The problem is not that assumptions exist. The problem is when assumptions are written with the same confidence as a sourced fact. A spreadsheet can make that mistake look professional. A single stabilized-rent number, expense ratio, renovation budget, interest rate, or exit value may travel through a model untouched until it produces a return that feels precise. The decimals do not make the underlying judgment more certain.
As an operator and investor, I want the model to help me ask a better question: what has to be true for this deal to work, and what evidence supports each part? That is different from asking a model to persuade me that a deal works. The first question exposes risk early. The second question can turn underwriting into confirmation bias.
This is why I separate the real estate decision from the document-processing work around it. Systems can make the file easier to review, but they cannot accept the risk for me. My AI underwriting support framework follows the same boundary: use tools to prepare information and surface gaps, then keep value, financing, terms, and capital allocation with the accountable owner.
The Four Buckets Every Number Needs
The simplest repair is to classify every material input. I use four buckets because they force a useful distinction without making the worksheet hard to maintain.
Verified source input. This is a figure tied to a document, an observed condition, or another source that has been reviewed for the decision. It still may need interpretation, but the source is visible. Examples include a current lease, bank statement, tax bill, executed contract, utility invoice, lender term sheet, or an inspection observation.
Reported input. This is information supplied by a broker, seller, manager, listing, or conversation. It belongs in the file because it may point toward a real answer. It is not the same as verification. Mark the source and date, then decide what document or observation would confirm it.
Operating estimate. This is a reasoned forecast used to test a future condition, such as a reserve, renovation scope, market rent, marketing period, or recurring expense. State the method, comparable context, or rule behind it. If a number exists only because it makes the return work, it is not an estimate yet. It is a wish.
Open question. This is a material item that does not yet have a trustworthy answer. It should not disappear into a footnote. An open question needs an owner, a requested source, a due date, and a decision gate. The honest answer may be that the deal cannot advance until the question is resolved.
| Input type | What belongs beside it | Example | Decision treatment |
|---|
| Verified source input | Document link, date, reviewer | Current rent roll reconciled to leases | Use in base case, note any limitations |
| Reported input | Who reported it and confirmation request | Broker says repairs are complete | Do not treat as closed until evidence arrives |
| Operating estimate | Method, comparable basis, range | Turn-cost estimate from scoped bids | Test in base and downside cases |
| Open question | Owner, due date, decision gate | Unknown insurance renewal cost | Hold, reserve conservatively, or decline |
The label matters because it changes the work. A verified number may be ready for an investment committee discussion. A reported number creates a diligence request. An estimate deserves sensitivity testing. An open question may prevent an offer, change a contingency, or tell you to walk away.
Build an Assumption Register Before the Model Gets Complicated
I prefer a short assumption register alongside the operating model. It can be a tab in a spreadsheet, a deal memo, or a database record. The format is less important than the discipline. Build it before the model becomes difficult to audit.
For each material number, record the input name, current value, unit, classification, source link or source note, source date, confidence level, impact if wrong, next verification step, owner, and decision deadline. Add one plain-language sentence about why the number is reasonable. This makes it much easier for another operator to review the deal without reverse-engineering a series of hidden cells.
Start with the inputs most likely to change the decision: income, occupancy, expenses, capital needs, taxes, insurance, debt terms, timing, and exit assumptions. The correct list varies by asset and strategy. The point is not to create a generic hundred-line checklist. The point is to expose the small number of assumptions that carry the most downside.
For example, projected rent growth should state whether it comes from executed leases, current in-place rent, comparable units, a market report, a broker opinion, or a planning target. Those are different inputs with different confidence. A good register does not flatten them. It tells the reader whether the model is describing today, extrapolating from observed evidence, or waiting on something material.
That same approach supports a cleaner pipeline. In my weekly real estate deal-review system, a deal should have a current state, a specific next action, and an owner. The assumption register gives that meeting a better agenda. Instead of saying a deal is still being underwritten, the team can say exactly which three inputs are unresolved and what would change the decision.
The Decision Table for Source Conflicts
Conflicting documents are not unusual. What matters is how the conflict is handled. The worst response is to select the most favorable figure and keep moving. The better response is to preserve the difference, name the source hierarchy, and decide whether the gap changes the economics or the deal structure.
| What the file shows | First response | What not to do | Possible decision gate |
|---|
| Seller income statement differs from bank deposits | Reconcile timing, scope, and source records | Average the two figures without explanation | Do not use the income claim until reconciliation |
| Rent roll differs from lease terms | Review executed leases and payment evidence | Assume the higher rent is collectible | Underwrite current verified rent or request cure |
| Expense estimate has no invoices or contract | Request source documents and test a range | Call the estimate fixed because it is in a package | Increase reserve or hold for confirmation |
| Inspection suggests work outside the budget | Scope the issue and price alternatives | Hide the concern in a general contingency line | Reprice, negotiate, or decline |
| Financing terms changed from initial quote | Re-run debt service and closing-cost assumptions | Keep the original return output | Revise offer, terms, or capital plan |
Source conflicts do not automatically kill a deal. They do create work. Sometimes the gap is a timing difference, a definition mismatch, or an outdated document. Sometimes it reveals a quality-of-earnings problem, a deferred maintenance issue, or a seller representation that needs more scrutiny. The record should make that distinction visible.
When a figure changes, keep the old number and explain why it changed. That preserves learning. It also lets the buyer see whether the underwriting is becoming more conservative as evidence improves, rather than quietly changing inputs until the output looks acceptable.
Run Downside Cases That Change a Real Decision
A downside case is not a ritual where every assumption gets reduced by the same percentage. It is a test of the conditions most likely to matter for this deal. Choose a few variables that could change cash flow, required capital, financing, timing, or exit value. Then decide in advance what would change if that case occurred.
For a rental asset, that might mean testing slower lease-up, lower effective rent, higher repairs, insurance movement, tax reassessment, or refinance constraints. For a value-add project, the material variables may be scope, permits, contractor timing, contingency, and the carrying cost while work is complete. For any strategy, model the cost of being wrong before deciding that a favorable scenario is the base case.
Each scenario should have a rule attached. If the downside case pushes the required reserve beyond available capital, the next action might be to renegotiate, raise the reserve, or pass. If it turns a projected return from acceptable to marginal, the buyer may request new evidence or reduce the offer. If the deal remains workable, the scenario becomes a reasoned risk statement, not an excuse to ignore risk.
That is how I think about operational leverage in investing. Good systems do not remove judgment. They make the relevant judgment easier to find before money and time are committed. My AI for real estate investors page explains where a support layer can help with this preparation without pretending that a tool can own the investment thesis.
Use AI for Preparation, Not False Certainty
AI can be useful around underwriting when it reduces repetitive reading and improves the questions that reach the operator. It can extract tables from documents, create a document index, compare two versions of a rent roll, summarize a lease clause for review, flag terms that appear inconsistent, draft a diligence request list, and turn call notes into an assumption register draft.
Those are preparation tasks. The output must still point back to the source. A model can misread a scan, omit a qualifier, or state an unsupported conclusion with confidence. That is why I do not use AI output as an unsourced fact in a deal model. I use it to find where I need to look next.
The boundary becomes especially important when a result affects price, leverage, legal rights, taxes, insurance, or a representation to another party. AI can make the first pass faster. The accountable owner and qualified specialists still review the evidence and make the decision. This is not a limitation to work around. It is a design rule that keeps speed from becoming untracked risk.
The Underwriting Handoff That Protects the Next Owner
A deal file becomes valuable when someone else can understand it. That may be a business partner, lender, attorney, property manager, investor, or future operating owner. A handoff should include the current model, the assumption register, source documents, unresolved questions, the latest decision, and the next action with an owner and due date.
Do not reduce the handoff to a folder link. Explain the decision logic in plain language: which facts are verified, which assumptions carry the most sensitivity, what has changed since the last review, and what would cause the buyer to renegotiate or stop. The next reviewer should not have to guess whether a number is fact, estimate, or hope.
This is also where investor discipline becomes a repeatable operating system. A good file makes it possible to compare future deals, audit why a prior decision was made, and learn from what actually happened after closing. It turns underwriting from a one-time performance into a body of evidence the next decision can use.
Frequently Asked Questions
What are the most important real estate underwriting assumptions?
The most important assumptions are the ones that can materially change cash flow, value, financing, timing, or required capital. Their source, date, confidence level, and downside case should be visible in the model.
How do I know whether an underwriting input is verified?
Treat an input as verified only when it is tied to a credible source that has been reviewed for the decision at hand. A seller statement, old listing, or copied figure can be useful context without being verified proof.
Can AI underwrite a real estate deal for me?
AI can organize documents, flag inconsistencies, and prepare questions. It cannot replace investor judgment about sources, market fit, financing, risk tolerance, negotiation, or an investment decision.
What should I do when the seller numbers and source documents conflict?
Log the difference, retain both sources, request clarification, and avoid building the decision around the more favorable number. A model should show the unresolved issue instead of hiding it.
How many downside cases should I run?
Use a small set of cases tied to the assumptions that matter most, such as revenue, expenses, occupancy, financing, timing, or capital needs. More scenarios are not helpful if no decision rule changes with them.
Build the Decision Record Before You Add More Automation
The goal of underwriting is not a prettier model. It is a decision that stays understandable when the initial excitement is gone. Start by making sources, assumptions, and open questions visible. Then use AI to reduce the preparation work around that standard.
I help owner-led businesses build cleaner systems for research, handoffs, documentation, and execution. If your acquisition process needs a clearer operating layer before you add more tools, start there.
Current Search Intent Check
Recent Search Console data shows people arriving through "deal flow automation commercial real estate". That changes the bar for this post: it needs to answer the operator question directly, name the workflow being improved, and give the reader a practical decision rule instead of another broad AI opinion.
Recent Search Console data shows people arriving through "automated real estate investing systems". That changes the bar for this post: it needs to answer the operator question directly, name the workflow being improved, and give the reader a practical decision rule instead of another broad AI opinion.