The binding constraint at most 4-20 person CRE firms is not deal flow. It is screening capacity. Broker blasts arrive faster than anyone can underwrite them, so the firm defaults to analyzing whatever landed most recently, or whatever came from the broker who calls the most. AI changes that math: a principal with a written buy box and a well-built prompt can triage an inbox in a fraction of the time it took by hand, and spend the recovered hours underwriting deals that deserve it. This playbook walks through the full funnel, from broker-blast triage to comp selection to model population, and is blunt about the one part that gets skipped most often: the specific places where AI output must be verified by a human before money moves.
The Screening Bottleneck Nobody Staffs For
A lean acquisitions shop sees somewhere between a handful and several dozen opportunities a week: listing blasts from CoStar and Crexi alerts, LoopNet listings, flyers from broker relationships, the occasional off-market teaser. Each one demands the same first question. Does this fit what we buy?
Answering that question manually costs 30 to 90 minutes per deal once you open the OM, find the rent roll, and rebuild enough of the numbers to form a view. At institutional firms, junior analysts absorb this cost. At a 10-person firm, the person doing it is often a principal whose time is worth more than the analysis. So screening gets rationed, and rationed screening means missed deals.
The industry has noticed. Deloitte’s 2026 commercial real estate outlook found 73% of CRE firms now view AI as crucial for advanced analytics and market-signal detection, while only 7% report transformative impact so far. That gap between belief and result is the interesting part. In our read, most firms bought tools before designing a workflow. This playbook is the workflow.
The Playbook at a Glance
Five stages, each with a clear owner:
- Triage broker blasts against a written buy box. AI does the reading; you make the kill decision.
- Screen survivors by extracting the OM’s actual numbers into a standard format.
- Select comps with AI assistance, then interrogate every comp before it enters the model.
- Populate the underwriting model from extracted data, keeping the model itself human-owned.
- Verify the specific outputs where language models fail most often on CRE documents.
Stages 1 and 2 are where the time savings live. Stages 3 through 5 are where deals are won or lost. Treat them differently.
Stage 1 — Triage Broker Blasts With a Written Buy Box
Write the Buy Box Down First
AI triage only works if your criteria exist outside someone’s head. Most small firms carry the buy box as tribal knowledge: “we do neighborhood retail and small-bay industrial, Southeast, $2M to $15M, value-add preferred.” That sentence is a start, not a screen.
A usable buy box specifies, at minimum: asset types and subtypes, markets and submarkets, deal size range, minimum going-in cap rate or maximum price per unit or square foot, hold strategy, deal-breaker conditions (ground lease, single-tenant with under five years of term, flood zone, deferred maintenance beyond a threshold), and the return hurdles a deal must clear at screening assumptions. Write it as a one-page document. You will paste it into prompts daily, and it doubles as training material for anyone new.
A Daily Triage Workflow in ChatGPT or Claude
The workflow we recommend to firms starting out costs nothing beyond a standard AI subscription:
- Once a day, collect new deal materials: OM PDFs, flyers, emailed teasers.
- Open a fresh conversation in ChatGPT or Claude, paste the buy box, and attach the documents.
- Ask for a structured verdict per deal: pass, decline, or insufficient information, with the two or three facts driving each verdict and a list of what is missing.
- Log every verdict in a simple tracker with a one-line reason.
The prompt matters less than the buy box. A mediocre prompt with precise criteria beats an elegant prompt with vague criteria every time. This is also the single highest-value exercise from LLM fluency training: once a team can write a buy box the model can execute against, the same skill transfers to lease summaries and market write-ups. If your team has not built that muscle yet, the CRE AI training playbook covers how firms get there in 90 days.
What Triage Should Not Decide
Triage kills deals; it should never green-light them. The model’s role is to say “this fails your stated criteria because the OM discloses a ground lease” and save you the read. When a deal passes triage, the output is not a buy signal. It is permission to spend real screening time. Firms that blur this line end up trusting a summary where they should be trusting a rent roll.
Off-market teasers deserve a special rule: when information is thin, the honest verdict is “insufficient information,” not a guess. Instruct the model to say so explicitly. Language models fill gaps confidently by default, and a confident guess about an unstated cap rate is worse than no answer.
Stage 2 — Structured Screening From the OM
Extract the Numbers, Not the Narrative
An OM is a marketing document. The screening step exists to separate the property’s numbers from the broker’s story about them. For each deal that survives triage, extract into a standard screening sheet: in-place NOI as stated, the T-12 revenue and expense lines if included, current occupancy, rent roll summary (tenant count, major tenants, WALT for commercial, average in-place rent for multifamily), asking price and implied cap rate, and the pro forma assumptions the broker used to make the deal look good.
That last item is deliberate. Asking an AI to list the OM’s pro forma assumptions, side by side with in-place figures, is the fastest way to see how much of the deal is real and how much is projection. A deal asking a 5.2% cap on pro forma NOI that is 30% above in-place is a different conversation than the flyer suggests.
Purpose-built platforms have moved fast here. Dealpath’s AI Extract ingests structured data from OMs and flyers and generates screening tear sheets in seconds, and its deal-screening tooling processes OMs, rent rolls, T-12s, and broker opinions of value (verified against Dealpath’s public product documentation at the time of writing; proptech AI features change quarterly, so re-verify before buying). Whether a platform subscription beats a general LLM for your volume is an economics question we address below.
Score Against Criteria, Then Rank
With a standard screening sheet per deal, ranking becomes mechanical: which deals clear the return hurdles at your assumptions, not the broker’s? Run screening-level numbers with your own expense ratios and your own exit cap. The AI can do this arithmetic in the same conversation, but the assumptions must be yours, written in the buy box document.
The output of Stage 2 is a short ranked list and a decision: which one or two deals earn a full underwrite this week. Screening more deals is only useful if it concentrates effort on better deals.
Stage 3 — Comp Selection You Can Defend
Where AI Genuinely Helps With Comps
Comp work has two halves: finding candidates and defending selections. Purpose-built data products are strong at the first; unsupervised AI is dangerous at the second.
For rent comps in multifamily, purpose-built data products now do the collection at a scale no small firm can match manually. HelloData surveys rents across more than 38 million units daily from public listing data, claims a 97% match rate against actual rent rolls, and reports its algorithmic comp picks overlap appraiser selections roughly nine times out of ten (vendor-stated figures, verified against HelloData’s public site at the time of writing). For sales comps, CoStar remains the default data source, and a general LLM is useful for summarizing and comparing candidate comps you feed it, not for sourcing them from its own memory.
That distinction is the rule to tattoo somewhere visible: never let a language model generate comps from memory. Models trained on public data will produce plausible-sounding sale transactions with plausible-sounding prices, and some of them will not exist. Comps come from a database or a broker; AI organizes and pressure-tests them.
Interrogating an AI-Proposed Comp Set
When a platform or a model proposes a comp set, run the same interrogation an appraiser would face:
- Which comps would a lender’s appraiser throw out, and why?
- What adjustments does each comp require (vintage, location quality, unit mix, deal timing), and which direction do they push value?
- Which single comp, if removed, moves the conclusion most?
Paste your comp set into ChatGPT or Claude and ask exactly those questions. This is the highest-value use of a general LLM in the comp process: it argues against your selection cheaply, before a lender or an LP does it expensively. The final set, and the responsibility for it, stays with the human who signs the memo.
Stage 4 — Populating the Underwriting Model
From Extracted Data to Excel or Argus
Most lean firms underwrite in Excel; office and retail shops with institutional capital partners often model in Argus. In both cases the practical AI win is the same: the transcription layer disappears. Rent rolls and T-12s arrive as PDFs, and retyping them is where analyst-hours go to die. This is document intelligence applied to deal analysis, and the extraction techniques are identical to the ones covered in depth in the CRE document intelligence playbook: AI reads the PDF, outputs structured tables, and the tables load into your model template.
Two implementation notes from automation work we have shipped in adjacent professional-services domains, where the documents are different but the failure modes are identical. First, extraction accuracy is document-dependent: clean institutional rent rolls extract nearly perfectly, while a scanned rent roll from a mom-and-pop seller with handwritten notes will produce errors, and the errors will be silent. Second, always extract to an intermediate table a human can eyeball before anything touches the model. The intermediate table is the checkpoint; skipping it converts a five-minute review into a corrupted IRR nobody catches until diligence.
Keep the Model Human-Owned
Do not ask AI to build or restructure the underwriting model itself. Your model encodes the firm’s judgment: how you treat vacancy, when you burn off concessions, what capex reserve you carry per door or per foot. A model an AI assembled is a model nobody at the firm fully understands, and the first time a capital partner asks “why is month-37 refinance sized this way,” that becomes obvious. AI populates inputs and drafts scenario commentary. The formulas, the structure, and the assumptions belong to a named human.
Scenario work is the exception where AI adds analytical value inside the model workflow. Once the base case is built and verified, asking a model to propose stress scenarios (what combination of exit cap expansion and rent softness breaks the deal covenant?) and then running them yourself is faster than inventing stresses from scratch, and occasionally surfaces a combination you had not considered.
Stage 5 — The Verification Protocol
This is the section most AI-in-CRE content waves at with “keep a human in the loop.” A loop is not a protocol. Verification only works when effort concentrates where language models fail on CRE documents specifically. Four failure modes account for most of the damage.
Footnotes and one-time items. Models read tables well and footnotes inconsistently. A T-12 with a one-time insurance settlement in other income, or an expense line footnoted as “excludes management fee, paid by affiliate,” will extract cleanly and mislead completely. Verification: a human reads every footnote and every “other” line in the source document. This is minutes, not hours.
Concessions and effective rent. OMs quote face rents; the economics live in concessions, abatements, and TI packages. Models summarize what the document states and rarely flag what it omits. Verification: reconcile the stated average rent against the rent roll math yourself, and ask the broker directly what concessions are outstanding.
Lease rollover masking. A healthy WALT can hide a cliff: three tenants representing 40% of income all expiring in the same 18 months. Extraction gets the individual dates right; the risk pattern requires someone to look. Verification: sort the extracted rent roll by expiration and eyeball the cumulative income at risk per year. AI can build that table; a human must read it.
False precision. The most dangerous outputs are the ones that look finished: a tear sheet with a cap rate computed from an NOI that silently mixed in-place revenue with pro forma expenses. Every derived number in an AI-produced summary needs its lineage checked once: which source figures produced it? If the answer is unclear, recompute it.
The full verification pass on a screened deal takes 20 to 30 minutes with the source documents open. That is the price of using AI on decisions of this size, and it is a bargain against the alternative, which is either not screening the deal at all or trusting an unverified summary into an investment memo.
Tooling Economics for a Lean Firm
Three tiers, honestly compared:
| Tier | What it looks like | Cost profile | Fits when |
|---|---|---|---|
| General LLM | ChatGPT, Claude, or Gemini subscriptions plus written buy box and prompt discipline | Tens of dollars per seat per month | Deal flow under ~10 serious screens a month; workflow still stabilizing |
| Purpose-built platform | Dealpath, HelloData, and similar per-seat proptech subscriptions | Hundreds to low thousands per month depending on seats and data | Consistent deal flow in one asset class; team wants extraction plus data in one place |
| Custom automation | A built-for-you pipeline from inbox to screening sheet to model template | Market range for custom CRE automation projects runs roughly $25K-$150K depending on scope | High or multi-channel deal flow; a workflow that is stable, proven manually, and worth compounding |
The ordering is the advice. Firms that jump to custom automation before proving the manual-plus-LLM workflow buy software that encodes a process they have not debugged. Firms that stay on general LLMs after deal flow scales pay in principal-hours instead. The decision framework for that progression, including when off-the-shelf proptech is simply enough, gets full treatment in the CRE AI buy-vs-build playbook.
Confidential Deal Data in AI Tools
Deal documents are confidential, frequently under NDA, and small firms are right to hesitate before uploading them anywhere. The workable posture has three parts.
Use business-tier AI accounts, not free consumer ones. ChatGPT, Claude, and Gemini business offerings contractually exclude customer inputs from model training by default; free tiers vary. Read the current data-use terms for the product you deploy, because they change.
Strip what you can when it costs nothing. For triage, a flyer rarely needs the seller’s name to be screened against a buy box. For full screening, document redaction usually costs more time than it saves risk, which is why the account-tier decision matters more.
Put the rule in writing. A one-paragraph internal policy (which tools are approved, which account tier, what never gets uploaded) prevents the real risk, which is a well-meaning team member pasting an NDA’d rent roll into a personal free account. Communication and CRM data raise the same questions with different documents, covered in the CRE communications playbook.
Where Deal Analysis Fits in the Bigger System
Deal analysis is one of four automation domains where small CRE firms are quietly out-operating larger ones, an argument made in full in the small CRE firm AI manifesto. The screening funnel described here gets stronger when its neighbors are built: lease-stack extraction feeds better rent roll data into underwriting, and the back-office side (rent rolls, CAM reconciliation, investor reporting, covered in the CRE back-office automation playbook) reuses the same document-extraction discipline after the deal closes.
The sequencing insight from firms that do this well: the buy box document, the screening sheet template, and the verification checklist are the durable assets. Tools will change; those three artifacts compound.
Frequently Asked Questions
Can I just use ChatGPT or Claude to screen CRE deals?
For triage and first-pass screening at modest deal flow, a business-tier ChatGPT or Claude subscription plus a written buy box handles the job well. The general LLM approach breaks down when volume grows (no persistent database, manual document handling every time) or when you need market data the model does not have, which is where platforms like Dealpath or data products like HelloData earn their subscriptions. Start with the general LLM; let the pain tell you when to upgrade.
How accurate is AI at pulling numbers from an OM or T-12?
Accuracy is document-dependent, which is why blanket percentages from vendors deserve skepticism. Clean, digitally generated OMs and institutional rent rolls extract with few errors. Scanned documents, unusual layouts, and footnoted adjustments produce silent mistakes. The operational answer is not a percentage; it is a workflow: extract to an intermediate table, have a human eyeball it against the source, then load the model. That review takes minutes and catches the errors that matter.
Will AI screening make me miss good deals?
The risk runs the other way in practice. Manual screening under time pressure already misses deals, silently, because nobody opens the twentieth OM of the week. A written buy box executed consistently by AI screens everything and documents why each deal died, which means you can audit your own criteria quarterly and loosen a filter that is killing deals you wish you had seen. The tracker of declined deals with reasons is the safety net most firms have never had.
Does AI replace an underwriting analyst?
At a 4-20 person firm the honest question is different: most of these firms never had a dedicated analyst to replace. AI gives principals and brokers analyst-level throughput on extraction, first-pass screening, and scenario drafting. Judgment work (final comp selection, assumption setting, deal structure, the decision to offer) stays human, and the verification protocol adds a task that did not exist before. The realistic outcome is more deals screened per person, not fewer people.
What should I always check by hand before trusting an AI-assisted underwrite?
Five things, every deal: footnotes and one-time items in the T-12, concession and effective-rent math, lease expiration concentration, the lineage of every derived number in the summary (especially NOI and cap rate), and any figure the source documents state twice with different values. Budget 20 to 30 minutes with source documents open. Anything that fails a check gets recomputed from source, not patched from memory.
Is it safe to upload confidential deal documents to AI tools?
With business-tier accounts and a written internal policy, the risk is manageable and comparable to other cloud software your firm already uses for the same documents. Business offerings from the major AI products exclude customer inputs from training by default; verify the current terms for your product and tier. The real-world failure mode is an employee using a personal free account, which is a policy problem, not a technology problem.
Can AI pick rent and sales comps for me?
AI can propose, organize, and stress-test comps; it must never invent them. Data products like HelloData automate rent comp collection at a scale no small team matches manually, and general LLMs are effective at arguing against a candidate comp set before a lender’s appraiser does. Sales comps come from CoStar or broker knowledge. Any comp a language model produces from its own memory is unverified and possibly nonexistent until you confirm it in a database.
What does AI deal-analysis tooling cost a small firm?
Three tiers. General LLM subscriptions run tens of dollars per seat monthly. Purpose-built proptech platforms run hundreds to low thousands monthly depending on seats and data access. Custom automation projects, where a firm’s specific inbox-to-model workflow gets built as a pipeline, run roughly $25K-$150K in the current market depending on scope. Team AI training workshops, if you need to build fluency first, run roughly $2K-$15K in the market. Sequence them in that order.
Can AI handle off-market deals with almost no documentation?
Thin-information deals are where AI discipline matters most, because the temptation to let the model fill gaps is strongest. The correct configuration makes the model state what is missing rather than estimate it: an off-market teaser with no T-12 should produce a checklist of what to request, not a guessed NOI. Used that way, AI speeds up off-market work by generating sharp question lists for the broker or seller in seconds.
How do we start if nobody at the firm is technical?
Write the buy box document first; it requires zero technology. Then run the daily triage workflow manually in ChatGPT or Claude for two weeks and count what it catches and what it gets wrong. That experiment costs a subscription and a few hours, and it tells you more about your firm’s real automation case than any vendor demo. If the team needs structured prompting fundamentals first, that is exactly what LLM fluency training covers, and the training playbook lays out the 90-day path.
Next Steps
Run the two-week experiment: buy box on paper, daily triage in a business-tier LLM, a tracker of verdicts, and a 20-minute verification pass on anything that reaches a model. What that experiment produces, beyond screened deals, is evidence about where your firm’s specific bottleneck sits, and that evidence is what makes the buy-vs-build decision rational instead of vibes-based.
If you want a structured read on where AI would return the most in your deal workflow before committing to tools or builds, we run a free AI-readiness assessment for small CRE firms. Book a discovery call and bring one recent OM; the conversation is more useful with a real deal on the table.
Dirk Jan van Veen, PhD
