Why AI Deal Analyzers Get NH Multifamily Wrong — and What to Do Instead

Question

I run an AI deal analyzer alongside every property I underwrite. It's a useful tool. It catches things — expense omissions, overstated rent assumptions, overlooked capital expenses. I'd recommend any serious investor use one.

I would not recommend outsourcing your judgment to one.

I recently ran my AI on a Manchester, NH four-unit listed at $850K. The AI flagged it as a pass and suggested an offer of roughly $588,000. That's $147K per door — close to my own $150K/door Manchester benchmark.

The math is technically sound.

The suggestion is practically useless in the current Manchester market.

What the AI Got Right

The AI flagged several things my manual analysis also caught: the missing expense line items in the seller's pro forma (no CapEx, no property management, no vacancy reserve), overstated rents on two of the four units relative to HUD Fair Market Rent, and the 21-year-old heating system as a near-term capital expense.

On the expense side, the AI did its job. It read the pro forma, identified the gaps, and recalculated NOI with a more realistic expense stack.

At $588K — roughly $147K/door — the deal would pencil well for an investor. That per-door logic is correct.

What the AI Got Wrong

The AI doesn't know what's actually trading in the Manchester market. And that's where it falls apart.

In my experience, not a single four-family in Manchester has sold on the open market anywhere near $588K in the last two years. The buyers, the sellers, and the transaction history all point to a different range. An offer at that price wouldn't be rejected — it would be ignored.

This is the core limitation of algorithmic underwriting in local markets: it processes national data and statistical averages. It doesn't have visibility into what specific assets in specific cities are actually transacting at.

A Manchester operator knows this. A national algorithm doesn't.

The HUD FMR Gap

The AI also flagged that two units were rented above HUD Fair Market Rent. That sounds like a problem on paper — if rents are "above market," shouldn't you underwrite them conservatively?

Not necessarily.

HUD FMR is a useful baseline, not an absolute ceiling. Manchester's rental demand has been sustained over several years (U.S. Census Bureau, American Community Survey). Owner occupied multifamilies with quality renovations and good locations frequently command rents above FMR.

The AI doesn't know the difference between inflated rents that will fall on lease renewal and strong rents that reflect a genuinely tight market. An operator does.

How to Use AI Tools the Right Way

The right approach is to run the AI output alongside your own analysis — not instead of it. Specifically, I look for where the two diverge.

If the AI says offer $588K and I think $750K makes sense, I want to understand why. Sometimes the AI has caught something I missed. Often, the divergence is the AI applying national logic to a local market where the dynamics are different.

The most useful thing an AI output produces isn't the offer number. It's the expense gaps — what's missing from the seller's pro forma, what capital expenses are about to hit, where the income is overstated. That part of the analysis is genuinely useful.

The offer strategy, the market positioning, the buyer fit — those require local knowledge the algorithm doesn't have.

The Bottom Line

AI deal analyzers are useful expense checkers and pro forma auditors. They're poor market analysts in geographically specific markets like Manchester, NH, where the transaction history doesn't match national averages.

Use them to verify your expense assumptions. Use local market knowledge to evaluate everything else. The edge in a market like this isn't running better spreadsheets — it's knowing what the spreadsheet can't tell you.

Landon Reid · Answer

This is one of the best breakdowns I've seen on where AI actually fails in real estate.

Your point about the offer price being "technically correct but practically useless" is exactly the problem. The math works. The market doesn't care.

I'd add one more blind spot most AI deal analyzers share: they completely ignore what you're actually allowed to build on the parcel.

I've seen investors run beautiful financial models on a property, get excited about the numbers, then discover the zoning doesn't permit their planned use. Or a flood overlay kills the buildable area. Or setbacks reduce the footprint so much the project doesn't pencil anymore.

That's a different kind of "wrong" than what you're describing — but it's equally expensive. Investors spend $2,000-$3,500 on feasibility studies before they ever discover the lot won't support their plan.

I built a system that catches those issues in about 20 seconds — zoning, overlays, setbacks, height limits, permitted uses — so you can kill unbuildable deals before you even run the financial model.

Your approach of running AI output alongside manual analysis is exactly right. AI for data extraction, human judgment for market context.

Masoud Arouni · Answer

The assumptions going in are where it falls apart regardless of the tool. Vacancy, rent growth, exit cap — those have to come from someone who actually knows the market. The formula itself is just grounded domain math. Cap rate, IRR, debt coverage — deterministic, universal, same inputs same outputs. No model needed there.

The interpretation, the trend-reading, the 'does this submarket actually support these numbers' — that's still human work and will Always be.

Alex Wright · Answer

yup — don’t disagree at all

I’ve noticed the same thing. the numbers can be technically right, but still pretty disconnected from what would actually get a deal done in that market

I’m in a more rural area and prices are still relatively high, but that nuance just doesn’t really show up in a lot of these models

feels like they’re best at catching what’s missing or overly optimistic in a deal, not telling you what to pay for it

that gap between “what the model says” and “what actually trades” is usually where most of the value is

Alex Wright · Answer

Yeah that’s a good way to put it

feels like after you’ve looked at enough deals in one area, you can tell pretty quickly what’s realistic.

the models help clean things up, but that day-to-day exposure is what really anchors it

Ryan D. · Answer

@Andrew Bosco — this is one of the better breakdowns I've seen on where AI tools actually help vs where they fall apart.

The expense audit use case is real. The offer price output is useless. You nailed why — real estate is insanely local and no national model captures what's actually trading in a specific submarket. I invest in the PNW and see the same thing. The algorithm doesn't know that a particular corridor softened because two major employers froze hiring, or that a specific neighborhood commands premiums the data doesn't reflect yet.

This is exactly why I don't let AI touch the math on my deals. I use it heavily for the upstream work — reading OMs, extracting expense line items, flagging gaps in pro formas. All the stuff you described it doing well. But the underwriting engine itself has to be deterministic. Same inputs same outputs every time. And the assumptions going in have to come from someone who actually knows the market, not a model trained on national averages.

@Masoud Arouni vacancy point is right too. The stress test that matters isn't "does it work at average" — it's how long do I survive when it doesn't. I always underwrite with multiple exit strategies for exactly that reason.

Michael K Gallagher · Answer

Thanks for putting this together. I grew up in the 603, in enfield, always liked Manchester and MHT.  I can see why demand Would be sustained there, proximity to Boston and the northern burbs of mass, more affordable eastern city but still easy access to the seacoast and mountains.

Masoud Arouni · Answer

The expense audit point is spot on — that's where most pro formas fall apart regardless of market. What I'd add: the same AI limitation applies to vacancy assumptions. A national algorithm plugs in 5-8% (up to 1month) without knowing that Manchester's tight rental market or a military-adjacent market like Clarksville TN can sustain sub-3% vacancy for years — until it can't, and then it hits 30-50% (up to 6months) overnight.

The stress-test that actually matters isn't "does this deal work at average vacancy" — it's "how long until I'm underwater if vacancy doubles for 6 months?"

Romina Gonzalez · Answer

The expense gap flagging point is spot on! That's where most pro formas fall apart. Do you use anything for the renovation scope and rent lift side, or is that still mostly manual?

Andrew Bosco · Answer

it's a great market so long as boston stays strong

Andrew Bosco · Answer

it takes daily analysis to know your market

Why AI Deal Analyzers Get NH Multifamily Wrong — and What to Do Instead

Most Popular Reply