How AI Can Reduce Multifamily Underwriting Time

Most acquisitions teams I've talked to underwrite somewhere between 20 and 100 multifamily deals for every one they actually pursue. The conversion rate isn't the problem — it's the cost per look. A junior analyst spends 6–14 hours pulling together a first-pass underwrite that, in 90% of cases, ends in a one-word verdict: pass.

The cost isn't the analyst's time in isolation. It's the deals you don't look at because you're looking at this one. Every team has a backlog of broker emails, off-market intros, and PocketList listings that go cold because nobody got to them in time.

AI doesn't replace underwriting judgment. But it can compress the mechanical first 80% of the work — the part that's the same on every deal — from a day to about ten minutes.

Where the time actually goes

Before talking about AI, it helps to be honest about where a multifamily first-pass underwrite consumes hours. In our experience working with sponsors analyzing $5M–$50M deals, time breaks down roughly like this:

1. Document intake (45–90 min). Open the OM PDF, the rent roll spreadsheet, the T-12, the survey, the inspection report. Skim everything once. Note unit count, year built, location, asking price.
2. Rent roll normalization (60–120 min). Reconcile actual rents to market rents. Identify loss-to-lease. Spot vacant units, downed units, model units, employee units, non-revenue units. Aggregate to unit mix with weighted average rents.
3. Expense scrubbing (90–180 min). T-12 actual vs broker pro forma. Identify items the broker has stripped out (management fees, replacement reserves, payroll allocations). Adjust for missing line items. Benchmark against market norms.
4. Market research (60–180 min). Pull rent comps. Look at submarket trends — population, employment, new supply pipeline, recent transactions. Form a view on rent growth assumptions.
5. Financing scenarios (45–90 min). Agency vs bank vs bridge. Run DSCR. Stress at +100 bps. Model refinance year 3 or year 5.
6. Returns modeling (45–90 min). IRR, equity multiple, cash-on-cash by year. Run sensitivity tables. Model sale at exit cap rates.
7. Memo writing (60–120 min). Synthesize all of the above into a 2–3 page IC pre-read.

Total: 6 to 14 hours, depending on document quality and analyst experience.

What AI compresses well

Three of those steps — intake, normalization, and memo writing — are mostly mechanical translation. They consume the most hours per analyst and they have the lowest decision-density. They're also the steps where AI delivers the largest time savings with the least risk.

Document intake. An LLM with a strict JSON schema can read an OM PDF and return purchase price, NOI, occupancy, unit count, year built, average rent, expense ratio, and broker assumptions in under 60 seconds. We've found that on well-formatted OMs, extraction accuracy on the headline financials hits 95%+ when the prompt forces the model to return null rather than guess. The 5% that fail — usually atypical formatting or scanned PDFs — fail loudly, so the analyst knows to double-check.

Rent roll normalization. This is harder than it sounds because rent rolls aren't standardized — every property management software exports a slightly different layout. But the underlying task is grouping rows by unit type and computing weighted averages, which a small set of structured prompts handles well. The trick is asking the model to flag concentration of vacant units or unusually long lease terms rather than just averaging them away.

Memo writing. This is the highest-value compression. Once you have the structured underwriting outputs, an LLM can produce a coherent first-draft IC memo in the house style in about 90 seconds. The analyst then spends 20 minutes editing — adding judgment, adjusting tone, flagging the two or three things that warrant committee discussion — instead of three hours assembling.

What AI compresses poorly

The middle steps — expense scrubbing, market research, financing — are where judgment compounds. AI can accelerate these but it can't replace the analyst.

Expense scrubbing is a pattern-matching task that humans are still better at, because the patterns are local. Knowing that a particular submarket has property tax reassessment triggers at sale, or that a specific PMC inflates payroll on take-overs, is institutional knowledge. AI can flag that expenses look low relative to a generic benchmark, but the diagnosis is human.

Market research is where AI gives you a confident-sounding answer that's wrong 15% of the time. Public population and rent data is reasonable. Submarket-specific supply pipelines, lease-up velocity at new construction, and concessions burning off — these are facts that live in CoStar, REIS, and broker conversations, not in a model's training data. We treat AI-generated market summaries as a starting point for an analyst's research, never as the research itself.

Financing scenarios require a model of the capital stack the sponsor actually has access to. Generic AI output here is dangerous because it sounds plausible but doesn't reflect quoted terms.

The honest math

If a typical first-pass underwrite is 10 hours and AI compresses it to 4 hours of human time (with 6 minutes of model time), the analyst's deal capacity roughly doubles. That's the conservative version of the impact.

The non-conservative version: many of those 6 hours saved get reinvested in more deals. A team that was looking at 30 deals to close one starts looking at 60 — and the conversion rate doesn't move much, because the pass deals were always going to be passes. But the one deal you'd otherwise have missed becomes a deal you actually run.

The honest caveat: this only works if the team trusts the AI output. Trust is built by failure mode discipline — strict JSON contracts, explicit null for missing data, and never asking the model to fabricate numbers it doesn't have. The moment the model hallucinates a $1.4M NOI it pulled from a similar deal three years ago, the whole workflow loses credibility. Design the prompts so they fail loudly.

What this looks like in practice

A workflow we've seen work:

• Analyst drops the OM PDF into the system.
• Within 60 seconds, structured JSON appears with extracted headline financials and flags for missing fields.
• Analyst reviews the extraction (this is the trust-building step — 2 minutes for an experienced user).
• Analyst hits "run underwriting" — five specialist agents (financial, market, risk, operations, IC) run in parallel and return a first-draft IC view in about 90 seconds.
• Analyst spends 20–40 minutes refining the memo: correcting assumptions, adding local knowledge, sharpening the recommendation.
• IC pre-read goes out the same day.

Total cycle: 30–45 minutes of analyst time per deal, down from 6–14 hours. That doesn't include the deeper diligence that happens after IC says "go deeper" — but the screening funnel that decides which deals get that deeper look is now 10–20x faster.

The deals don't change. The number of deals you can actually evaluate does.