Stop putting every deal through a full underwriting model. Screen the entire inbox with AI to a ranked short list, then spend your underwriting hours only on the deals that survive the cut. A small firm’s constraint was never underwriting speed — it is underwriting discipline: deciding which deals earn a real model and which never had a chance. When a principal underwrites everything, the two deals worth winning get the same tired half-day as the thirty that were dead on arrival. The fix is not a faster spreadsheet. It is a two-stage funnel — cheap, fast, AI-assisted triage across everything, then careful human underwriting on the few that clear the bar. This piece is about how to draw that line and where the funnel breaks if you draw it wrong.
The bottleneck is discipline, not speed
Ask a principal at a ten-person shop why good deals slip, and the answer is rarely “our model is too slow.” It is “we were buried.” Forty broker blasts a week, each one a PDF that wants an hour, and the arithmetic never works. So the team does the human thing: it skims, it underwrites the ones with a familiar broker or a round cap rate, and it lets the rest sit until they are stale. The deal that would have cleared committee gets a five-minute skim; the deal that never had a chance gets a full afternoon because it happened to land on a slow Tuesday.
That is a discipline problem wearing a capacity costume. The scarce resource at a lean firm is not software budget or even hours in the abstract — it is the principal’s underwriting attention, the deep, judgment-heavy work of pricing a bid you would defend to a partner. Spend that attention evenly across every inbound and you have spread it thin enough to be worthless. The institutional shops the small-firm manifesto argues you can out-run do not underwrite everything either — they have an analyst bench to triage first. You do not. AI is how a lean team buys back that triage layer without the payroll.
Screening and underwriting are two different jobs
The category blurs the two on purpose, because “AI underwriting” sells better than “AI triage.” But they answer different questions.
Screening answers is this worth my time? It reads the offering memorandum, the T-12, the rent roll, and the broker one-pager, pulls the terms into a common shape, and ranks the deal against your buy box so you know what to open first. It is fast, cheap, and approximate, and it is fine that it is approximate, because nothing about a screen moves money.
Underwriting answers what is it worth, and what do I bid? It models the cash flows, stress-tests the assumptions, and produces a number a person will defend at committee. It is slow, expensive, and exact, and it has to be, because a bid is a commitment.
Confuse the two and you make the most expensive mistake in this category: you either buy a triage tool and trust its score as a verdict, or you drag a full modeling platform across your whole inbox and drown in precision you did not need. The two-stage funnel keeps each job in its lane. We covered the failure modes of over-automating the deep end in why most underwriting automations fail; the point here is the opposite end of the pipe — get the triage right so the deep end sees only deals that deserve it.
Screen wide and cheap: what AI does well up top
At the top of the funnel, current AI is genuinely good at the boring, high-volume reading a person hates. A well-built screen will take an inbound deal package and, in a minute, do three things a human takes an hour to do:
- Extract the terms into a common shape. Asking price, in-place NOI, unit or square-foot count, occupancy, year built, submarket — pulled from a messy PDF into the same fields every time, so deals become comparable instead of forty different formats.
- Rank against your buy box. Given a written statement of what you buy — target markets, asset classes, size, minimum yield, hold — the model orders the inbox so the deals worth opening rise to the top.
- Draft a first-pass summary. A three-line “here is what this is and why it ranked where it did,” so a principal decides what to open in ten seconds instead of ten minutes.
General assistants like ChatGPT, Claude, or Gemini already do most of this over a document you paste in, which is why the honest starting point is a prompt and a spreadsheet, not a purchase. A step up in volume and you want the screen wired to your inbox and pipeline so it runs without copy-paste — the shape of that build is laid out in our walk-through of a deal-screening framework that kills deals in minutes, not meetings. Either way, the job is the same: turn a flooded inbox into a ranked short list.
What AI does not earn at this stage is trust in its numbers. It will read a rent off the wrong column and present it as cleanly as a right one. That is a feature of the funnel, not a flaw — the screen’s job is to order the inbox, and a person verifies the numbers that matter only on the deals that survive.
Underwrite narrow and deep: what stays human
Below the cut line, the work changes character. Now you are pricing a bid, and the failure mode is no longer “we were too slow” but “we were confidently wrong.” This is where the model stays in the spreadsheet your firm already trusts and where a person owns every assumption.
That does not mean AI leaves the room. It can still pull comps, draft the market narrative, sanity-check an expense ratio against the submarket, and flag an assumption that looks aggressive. But the structure of the deal — the rent roll you have verified line by line, the exit cap you will argue for, the capital plan — is human work, because those are the numbers a partner will hold you to. The right division of labor is the one the deal-analysis playbook builds the whole workflow around: AI clears the reading and the busywork, human judgment owns the decision and the model behind the bid.
The tell that you have drawn the line correctly: a screening score never appears in an investment memo. It got the deal into the room. It has no business pricing it.
Where to put the cut line
The cut line is the threshold that decides which screened deals graduate to a full underwrite. Set it too loose and you are back to underwriting everything; set it too tight and you starve your own pipeline. Three inputs set it honestly:
- Your real underwriting capacity. Count how many deals your team can properly model in a week without cutting corners — usually a small number at a lean firm. The cut line should pass roughly that many deals, not more. The funnel exists to protect this capacity, so size the mouth of the deep end to what it can actually swallow.
- A written buy box. The screen can only rank against criteria you have written down. Most small firms carry the buy box in a principal’s head; get it onto one page — what you buy, what you never buy, the deal-breakers — and the ranking becomes something you can trust and tune.
- A deliberate false-negative check. The scariest deals are the good ones the screen ranks low on a misread. Once a week, spot-check a few deals just below the line. If good ones are getting cut, the buy box or the extraction is off, not the deal.
The line is a dial, not a law. Slow month, open it wider and let more deals through to a light underwrite. Flooded, tighten it and protect the hours. What matters is that the dial exists and someone owns it, instead of the team silently rationing attention by whoever emailed most recently.
Four ways the funnel breaks
A two-stage funnel is only as good as its discipline. Four failure modes turn a clean pipeline into a confident mess.
Silent number errors. The dangerous mistake is not the deal ranked wrong — that gets caught. It is the number extracted wrong and shown cleanly: an in-place rent read off the wrong column, an expense line quietly dropped, a T-12 total that does not tie. The guard is grounding — every extracted figure should link back to the page and line it came from, so verification is a one-click check on the survivors, not a re-keying of the whole inbox.
Buy-box drift. Your criteria change — a new market, a shifted yield target, an asset class you have soured on — and the screen keeps ranking against last quarter’s rules. A ranking against stale criteria is worse than no ranking, because it looks current. Revisit the buy box on a schedule, not when something breaks.
False negatives. The funnel’s quiet tax. A screen that surfaces a bad deal wastes ten minutes; a screen that buries a good one on a misread costs you the deal and you never know it happened. This is why the below-the-line spot-check is not optional — it is the only way to see the errors that do not announce themselves.
Score creep. The slow failure. A triage score is useful precisely because it is disposable, but the moment it works well enough, people start treating “it scored an 82” as a decision instead of an invitation. The number that was meant to order the inbox quietly becomes the verdict, and human underwriting erodes without anyone deciding to erode it. Keep the score out of the memo, on purpose, forever.
What it costs and what it returns
The spend has two shapes. Getting fluent and running a prompt-plus-spreadsheet screen costs the tools you already pay for and a few days of learning — effectively the price of a short fluency workshop, which for a small team runs in the low thousands (market rates for hands-on LLM training land roughly in the $2,000–$15,000 range depending on scope and headcount). Wiring a custom screening workflow into your inbox and pipeline is a project, generally in the tens of thousands to low six figures — the full cost model, including the maintenance tail most firms forget to price, is broken down in how much custom underwriting automation costs.
The return is not “documents processed.” It is recovered underwriting attention, redirected to deals that can actually clear. If a screen turns a week of triage into an afternoon, the payoff is the three deals your principal now underwrites properly instead of skimming — and the one of those three that becomes a bid you would not have written at all. Measure the funnel by that, not by pages read. A lean team wins by pointing its scarcest judgment at the deals that deserve it, and the two-stage funnel is the mechanism that makes that happen every week instead of by luck.
How to start without buying anything
You do not need a platform to run this funnel next Monday. Start with the parts you own:
- Write the buy box. One page: what you buy, what you never buy, the deal-breakers. Nothing downstream works without it.
- Build one screening prompt. Feed a recent OM and your buy box to ChatGPT, Claude, or Gemini and ask for the terms in a fixed shape plus a rank and a two-line rationale. Refine it against ten real deals, including your two ugliest.
- Set a provisional cut line. Pick the number of deals a week your team can truly underwrite, and let only that many through.
- Verify on the survivors only. Check the numbers that matter on the short list, never on the whole inbox.
- Run the false-negative check. Once a week, look just below the line for good deals that got cut.
Run that for a month and you will know two things the vendor demos cannot tell you: whether your process needs a purchased tool at all, and exactly what that tool would have to do. Many small firms find the prompt-and-spreadsheet version covers their volume. If yours does not, you are now the sharpest possible buyer, because you know precisely what you are replacing — the discipline you should bring to any purchase, spelled out in our rules for buying deal-screening AI.
Frequently asked questions
What does “screen with AI, underwrite the survivors” actually mean?
It means running your deal flow as a two-stage funnel. In the first stage, AI reads every inbound deal — offering memorandums, T-12s, rent rolls, broker one-pagers — extracts the key terms into a common shape, and ranks each deal against your written buy box, producing a short list. In the second stage, a human underwrites only the deals that survive the ranking: modeling the cash flows, verifying the numbers, and pricing a bid. The point is to stop spreading scarce underwriting attention evenly across every deal and concentrate it on the few that can actually clear.
Why not just underwrite every deal properly?
Because a lean team cannot, and pretending otherwise is why good deals slip. A ten-person shop might see forty deals a week and have the capacity to properly model a handful. Underwriting everything means each deal gets a thin, rushed pass, so the two deals worth winning get the same treatment as the thirty that were never going to clear. Screening first lets you spend full underwriting effort where it changes the outcome, instead of rationing it by whoever emailed most recently.
Isn’t AI screening just AI underwriting by another name?
No, and conflating them is the most expensive mistake in this category. Screening ranks deals so you know what to open; it is fast, cheap, and approximate, and nothing about a screen moves money. Underwriting models cash flows to price a bid; it is slow, exact, and human-owned. Many vendors market triage tools as “AI underwriting,” which leads firms to trust a screening score as a verdict. Keep the jobs separate: the score gets a deal into the room, a human decides what it is worth.
Can I run this with ChatGPT and a spreadsheet, or do I need to buy a tool?
You can start with tools you already pay for. A well-built prompt over ChatGPT, Claude, or Gemini, plus your buy box and a spreadsheet, will extract terms and rank many deals at low volume. You need a purchased tool only when copy-paste becomes the bottleneck or when you want the screen wired directly to your inbox and pipeline. Run the manual version for a month first — it tells you whether you need a platform at all and makes you a sharper buyer if you do.
How accurate is AI at reading a T-12 or an offering memorandum?
Good on clean documents, much weaker on the scanned, inconsistent, typo-ridden files that fill a real inbox. The accuracy that matters is on your worst documents, not the vendor’s demo deal. The failure mode to watch is not a visible error — those get caught — but a clean-looking number pulled from the wrong column. That is why screening stays approximate by design, a person verifies the numbers only on the survivors, and every extracted figure should link back to its source page for a one-click check.
Where should I set the cut line between screening and underwriting?
Set it to your real underwriting capacity — the number of deals your team can properly model in a week without cutting corners. The cut line should pass roughly that many deals, not more, so the deep end never takes in more than it can handle well. Treat it as a dial: open it wider in a slow month, tighten it when you are flooded. And check just below the line weekly for good deals that got cut, because a threshold that is too tight starves your pipeline as surely as one that is too loose overwhelms it.
What is the biggest risk with an AI screening funnel?
False negatives and score creep. A false negative is a good deal the screen ranks low on a misread — it costs you the deal and you never know it happened, which is why a weekly spot-check below the cut line is essential. Score creep is the slow failure where a triage score, useful precisely because it is disposable, gets treated as a decision once it works well enough; human underwriting erodes without anyone choosing to erode it. Keep the score out of the investment memo, on purpose, to foreclose both.
Is it safe to put confidential deal packages into an AI tool?
Only if the terms say so. Deal packages are confidential and often under NDA, so before uploading anything, settle three questions in writing: where the data is stored, whether your uploads train shared models, and how you delete and export your history. A small firm without an IT department has the contract as its main safeguard, so treat vagueness as a reason to walk. For the manual approach, use business-tier accounts whose terms exclude your inputs from training, and confirm that before you paste a live deal.
How much does it cost to run this way?
The manual version costs the tools you already pay for plus a few days of learning — roughly the price of a short LLM fluency workshop, which for a small team runs in the low thousands. A custom screening workflow wired to your systems is a project, generally in the tens of thousands to low six figures depending on scope, plus an ongoing maintenance cost in someone’s time. Price the full first year with your own hours included, and measure the return in recovered underwriting attention, not documents processed.
Where to start
The first question is not which screening tool to buy. It is whether your buy box is on paper, whether your team can already run a first-pass screen by hand, and how many deals a week your firm can genuinely underwrite well. Answer those and the funnel mostly designs itself — and the tool decision, if you need a tool at all, gets easy. A free AI-readiness assessment gives you that read: a short working session that looks at your deal flow, your asset mix, and where your underwriting hours actually go, then returns an honest recommendation on whether a fluency-first workflow with the tools you already own, a custom build, or a purchased platform is the right next step. Book a free AI-readiness assessment before you spend another quarter underwriting deals that were never going to clear.
Arthur Wandzel
