An underwriting model is the spreadsheet that turns a property into a decision. It takes what a building earns today, projects that income across the years you plan to hold it, subtracts the costs and the loan, adds what you expect to sell it for, and boils the whole thing down to a handful of return numbers you can compare against every other deal on your desk. Strip away the jargon and a model does one job: it answers “if I pay this price, what do I make?” This piece walks the model part by part — every major line, what it does, and, for a firm running on Excel with no analyst bench, which parts AI can safely assemble and which parts a human must own.
What an underwriting model actually is
Underwriting, in commercial real estate, means deciding what a deal is worth and what could go wrong with it before you commit money. The underwriting model is the tool that carries out that decision in numbers. At its simplest it is a projection of cash flows: a row for each year of your intended hold, income coming in at the top, costs and debt payments coming out below, and a sale at the end. The bottom of the model converts those cash flows into return metrics — how fast your money comes back and how much of it.
Most small firms build this in Excel. Larger shops may use Argus for complex, multi-tenant properties, but a spreadsheet handles the majority of deals a 4-to-20-person firm sees. The software is not the point. The anatomy is the same whether the model lives in Excel, Argus, or a purpose-built tool: the same parts, in the same order, feeding the same outputs.
The critical thing to understand before we open it up is that a model is not a fact machine. It does not tell you what a property is worth. It tells you what a property is worth if your assumptions hold. Change the rent-growth number or the sale-year cap rate and the same building produces a very different answer. A model is a structured argument, and its power comes entirely from the honesty of the inputs you feed it.
The two layers hiding inside every model
Every line in an underwriting model belongs to one of two layers, and telling them apart is the single most useful lens for a small firm — because it decides where AI belongs.
The mechanical layer is arithmetic and transcription. It only has to be right. Reading in-place rents off a rent roll, summing operating expenses from a trailing statement, calculating a loan payment, compounding a growth rate across ten years — none of this requires judgment. It requires accuracy and time, and it is where a lean team quietly loses hours on every deal.
The assumption layer is judgment. It can only be defended, never proven. What will market rents do over your hold? What cap rate will a buyer pay when you sell in year five? How long will that vacant suite sit before it leases? These are opinions about an uncertain future, and the quality of a model is almost entirely the quality of these calls.
Keep this split in mind as we walk the parts. The mechanical layer is where AI genuinely earns its keep. The assumption layer is where it is dangerous. Most of the confusion about “AI underwriting” comes from failing to draw the line between them.
The revenue build: the top of the model
The model starts with income, and income starts with the rent roll — the document at the center of every deal. The rent roll lists every tenant, their space, their current rent, and their lease expiration. That is the raw material for the revenue build: the projection of gross rental income across your hold.
Building revenue means taking today’s in-place rents and rolling them forward. Contractual bumps written into existing leases are close to fact — you can read them off the lease. Everything past a lease’s expiration is judgment: when the space rolls, you assume a renewal or a new tenant, a market rent, some downtime while it sits vacant, and a leasing cost to fill it. On top of contractual rent, the revenue build adds other income — parking, storage, expense reimbursements from tenants — and subtracts a vacancy and credit-loss allowance, because no building stays fully leased and fully paid forever.
The result is effective gross income: what the property realistically collects in a year. Notice the layers already tangled together here. The current rents are mechanical. The rollover assumptions, the market rents, the vacancy factor — those are the assumption layer, and they are where two honest underwriters looking at the same building will disagree.
Expenses and NOI: the operating result
Below income sits the operating expense build: property taxes, insurance, utilities, repairs and maintenance, management fees, and the general cost of running the building. For an in-place year, these come off the trailing twelve months of operating statements — mechanical work, though it demands care, because expenses are the easiest thing for a seller to understate. A management fee left off because the owner self-manages, a tax line that ignores the reassessment a sale will trigger: both quietly inflate the result.
Effective gross income minus operating expenses gives you net operating income — NOI, the most important single number in the model. NOI is the property’s unlevered operating profit, before any loan payment and before capital spending. It is the number the whole valuation hangs on, because NOI divided by a cap rate is how the market prices the asset. That relationship — and how a quoted one gets manipulated — is the subject of the guide to what a cap rate is and how AI helps sanity-check it.
One trap worth naming: NOI is not cash flow. NOI is the operating result. Cash flow is what is left after you also subtract capital costs and the mortgage, which is where we go next.
Capital and debt: the money layer
Two things stand between NOI and the cash that actually reaches your pocket.
The first is capital expenditure — the big, non-operating costs of owning the building. Tenant improvements to build out space for a new lease, leasing commissions to the brokers who bring tenants, a reserve for the roof and the HVAC and the parking lot that wear out over the hold. These are lumpy and easy to underweight, and a model that ignores them overstates what the deal returns. A realistic capital plan is often the difference between a deal that pencils and one that only appears to.
The second is debt. Most deals use a loan, and the financing layer models it: the loan amount, the interest rate, the amortization, and the annual debt service that comes out of cash flow. Debt is where an unlevered deal becomes a levered return. A property might produce a modest unlevered yield, but borrowing at a rate below that yield magnifies the return on the cash you actually invested — and magnifies the downside just as hard if the deal underperforms. The debt terms are mostly mechanical once you know them; the question of how much borrowing is prudent is judgment.
NOI, minus capital costs, minus debt service, gives you levered cash flow — the real annual cash the deal throws off to your equity. That is the number that matters to your return.
The exit: where the deal is really decided
Here is the part that surprises people new to modeling: for a typical hold, most of the return does not come from the annual cash flow. It comes from the sale at the end. The exit is modeled by taking the NOI in your sale year and dividing it by an exit cap rate — the yield you assume a future buyer will accept — to produce a projected sale price. Subtract selling costs and pay off the remaining loan balance, and what is left is the equity you walk away with.
The exit cap rate is the most consequential assumption in the entire model, and the least knowable. It is a guess about market conditions years in the future, and a swing of half a percentage point can move the whole return by more than any operating decision you make. This is precisely the kind of number a spreadsheet will happily accept and an AI tool will happily invent — and precisely the number a human must set from a defensible view of where the market is heading. Everything mechanical in the model funnels into this single act of judgment.
The outputs: IRR, equity multiple, cash-on-cash
The bottom of the model translates the projected cash flows — the annual levered cash flow plus the lump of equity from the sale — into three return metrics that let you compare this deal against any other.
Internal rate of return (IRR) is the annualized return on your invested equity, accounting for the timing of every cash flow. It rewards money that comes back sooner and penalizes money that comes back later, which makes it the metric most investors lead with.
Equity multiple is the simplest: total cash returned divided by cash invested. A 2.0x multiple means you doubled your money over the hold. It ignores timing entirely, which is why it is read alongside IRR, not instead of it.
Cash-on-cash return is annual pre-tax cash flow divided by the cash invested — a year-by-year read on income, useful for a buyer who cares about ongoing yield rather than just the exit.
None of these three is sufficient alone. A deal with a strong IRR but a thin equity multiple is telling you the return depends on a fast exit; a strong multiple with a weak IRR is telling you the money is tied up a long time. Reading them together is the judgment the model exists to inform.
The assumption layer behind everything
Step back and the model resolves into a small set of assumptions doing almost all the work: market rent growth, the vacancy factor, rollover timing and downtime, the capital reserve, the amount of debt, and — above all — the exit cap rate. Everything else is arithmetic wrapped around these calls.
This is why two underwriters produce different values for the same building and why “the model says it’s worth X” is never a complete sentence. The right response to any model is to ask what it assumes, then stress-test those assumptions — running the return at a higher exit cap, slower rent growth, and a longer lease-up to see how fragile the answer is. A deal that only works under optimistic assumptions is a deal disguised as an opportunity. The maturity of the tools now available to run that analysis, and where they still fall short, is covered in the overview of the state of AI in CRE investment analysis.
Where AI helps build the model
Map AI against the anatomy and its role is clear: it belongs on the mechanical layer, and it is genuinely valuable there. The input side of a model is hours of document transcription, and that is work a language model does in seconds.
Point a general assistant — ChatGPT, Claude, or Microsoft Copilot — or a purpose-built deal tool at a rent roll, and it reads tenants, in-place rents, square footage, and lease expirations into a clean table you can drop into the revenue build. Feed it the trailing operating statements and it structures the expense lines, flags the ones a seller tends to omit, and drafts a starting expense schedule. Hand it a lease and it pulls the rent steps and renewal options that drive rollover. It can lay out the skeleton of a model, wire up the arithmetic, and check that NOI, debt service, and cash flow tie out — the assembly a lean firm keeps deprioritizing when the pipeline is full.
Used this way, AI collapses the setup time that keeps a small team from modeling every deal it sees. The firm underwrites more opportunities at the same headcount, which is the whole argument of the small-firm playbook for out-operating institutional competitors. When the same model gets rebuilt on every deal, that repeatable assembly is also exactly what a custom underwriting automation is built to handle.
Where AI must not go
The line is the two-layer split, and holding it is what makes the tool safe: AI assembles the mechanical layer, and a human owns every assumption and the final number.
Ask a language model for a market rent-growth rate, an exit cap rate, or a return target and it will produce a confident, specific figure with nothing behind it. It has no genuine read on your submarket’s next five years and no memory of what traded down the street. In a model, a baseless assumption does not look baseless — it looks exactly like a defensible one, and it flows straight through to a return number that moves a bid. The model can hold the assumptions; it cannot form them.
The same caution applies to the extractions. AI can read a figure off the wrong row of a messy scanned rent roll, and the wrong number looks just as clean in the output table as the right one. Every input that will move money gets verified against its source before you trust it. And one guardrail specific to a firm with no IT department: a rent roll and a trailing operating statement are confidential financial data. Before you upload a deal’s documents to any AI tool, get three answers in writing — where the data is stored, whether your inputs train shared models, and how you delete your history. Major providers state that business-tier and API data is not used for training by default, but the contract is your safeguard, not the marketing copy. For how model-building sits alongside screening, comps, and market work in one workflow, the deal-analysis playbook for lean teams sequences the whole stack.
FAQ
What is an underwriting model in commercial real estate?
An underwriting model is a spreadsheet that projects a property’s cash flows over your intended hold and converts them into return metrics you can compare across deals. It builds revenue from the rent roll, subtracts operating expenses to reach net operating income, then subtracts capital costs and debt service, adds a projected sale at the end, and outputs figures like IRR and equity multiple. It does not tell you what a property is worth outright — it tells you what it is worth if your assumptions about rent, vacancy, and the eventual sale hold true.
What are the main parts of a CRE underwriting model?
The main parts, top to bottom, are the revenue build (rents from the rent roll plus other income, minus vacancy), the operating expense build, net operating income (NOI), capital costs and reserves, the debt or financing layer, the exit (a projected sale using an exit cap rate), and the return outputs (IRR, equity multiple, cash-on-cash). Behind all of them sits an assumption layer — rent growth, vacancy, the exit cap rate — that drives most of the answer.
What is the difference between NOI and cash flow in a model?
NOI, net operating income, is the property’s operating profit: income after operating expenses but before any loan payment or capital spending. Cash flow is what remains after you also subtract capital costs — tenant improvements, leasing commissions, reserves — and the mortgage payment. NOI describes the building’s operations and is what the market prices with a cap rate; cash flow describes what actually reaches your equity. A model that treats NOI as the money you pocket overstates the return.
What return metrics does an underwriting model produce?
The three standard outputs are IRR, equity multiple, and cash-on-cash return. IRR is the annualized return on invested equity, weighted for the timing of cash flows. Equity multiple is total cash returned divided by cash invested, ignoring timing. Cash-on-cash is annual pre-tax cash flow divided by cash invested, a year-by-year yield read. None is sufficient alone — read together, they show whether a deal’s return depends on ongoing income, a fast exit, or a large but slow payoff.
What assumptions matter most in an underwriting model?
The exit cap rate matters most — the yield you assume a future buyer will accept, which sets the sale price and can swing the whole return by more than any operating decision. After it come market rent growth, the vacancy and credit-loss factor, rollover timing and lease-up downtime, the capital reserve, and how much debt the deal carries. These are the judgment calls that separate two honest underwriters valuing the same building. The rest of the model is arithmetic wrapped around them.
Can AI build an underwriting model for me?
AI can build the mechanical parts quickly. A general assistant or a deal tool can read a rent roll and operating statements, structure the revenue and expense builds, lay out the model’s arithmetic, and check that NOI, debt service, and cash flow tie out. What it cannot do is supply the assumptions that drive the answer — rent growth, exit cap rate, return targets — or be trusted without verification of the figures it extracted. Treat it as a fast analyst for the transcription and structure, with a human owning every judgment call.
Should AI set the assumptions in my model?
No. Setting a market rent-growth rate, an exit cap rate, or a return target is judgment about a specific submarket at a specific time, and a model has no genuine basis for it. Asked for one, a language model produces a confident number with nothing behind it, and a baseless assumption flows straight through to a return that moves a bid. Let AI assemble the inputs and the arithmetic; set every assumption yourself from a defensible view of the market, and stress-test the model against less optimistic ones.
Is it safe to upload a rent roll and operating statement to an AI tool?
Only under the right terms. A rent roll and a trailing operating statement are confidential financial data, so before uploading them to any AI tool, get three answers in writing: where the data is stored, whether your inputs train shared models, and how you delete and export your history. Major providers state that business-tier and API data is not used for training by default, but for a firm with no IT department the contract is the safeguard, not the marketing copy. A vague answer is a reason to keep proprietary deal data out of that tool.
Do I need Argus, or is Excel enough for a small firm?
For most deals a small firm underwrites, Excel is enough — the anatomy of the model is identical regardless of the software. Argus earns its cost on complex, multi-tenant properties with intricate lease structures where its cash-flow engine saves real time. A 4-to-20-person firm doing straightforward acquisitions rarely needs it to start. The bigger gain is usually not switching platforms but speeding up the mechanical assembly of the model you already build in Excel.
How is an underwriting model different from an appraisal?
An underwriting model is a buyer’s or lender’s forward-looking analysis of what a deal returns under their assumptions and strategy; it is built to make a decision. An appraisal is an independent, standardized estimate of market value at a point in time, prepared by a licensed appraiser to a defined methodology, often for a lender’s records. The two overlap on inputs like NOI and cap rate, but the model reflects your specific plan and return targets, while the appraisal aims for a defensible, neutral value.
Key takeaways
- An underwriting model projects a property’s cash flows over your hold and converts them into return metrics — it tells you what a deal is worth if your assumptions hold, not as a fact.
- The core anatomy runs top to bottom: revenue build, operating expenses, NOI, capital costs, debt, the exit sale, and the return outputs of IRR, equity multiple, and cash-on-cash.
- Every line belongs to one of two layers — a mechanical layer that only has to be accurate, and an assumption layer that can only be defended. The exit cap rate is the most consequential assumption in the whole model.
- AI belongs on the mechanical layer: reading rent rolls and operating statements into the revenue and expense builds, structuring the arithmetic, and checking the math ties out — the setup work a lean firm keeps skipping.
- AI must never supply the assumptions or go unverified on extractions. The model can hold the judgment; it cannot form it. Keep every assumption and the final number with a human.
Want to know where your firm’s hours actually go in deal analysis, and which parts AI can safely take off your plate? A short assessment maps that against your deal flow, your documents, and where AI belongs without introducing risk. Book your free AI-readiness assessment →
Dirk Jan van Veen, PhD
