Most small commercial real estate firms already underwrite with two tools open: an Excel model on one screen and ChatGPT on the other. The chatbot drafts the investment memo, researches a submarket, and untangles a broken formula; the spreadsheet holds the numbers. It is a good stack, it costs almost nothing, and for a firm doing a handful of deals a month it is often all you need. This article answers a narrower, more useful question than “which is better”: at what point does the Excel-plus-chatbot stack stop scaling, and what does a purpose-built underwriting copilot do that the free version cannot? The honest answer is that most lean firms are not at that point yet — but the ones that are can name exactly what got them there.
The short answer
Stay on Excel plus ChatGPT when your deal volume is modest, your model changes shape from deal to deal, and one person can eyeball every number. Commission a custom underwriting copilot when you screen enough deals that manual pre-fill is a bottleneck, you run essentially the same model every time, and more than one person underwrites — so consistency and an audit trail have become real costs rather than nice-to-haves.
The dividing line is not intelligence. The chatbot is smart enough to reason about a deal. The line is where three things converge: throughput, repeatability, and reproducibility. A copilot wins when you need the same reviewed model, populated the same way, produced fast, every time. Below that, it is an expensive fix for a problem the free stack already handles.
What “Excel plus ChatGPT” actually is
Two tools do two distinct jobs, and conflating them is where firms get into trouble.
Excel is the model. The rent roll, the ten-year cash flow, the debt sizing, the returns waterfall — all of it lives in cells and formulas, most of it on a template from a source like A.CRE or one your firm wrote years ago. This is the system of record. Its numbers go to committee.
ChatGPT — or Claude, or Microsoft Copilot inside Excel — is the assistant beside the model. It does not hold the deal; it helps you work on it. You paste a paragraph and ask for a memo; you describe an assumption and ask whether it is market; you show it a broken formula and ask why it errors. Microsoft Copilot goes further by living inside the workbook, writing formulas and explaining ranges without a copy-paste round trip.
The assistant reasons in language; your model runs on arithmetic. It is genuinely useful right up to the moment you ask it to move numbers into cells — which is exactly where the trouble starts.
Where the free stack works well
The strengths are real, and most firms under-use them.
- Investment memo drafting. Hand the chatbot a finished model and your bullet points and it produces a first-draft memo — summary, thesis, risks, recommendation — in minutes. This is the single highest-return use of AI in a small shop’s deal process.
- Assumption research. Sanity-checking an exit cap, an expense ratio, or a re-tenanting timeline is a fast starting point you then verify against real comps — the same discipline that governs choosing dedicated deal-screening tools.
- Formula help. “Why does this IRR cell error?” or “write a formula that caps rent growth at 3%” is squarely in the assistant’s wheelhouse.
- Scenario narration. Once your model runs three cases, the chatbot turns the output into plain-English explanation an investor can read. It describes what the model computed — it does not compute it.
The pattern: every strong use has the assistant working on text or logic, with the numbers under human control in Excel. That is the free stack at its best.
Where the spreadsheet-plus-chatbot stack stops scaling
Three failure modes appear as volume climbs, and each is quiet until it is expensive.
Silent numeric errors. Ask a chatbot to read a rent roll or a T-12 and hand back the numbers, and it will — some of them wrong, and none flagged. A language model predicts plausible text; a figure that looks right (a $4,200 rent that should be $4,020) is the most dangerous error because nothing about it looks off. General AI tools are not reliable at extracting structured numbers from documents without validation, and a wrong number that flows into a returns calc corrupts the model silently. One deal, you catch it. Twenty a month, you will not.
Non-reproducibility. Underwrite the same building twice through the chatbot and you can get two different models, because each session starts from scratch and the assistant is not deterministic. There is no canonical “our model” — only whatever this conversation produced. When committee asks why last quarter used a different reserve assumption, “the AI did it differently that time” is not an answer.
It does not compound. This one matters most and gets noticed least. Every deal starts over. The chatbot has no memory of your exit-cap logic, your stabilization rules, your standard reserve, or the last hundred deals you underwrote. That knowledge lives in analysts’ heads, not the tool. A larger competitor with a house model gets faster and more consistent with every deal; the free stack resets to zero. We map that compounding gap across the whole deal process in our deal-analysis playbook for lean teams.
What a custom underwriting copilot actually is
A custom underwriting copilot ingests a deal’s raw documents and hands your analyst a model that is already most of the way built — in your format, using your assumptions. It is more modest than “custom AI” sounds, and it has three parts.
A document-ingestion layer. It reads the offering memorandum, rent roll, and trailing financials — PDFs and scans included — and pulls the structured facts: units, in-place rents, lease expirations, historical income and expense lines.
A model-population engine. It writes those facts into your firm’s template, not a generic one. In-place rents land in the rent-roll tab; T-12 lines map to your expense categories; your standard growth, vacancy, and reserve assumptions pre-fill, so the analyst starts from a roughly 70% complete model instead of a blank sheet.
Your encoded standards. This is the asset. The copilot holds your exit-cap logic, stabilization timeline, reserve-per-unit, and debt-sizing conventions. Two people running two deals produce models built the same way, because the standard lives in the tool rather than in whoever built it.
Built well, it cuts the mechanical pre-fill — the hours of transcribing a rent roll and wiring a template — from most of a day to minutes of review, and it leaves an audit trail of which document produced which number. It is a scoped project, not an open-ended platform, and we size that build in our breakdown of what custom underwriting automation costs.
Why “structurally correct formulas” is the whole game
Here is the distinction a principal needs to internalize, because it is the reason a copilot beats a chatbot dumping numbers into a sheet.
Ask a general chatbot to build a cash-flow projection and it may return a table that looks reasonable. The problem is that the entries are often values, not formulas. Year 2 NOI is typed as $1,040,000 — hardcoded — rather than written as Year 1 NOI times your growth rate. It looks identical on screen. It behaves nothing alike.
Change one assumption — bump growth from 3% to 3.5%, extend the hold, adjust vacancy — and a formula-driven model recalculates while a hardcoded one silently stays wrong. Every downstream number now lies, and nothing on the surface tells you. A model full of plausible hardcoded numbers is more dangerous than one with an obvious error, because the obvious error gets caught and the plausible one ships to committee.
A purpose-built copilot generates models where the formulas are structurally correct — Year 2 references Year 1, debt service references the loan terms, returns reference the cash flows — so the model stays honest when you flex it. That structural correctness, not raw speed, is what you are buying.
The five variables that decide it
Score your firm on five questions. The more that point to “build,” the sooner a copilot earns its cost.
- Volume. How many deals do you fully underwrite a month? A handful favors the free stack. Dozens, where pre-fill is a bottleneck, favors a copilot.
- Model repeatability. Do you run one house model across deals, or a bespoke structure each time? A stable, repeated model is what a copilot automates well; genuine one-offs are not.
- People underwriting. One person holds consistency in their head. Three cannot — and drift across analysts is exactly what encoded standards remove.
- Audit-trail need. Do investors or lenders require you to show how a number was derived? A copilot’s document-to-cell trace answers that; a chat session does not.
- Data sensitivity. How much confidential deal data are you pasting into a chat window today? The more sensitive the terms, the stronger the case for an owned environment.
Three or more pointing to “build” and the rekeying, drift, and lost memory almost certainly cost more than a scoped copilot would. Two or fewer, and the free stack is still the right tool — a build would solve a problem you do not yet have.
What each option costs
Excel plus ChatGPT is close to free: business-tier AI accounts run roughly $25–35 per user a month, and you already own Excel. That price is why it is the correct starting point for almost every small firm.
A custom underwriting copilot is a project-based build. Small-to-mid custom automation engagements sit in a market range of roughly $25–150K depending on scope — how many document types it ingests, how complex your model is, and how much of your standard it encodes. Off-the-shelf platforms exist between the poles: lending-oriented tools such as Blooma automate credit underwriting, and Argus Enterprise remains the institutional standard for cash-flow and valuation modeling, though its price and complexity aim well above a 4–20 person shop.
The real comparison is not the sticker. It is the build cost against the loaded hours your team spends transcribing documents and rebuilding models, plus the risk cost of silent errors and drift. High volume tips that math; low volume does not. The same buy-versus-build logic governs the pipeline layer, which we work through in our comparison of Dealpath against a custom pipeline build.
The confidential-terms question
Underwriting means handling price, seller motivation, and financing terms you do not want leaking. Two rules keep the free stack safe.
First, never paste live deal terms into a free or personal AI account, where inputs can be used for model training by default. Use a business-tier account — ChatGPT Business, Claude Team, or the API — which does not train on your data by default. That one change removes most of the exposure at almost no cost.
Second, a custom copilot’s advantage here is structural: deal data flows through your own environment rather than a third-party chat window, so the question of what a vendor might do with your inputs largely disappears. For a firm handling institutional partners’ confidential deals, that control is sometimes the deciding variable on its own.
The threshold: when to build
For most 4–20 person firms, the recommendation is to stay on the hybrid stack and use it better — business-tier accounts, the chatbot for memos and research and formulas, every number entered and checked by a human in Excel. That is not a consolation prize. It is the correct answer until your volume and a stable house model justify more.
Build a custom copilot when the five variables converge: real throughput, a repeatable model, multiple underwriters, an audit-trail requirement, and confidential data you would rather keep in-house. At that point the free stack is not failing because the AI got worse — it is failing because you outgrew the shape of the tool. The copilot turns your underwriting standard from tribal knowledge into an asset your firm owns and compounds, the way an institutional shop does with a fraction of the headcount — a theme we develop across the small-firm AI manifesto.
FAQ
Can ChatGPT underwrite a commercial real estate deal on its own?
No, not reliably. ChatGPT is strong at the language parts of underwriting — drafting the memo, researching assumptions, explaining scenarios, fixing formulas — but it is not reliable at extracting structured numbers from a rent roll or T-12 and populating a model without human validation. It reasons in text; a model runs on arithmetic. Use it beside your Excel model, not as the model.
What is the difference between Excel plus ChatGPT and a custom underwriting copilot?
Excel plus ChatGPT is a general spreadsheet next to a general assistant: you enter and check every number, and the assistant helps with narrative and logic. A custom copilot ingests a deal’s documents and pre-fills your specific model with structurally correct formulas and your standard assumptions, so the analyst starts from a mostly-built model. The first is nearly free and manual; the second is a build that automates the mechanical pre-fill.
Why are hardcoded numbers in an AI-generated model dangerous?
Because they look identical to a real formula but do not recalculate. If a chatbot types Year 2 NOI as a fixed value instead of referencing Year 1 times your growth rate, the number looks right until you change an assumption — then every downstream figure is silently wrong with nothing to flag it. A plausible hardcoded number is more dangerous than an obvious error, because the obvious error gets caught.
When does the Excel plus ChatGPT stack stop scaling?
When three conditions arrive together: deal volume rises enough that manual document-to-model transcription is a bottleneck; you run essentially the same house model every time; and more than one person underwrites, so inconsistency and a missing audit trail turn into real costs. Below that, the free stack is the right tool. Above it, the rekeying and drift cost more than a copilot would.
How much does a custom underwriting copilot cost?
It is a project-based build, not a subscription. Small-to-mid custom automation engagements fall in a market range of roughly $25–150K depending on how many document types it reads, how complex your model is, and how much of your standard it encodes. Weigh that against the loaded hours your team spends transcribing documents and rebuilding models, plus the risk cost of silent errors.
Is it safe to paste deal terms into ChatGPT?
Not on a free or personal account, where inputs can be used for model training by default — and deal terms are exactly what you do not want in a training set. Use a business-tier account (ChatGPT Business, Claude Team, or the API), which does not train on your data by default. A custom copilot avoids the question by keeping deal data in your own environment.
Should a small CRE firm build a copilot or keep using ChatGPT?
Most 4–20 person firms should keep the hybrid stack and run it well — business-tier accounts, the assistant for memos and research, every number checked by a human in Excel. Build only when throughput, a repeatable model, multiple underwriters, and an audit-trail need converge. Building before that solves a problem you do not yet have.
Does Microsoft Copilot in Excel replace a custom underwriting copilot?
No. Microsoft Copilot in Excel is a strong in-workbook assistant — it writes formulas, explains ranges, and finds patterns without a copy-paste round trip — but it is a general tool that does not know your underwriting standard or reliably ingest a deal’s documents into your specific model. It improves the free stack; it does not encode your firm’s house model the way a custom copilot does.
Can a copilot and ChatGPT be used together?
Yes, and the best teams do. A custom copilot handles the deterministic, high-stakes work — document ingestion and structurally correct model pre-fill — while a general assistant handles the language work of memos, research, and scenario narration. They are complements: one produces the model, the other explains it.
Key takeaways
- Excel plus ChatGPT is the right starting stack for almost every small CRE firm: near-free, and strong at memos, assumption research, formula help, and scenario narration — with every number checked by a human in Excel.
- The free stack stops scaling at three failure modes: silent numeric errors from chatbot-transcribed figures, non-reproducible models, and no compounding — every deal starts from zero because the tool holds no institutional memory.
- A custom underwriting copilot ingests a deal’s documents and pre-fills your specific model with structurally correct formulas and your encoded standards, so the analyst starts from a mostly-built model.
- Structural correctness, not speed, is what you buy: a model whose formulas recalculate when you flex an assumption stays honest; one full of plausible hardcoded numbers does not.
- Score five variables — volume, model repeatability, people underwriting, audit-trail need, and data sensitivity. Three or more pointing to “build” and a copilot earns its cost; two or fewer, and the hybrid stack is still right.
Not sure which side of the threshold your firm sits on? A short, free AI-readiness assessment will map your deal volume, model repeatability, and data-handling needs and tell you exactly where the spreadsheet-plus-chatbot stack stops paying off. Book your free AI-readiness assessment → and we will size it for your firm.
Dirk Jan van Veen, PhD
