a plain recommendation on whether a general-model method, a purpose-built platform, or a month of fundamentals first is your next step. Book a free AI-readiness assessment before the next data room lands.
Dirk Jan van Veen, PhD
