Investment analysis
Underwriting software that shows its work
An investor deciding on a property needs a return number and the reasoning behind it. A model that guesses at a cap rate will produce one that looks right and is not.
What decides an investment analysis build
4 things that decide this
- 01The math has to be code, not a model call. ROI, cash flow and cap rate are deterministic formulas, and a language model asked to compute one will hallucinate a plausible-looking answer.
- 02Comps are a judgment call wearing a number. Two analysts picking different comparable sales get two different valuations from the same address.
- 03An AVM gives a point estimate with no confidence range, and that is the wrong output for a purchase decision. Show the range and the inputs behind it.
- 04The tool has to serve the decision, not just the deal. Investors compare scenarios before they commit, and the platform needs a place to hold more than one.
Spreadsheets do not fail at math. They fail at trust
Most investors already have a spreadsheet that computes ROI. The problem is not the formula. It lives on one laptop and breaks when a new expense line gets added. Nobody else has a reason to trust the number.
A platform fixes that by making the inputs structured and the output repeatable. Purchase price, renovation cost, projected rent, taxes and operating expenses go in as typed fields. Cash flow, total cost and return come out the same way every time, for every property, for every analyst who logs in.
- 01Structured inputs replace a spreadsheet's open cells, so a formula error cannot enter through a typo.
- 02The same engine prices every property, so two analysts get the same answer from the same numbers.
- 03Scenario comparison lets an investor hold two or three versions of a deal side by side.
- 04A change to one assumption recalculates the return instantly, instead of waiting on a rebuilt spreadsheet.
Sector context
The numbers that shape this build
489
functioning MLS systems in the US, licensed one at a time (RESO)
90%+
of MLSs have RESO-certified Web API services (RESO)
1
real estate investment analysis platform we built and run: SmartREI
22
published case studies across the firm
Four places an underwriting build goes wrong
These show up whether the property is a single rental or a fifty-unit portfolio.
The model computes the return
Ask a language model for a cap rate and it will produce one, formatted like an answer. It is not doing arithmetic. It is predicting the next plausible token. The return has to come from code that runs the formula, every time, the same way.
A point estimate hides the uncertainty
An automated valuation model gives one number with no range around it. A property worth 380,000 to 420,000 shown as a single figure looks more certain than the data supports, and an investor acts on that false confidence.
Comps are chosen, not computed
Picking which recent sales count as comparable is a judgment call. Distance, condition and timing all move the answer. A platform that hides which comps it used, and why, is asking to be trusted blind.
Renovation and expense assumptions decay
A rehab budget entered at analysis time drifts from the real cost by the time work starts. A platform that treats that number as fixed forever quietly overstates every return it produces after month one.
- InputsPrice, rent, tax, expenses
- CompsNamed, not hidden
- EngineDeterministic formulas
- RangeNot a single point
- DecisionInvestor compares scenarios
The engine step is code, never a model call. A language model can summarise a lease or draft a follow-up email. It should never be the thing computing a cap rate.
AI reads the lease. It does not do the math
The strongest real use of AI in this sector is extraction, not calculation. Someone is currently keying rent rolls into a spreadsheet by hand. A missed escalation clause quietly invalidates the return on the whole deal.
A model can read that lease and pull out the rent, the term and the escalation schedule. It should not then compute what those numbers mean for cash flow. That step runs through the same deterministic engine every property goes through, with the extracted fields feeding in as typed inputs a person can check.
- Every extracted number links back to the clause it came from, so a reviewer can check it in seconds.
- Low-confidence extractions route to a person instead of silently entering the model.
- Ask any vendor what happens after extraction: does a formula compute the return, or does the model?
The system this argument is built on
A generic calculator against a built platform
Both output a return number. Only one of them shows how it got there.
| Criterion | The generic tool | What a built platform does |
|---|---|---|
| The math | A model estimates the return from a prompt. | A deterministic engine computes it from typed fields. |
| Valuation | A single price, no range. | A range, with the comps that produced it named. |
| Lease data | Keyed by hand into a spreadsheet. | Extracted with a confidence score and a link to the clause. |
| Comparing deals | One property, one tab, no side by side. | Scenario comparison before the investor commits. |
| Advice at the moment of decision | The tool stops at the number. | Scheduling connects the investor to a professional right there. |
The stack this work runs on
Platform
Integrations
Questions investors and founders ask
01Can AI calculate a property's ROI or cap rate directly?
No, and it should not be asked to. A language model predicts plausible text, not arithmetic, so a return it computes can look correct and still be wrong. Deterministic code runs the ROI, cash flow and cap rate formulas from typed inputs. AI's job is reading documents and drafting summaries, never the calculation itself.
02What is an AVM and can I trust the number it gives me?
An automated valuation model estimates a property's worth from data alone, with nobody inspecting it. It cannot see condition, recent renovation or anything outside the data it was trained on. Trust the range it implies, not the single number most tools display. A point estimate with no confidence band is a guess with a decimal point attached.
03How does AI help with rent rolls and lease review?
It extracts terms, not conclusions. A model reads a lease or rent roll and pulls out rent, term length and escalation dates into fields a person can verify. Every extracted number should link to the clause it came from. Anything below a confidence threshold goes to a human reviewer instead of straight into the return calculation.
04How do comps get chosen, and does that change the valuation?
Yes, meaningfully. Selecting comparable sales is a judgment call. Distance, condition and timing all move the number, and two analysts can reach two different valuations for the same property. A platform should show which comps it used and let an analyst swap one out, rather than hiding the selection inside a single number.
05How do you price a real estate investment analysis build?
Scoping calls cost nothing. Where the honest answer needs us inside an existing codebase, we run a paid two-week diagnostic first, and it ends in a fixed price. The main driver is how many data sources feed the underwriting engine, from manual entry through to MLS and lease extraction.
Go deeper
- Real estate software development →The data-model decisions behind property software, across every buyer type.
- SmartREI →The full build: ROI engine, investor and realtor dashboards, CRM sync.
- AVM →Automated valuation models, and why a point estimate is the wrong output.
- Predictive analytics →The forecasting and modeling work behind a returns engine.
