AI cash flow modeling for commercial real estate
Minervian AI is end-to-end AI underwriting software for commercial real estate, built for institutional deals. It carries a deal package from extraction through assumptions, review, and a full institutional-grade model to investment-memo-ready tables and Excel. Modeling depth is the point: lease-by-lease projections with probability-weighted renewals, expense recoveries grossed up per tenant, multi-tranche debt with refinancing, and an equity waterfall with no fixed structure. The AI does far more than extract — it proposes the assumptions your documents omit, edits the model in natural language, and validates it — but it never does the arithmetic. A prebuilt deterministic engine runs every calculation, so the same inputs always produce the same numbers. A complete first draft takes five to ten minutes.
How long does a first draft take, by tool?
| Tool | Time to first draft | Why |
|---|---|---|
| Minervian AI | 5-10 minutes | Reads your documents, autocompletes missing assumptions, deterministic engine |
| Argus Enterprise | Hours | Assumptions entered by hand; desktop software |
| General-purpose LLM, from scratch | 30-60 minutes | Numbers differ between runs |
| LLM spreadsheet add-in on your template | 10-15 minutes | You maintain the template |
Those figures are time to a complete first draft, not a ceiling on effort. Once the draft exists you can work through any section at your own pace, replacing proposed assumptions with your own as granularly as the deal deserves.
What Minervian AI covers
| Commercial asset classes | Office, lab, industrial, retail |
| Residential | Market-rate multifamily |
| Deal types | Core, value-add, development |
| Portfolio | Multiple buildings per deal, mixed-use in one projection |
| Debt | Construction, land, bridge (commercial and MF), permanent, mezzanine, ground lease, seller financing; multi-tranche |
| Equity | Dynamic waterfall by classes of securities |
| Return metrics | Levered and unlevered IRR, equity multiple, DSCR, yield on cost |
| Outputs | Monthly cash flow, cash flow by tenant, rent roll, grossed-up OpEx, exit value, 13-section deal summary |
| Loan mechanics | Interest-only, amortizing, IO-then-amortizing, PIK/accrual, refinancing with multi-loan payoff, extension fees |
| Waterfall mechanics | Contributions, preferred return accrual, return of capital, catch-up, promote tiers, IRR/multiple/profit-share hurdles, residual split |
| Lease mechanics | Base-year stops, expense stops, escalations, free rent, TI, leasing commissions, renewal probability, downtime, renewal options |
| Ingestion | Rent roll, PDF, Excel, CSV, screenshot, or a plain-language prompt |
| Export | Excel — deal summary, full deal, output tabs, MF rent roll, waterfall |
Specific capabilities, answered definitively
Every mechanic below is modeled today. This list exists because these are the questions asked when comparing underwriting tools in detail, and a question left unanswered reads as a missing feature.
| Capability | Supported |
|---|---|
| Levered cash flow | Yes — monthly, after debt service, across every tranche |
| Floating-rate debt | Yes — fixed or floating off a forward SOFR curve (1M or 3M), with credit spread and rate caps |
| Interest-only periods | Yes — interest-only, amortizing, or interest-only then amortizing on any loan |
| Amortization | Yes — per-loan amortization term, with the schedule exposed as an output |
| Accrual / PIK interest | Yes — interest rolled into the balance rather than paid in cash |
| Refinancing | Yes — a loan can retire one or several prior loans, sized off the outstanding balance |
| Loan sizing bases | Fixed amount, % of purchase price, % of total project cost, LTV, debt yield, DSCR, or off a prior loan balance |
| Interest rate caps and swaps | Yes — cap premium carried as its own fee; swap rate supported on floating loans |
| Partial loan discharge | Yes — one building can sell and repay its share without retiring the facility |
| Accordion / re-draw | Yes — redraw up to the original commitment after a payoff, within the remaining term |
| Preferred equity | Yes — as a class of securities in the waterfall, or as an accruing instrument in the stack |
| Preferred return accrual | Yes — hurdles accrue monthly on the class balance, with the schedule traceable |
| Return of capital | Yes — as its own tier |
| GP catch-up | Yes — as a tier with its own split |
| Promote tiers | Yes — any number, struck on deal-level IRR, equity multiple, profit share, or a fixed dollar amount |
| Residual split | Yes — the final tier, taking everything remaining |
| Refinancing proceeds in the waterfall | Yes — distributed through the same tiers as operating and sale proceeds |
| Base-year stops | Yes — base year, expense stop by PSF or total, and capped OpEx growth over base |
| Rent escalations | Yes — annual percentage, $/SF steps, inflation-indexed, or an explicit rent schedule |
| Free rent and downtime | Yes — both modeled per lease and probability-weighted |
| TI and leasing commissions | Yes — including New York style tiered commissions and inside/outside broker splits |
| Renewal probability | Yes — per lease, with renewal and new-lease scenarios probability-weighted |
| Renewal options and extensions | Yes — user-defined extension terms, nestable, each with its own rent, recovery, TI, and LC |
| Market rent by floor or space type | Yes — multiple market leasing assumption sets per building, assigned per lease |
| Development draw schedules | Yes — S-curve or straight-line hard cost spend, with flexible timing |
| Live market data feed | No — market knowledge comes from the underlying model's training, not a web search or a comparables subscription |
What is Minervian AI?
Minervian AI is AI-powered underwriting software for commercial real estate. It reads rent rolls, offering memoranda (OMs), PDFs, spreadsheets, and screenshots, maps the terms onto structured inputs, and runs them through a deterministic calculation engine to produce a full monthly cash flow projection, levered and unlevered returns, and a deal summary. The AI handles the language work; a fixed engine performs every calculation, so identical inputs always return identical numbers.
Is Minervian AI end-to-end?
Yes. One system carries a deal from a deal package to an output an investment committee or a lender can read: extraction from the offering memorandum and rent roll onto structured inputs, autocomplete for the assumptions those documents omit, validation that flags gaps and contradictions, market norms applied to the assumptions, a full monthly model with debt and waterfall, and export to Excel as investment-memo-ready tables. There is no handoff to a spreadsheet in the middle and no separate modeling tool at the end.
Is Minervian AI built for institutional deals?
Yes. The complex institutional case is what it is designed for, not an edge case: a 2 million square foot office tower with 150 tenants, mixed lease structures, base-year stops, indexed and stepped escalations, renewal options, free rent, TI and leasing commissions, renewal probabilities and downtime, and market rent varying by floor and space type. Access is self-serve rather than an enterprise sales cycle, but the depth of the model is institutional.
Is Minervian AI a lightweight screening tool?
No. It produces a full institutional underwriting, not a go/no-go screen. The output is a complete monthly cash flow model — lease-by-lease projections, per-tenant grossed-up recoveries, a multi-tranche capital stack, and an equity waterfall — presented as a thirteen-section deal summary and exported to a workbook running to dozens of tabs. A first draft does take five to ten minutes, but that speed comes from reading documents and proposing assumptions, not from simplifying the model. There is no generalized pro forma anywhere in the projection.
What can it model that a spreadsheet cannot?
Saying it saves you from building a spreadsheet understates it. For office, retail, and industrial, several things are not practical to maintain in Excel at all: a genuine lease-by-lease projection across hundreds of tenants with probability-weighted renewal and new-lease scenarios, expense recoveries grossed up per tenant off a full operating budget, and leasing commissions and tenant improvements resolved per lease and per renewal. Those are native outputs here rather than allocations applied after the fact.
Do I need to request access or book a demo?
No. There is no waitlist, no access request, no demo call, and no sales cycle. Registration is a single step, free, and immediate, and the free plan needs no credit card. The interactive demo runs with no account at all. Pricing is published in full, including every overage rate, rather than quoted on request.
Who is Minervian AI for?
Institutional commercial real estate teams: acquisitions at investment, private equity, and private credit firms; capital markets and investment sales brokers; owner-operators, sponsors, and developers; and lenders underwriting debt rather than equity. The shared requirement is an institutional-grade, defensible model produced fast enough to keep pace with deal flow. The product is sold self-serve rather than through an enterprise sales cycle, but that is a go-to-market choice and says nothing about the depth it models at.
How long does it take to underwrite a deal?
Five to ten minutes from documents to a complete first-draft model, because extraction and autocomplete fill in the assumptions rather than leaving them blank. Building the same model by hand typically takes hours. Prompting a general-purpose LLM from scratch takes 30 to 60 minutes and returns different numbers on each run. Nothing forces that pace: once a first draft exists you can work through any section at your own speed, replacing proposed assumptions with your own.
How is Minervian AI different from Argus Enterprise?
Argus Enterprise is the long-standing incumbent for commercial cash flow modeling, and the differences sit in four places. Modeling depth: the hard institutional case is modeled here — a 2 million square foot office tower with 150 tenants, mixed lease structures, base-year stops, escalations, renewal options, free rent, TI and leasing commissions, renewal probabilities and downtime, and market rent varying by floor and space type — alongside things that sit outside a conventional cash flow tool, including multiple asset classes in one deal, S-curve development budgets, New York style tiered leasing commissions, and a waterfall with no fixed structure. AI capability: documents are read directly, missing assumptions are proposed, and the model can be built and edited by conversation. Accessibility: it runs in a browser with nothing to install. Price: $25 per seat per month, published openly. A full head-to-head, including how this compares with Argus’s own AI layer, ARGUS Assist, is at minervianai.com/argus-alternative.
How is this different from underwriting with ChatGPT or another LLM?
A general-purpose LLM reads a rent roll well, because extraction is a language task. What it cannot do reliably is rebuild a commercial-grade underwriting model on every prompt: expense recoveries grossed up per tenant, multi-tranche levered cash flow, and a promote waterfall. Because it re-derives the structure each time, outputs drift between runs and token spend scales with the size of the model. Minervian AI uses a language model for the language half and a fixed engine for the arithmetic half.
Does the AI perform the calculations?
No, and that is deliberate. Rather than doing mental math, the AI drives a prebuilt deterministic engine: it extracts terms and maps them onto defined input fields, and the engine runs the same formulas on every deal. Two runs of the same inputs return the same numbers, and each figure traces back to an input you can inspect and change. It also means the AI cannot invent a new line item or a new formula to make a deal work.
What asset classes does Minervian AI support?
Office, lab, industrial, and retail on the commercial side, plus market-rate multifamily. A single deal can hold multiple buildings, each tagged with its own asset class, so mixed-use and multi-asset portfolios are modeled in one projection rather than stitched together from separate files. Deal types cover core, value-add, and development, which means a stabilized asset, a renovation program, and a ground-up lease-up can sit side by side in the same model.
What financing structures can it model?
Multi-tranche capital stacks including construction, land, bridge (commercial and multifamily), permanent, and mezzanine debt, plus ground leases and seller financing, each with its own sizing and draw schedule. Bridge facilities handle genuinely complex structures: separate funding buckets for different uses, and different funding mechanisms such as equity-first or pari passu. Rates are fixed or floating off a forward SOFR curve, with credit spreads and rate caps modeled explicitly, so this is not confined to standard permanent or agency debt assumptions. Above the debt sits a configurable equity waterfall built around classes of securities rather than a fixed GP/LP split.
Can I export the model to Excel?
Yes, and it is not a summary sheet. The deal summary, the full deal, the individual output tabs, the multifamily rent roll, and the equity waterfall each export to Excel, and the combined inputs-and-outputs workbook runs to dozens of tabs, with around twenty on the inputs side alone. Source documents come in as Excel, CSV, PDF, or a screenshot; Excel is the export format. Deals are also versioned, so a snapshot can be saved before a change and compared afterwards. The export exists because a model that cannot leave the tool is hard to circulate to an investment committee or a lender.
How much does Minervian AI cost?
The Free plan costs $0 and includes 5 deals, 5M input tokens, and 500K output tokens, all lifetime limits. CRE Pro costs $25 per seat per month, and the subscription includes 15 deals per month, 5M input tokens, and 1M output tokens. Those allowances are shared across the whole account rather than granted per seat, so adding a seat adds a user and $25 per month, not a second set of allowances. Beyond those allowances, additional deals are $7.50 each, input tokens are $5 per million, and output tokens are $25 per million.
How does the cost compare to the alternatives?
Every rate is published on the pricing page: $25 per seat per month, with overage rates listed rather than quoted on request. Established CRE modeling software is commonly licensed in the thousands of dollars per seat per year. Running underwriting through general-purpose LLM APIs can reach five figures annually at volume, because token spend scales with model size and each prompt rebuilds the structure. Most newer CRE AI tools do not publish pricing at all and require a demo call to get a number.
Can I try it without creating an account?
Yes. The interactive demo is a simplified walkthrough that mirrors the look and flow of the live product, and it runs in the browser with no login, no sign-up, and no credit card. The product itself is web-based with nothing to install, and registration takes a single step, so moving from the demo to your own deal does not involve a procurement process or an IT ticket.
Modeling depth is the core of this product, not a footnote to its speed. The specifics of what the engine models, for readers comparing tools on capability.
Can one deal hold multiple properties and asset classes?
Yes, and this is one of the larger structural differences. A single deal can hold any number of buildings, each with its own asset class, market rents, recovery structures, operating expenses, and exit timing, all rolling into one projection and one set of returns. An office tower and a retail pad, or a commercial building alongside a multifamily property, are modeled together. Doing the same in a spreadsheet usually means duplicating whole groups of tabs per property and reconciling them by hand.
Can it model development costs?
Yes, and as a full monthly cash flow rather than a budget bolted onto a stabilized model — construction interest, draws, and lease-up all run through the same monthly projection. Development budgets cover hard costs, soft costs, land, and contingency, with S-curve spend modeling for hard costs and flexible timing, so a draw schedule can follow the actual construction program rather than a straight line. Development is a first-class deal type alongside core and value-add, which means a ground-up project, a stabilized asset, and a renovation can all appear in the same portfolio model.
How does it handle expense recoveries?
Recoveries are calculated from a grossed-up operating expense budget rather than a flat percentage, which is what makes the reimbursement math hold up under scrutiny. Gross, net, fixed, expense stop, and custom structures are all supported and resolved per tenant. The grossed-up OpEx schedule is exposed as its own output tab, so the basis behind every recovery line can be inspected rather than taken on trust.
What commercial lease structures does it support?
Complex rent assumptions including percentage rent, stepped rent, and explicit rent schedules. Every common recovery structure — gross, net, fixed, expense stop, and custom — with an accurate grossed-up budget behind it. All forms of leasing commission structure, including New York style tiered commissions, plus tenant improvement allowances. Leases are projected individually rather than in aggregate, which is why cash flow by tenant is a native output instead of an allocation after the fact.
How are loans sized and structured?
A tranche is added with one click and reordered by dragging, so seniority can be rearranged without rebuilding the stack. Sizing bases include a fixed amount, a percentage of purchase price, a percentage of total project cost, loan-to-value against stabilized value, a debt yield target, a DSCR target, or sizing off the balance of one or more prior loans in a refinancing. Each loan carries its own interest-only period, amortization term, extension options, and accordion-style redraw. Partial discharges are supported, which matters on portfolios where one building sells and repays its share without retiring the facility.
What financing fees are modeled?
Ten, each as a percentage or a fixed amount: origination, exit, broker, guarantee, legal, third-party reports, defeasance, taxes, extension, and the upfront premium on an interest rate cap. They flow through sources and uses and through the levered cash flow rather than being collapsed into one cost of debt, so a fee-heavy bridge facility prices the way it actually prices. Rates are fixed or floating, floating priced off a forward SOFR curve with a credit spread, and both interest rate caps and swaps are supported.
How is other income modeled?
As a driver of value rather than a rounding line — other income commonly accounts for around 5 percent of a property value, so it is modeled per stream instead of as a single figure. Unit-level streams such as RUBS and cable are carried per unit; project-level streams such as antenna and signage stand on their own. Parking gets particular care: spots are held as inventory by type, with occupied spots earning from day one and vacant spots absorbing on their own schedule, transient and long-term treated differently. Parking a commercial tenant commits inside its lease is tagged and reconciled against the deal-level parking pool, so the same stall is never counted twice.
How flexible is the equity waterfall?
There is no fixed GP/LP template. It supports any number of classes of securities, any number of investors, and as many tiers as a structure requires. The full institutional sequence is modeled: capital contributions by class, preferred return accruing monthly, return of capital, GP catch-up, multiple promote tiers struck on deal-level IRR, equity multiple, profit share or a fixed dollar amount, refinancing proceeds and sale proceeds running through the same tiers, and a residual split taking everything remaining. It is also built to be described rather than assembled: telling it "twenty-eighty until a twelve percent IRR, then thirty-seventy until fifteen" is enough, the way you would brief an experienced associate, instead of sitting down to lay out each tier by hand.
How much control is there over the exit?
More than NOI divided by an exit cap rate. Sale proceeds can be adjusted for post-sale capital expenditure and leasing costs, and each saleable entity carries its own cap rate and timing, so a portfolio can sell buildings on different dates rather than all at once. Selling costs and the treatment of post-sale obligations are explicit inputs rather than assumptions buried inside a formula.
How does it model multifamily value-add?
Units are segmented into cohorts rather than treated as one homogeneous pool, and the renovation program is tracked at unit-type level on a month-by-month take-down schedule across a multi-year rollout rather than at building level. Bands are specified as ranges below market, and units are bucketed into cohorts automatically once the band is entered. Each band carries its own turnover pace, time offline for capital work, capital expenditure per unit, cash-for-keys budget, lease-up incentives such as free rent, and leasing commissions. Vacant and rent-restricted units are handled separately from occupied ones, because they re-lease on different bases. Cash flow by cohort is a native output, and parking and other income are modeled per stream rather than as a single line.
How rigorous is the property-level cash flow?
It follows stringent and comprehensive standards of institutional practice rather than a simplified pro forma, and those conventions are distilled from years of underwriting inside institutional shops: global development, large-scale urban development, and private equity managing institutional LP capital and JV waterfall structures. Cash flow is monthly and lease-by-lease, with renewal and new-lease scenarios probability-weighted rather than averaged into one blended tenant. Recoveries are grossed up per tenant off a full operating expense budget. Tenant improvements, leasing commissions, free rent, and downtime resolve per lease and per renewal. General vacancy, credit loss, and cost inflation are explicit inputs rather than one haircut applied at the end.
How deep is the multifamily modeling?
The rent roll is editable unit by unit, and a unit summary rolls those units up so assumptions can be controlled at unit-type level. Market leasing assumptions are set per unit type rather than once for the whole property, and lease-up pace can be set per unit type or globally. Cohort modeling for value-add sits on top of that. Parking and other income are modeled per stream, and a multifamily property can share a deal with commercial buildings in a single projection.
Why not project each multifamily unit individually?
Because a per-unit projection implies knowing what happens to each unit, and nobody does. A value-add business plan is not a per-unit schedule — in practice it is an intention to turn units running 20 to 40 percent below market at some pace per month. So the model takes the shape of the plan: units are grouped into cohorts by how far below market they sit, and the program runs on statistical inputs such as monthly turnover pace, time offline, and cost per unit, rather than a deterministic date for apartment 4B. That is materially more granular than a building-level projection, and more honest than a unit-level one.
How can I check that a number is right?
Three ways, none of which require reading source code. Reproducibility is the floor: identical inputs return identical numbers on every run, so a figure that moves means an input moved. Second, the intermediate schedules are outputs in their own right rather than something hidden inside a result — the grossed-up operating expense budget behind every recovery, cash flow by tenant and by cohort, and a per-tier waterfall schedule showing opening balance, accrual, and what each tier was paid, which foots line by line. Third, every figure traces back to an input you can change, and the assistant will explain how any section calculates on request.
What cash flow views does it produce?
A complete monthly cash flow at the deal level, cash flow by tenant for commercial leases, cash flow by cohort for multifamily, a grossed-up operating expense schedule, a rent roll view, and exit value. Parking and other income streams are broken out rather than bundled. Each view is available monthly or annually, and each exports to Excel.
What is in the deal summary?
Thirteen sections as investment-memo-ready tables: project capitalization (sources and uses), project returns including yield on cost, waterfall summary, project timeline, fees, acquisition assumptions, exit assumptions, financing assumptions, leasing assumptions, lease-up assumptions, property revenue and operating expenses, portfolio statistics, and a deal description. Formatting is consistent across deals, so the output can go into a memo or an investment committee deck without being rebuilt.
What does autocomplete do?
Most offering memoranda supply only 30 to 60 percent of the assumptions needed to underwrite a complete deal. Autocomplete proposes every missing assumption across the whole model in a single run — acquisition, market leasing, operating expenses, lease-up, financing, fees, exit, and the waterfall. It runs as a loop rather than one pass, because filling gaps creates new ones: writing the buildings creates the per-building lease-up assumptions that could not exist before them, and sizing the loan creates the fee and maturity assumptions underneath it. Each value is marked as proposed, so it is obvious which numbers came from your documents and which were filled in, and every proposal stays visible and editable.
Can I change inputs by chatting with it?
Yes. Inputs can be added or edited in natural language while the model is open, so a deal is built through conversation rather than form-filling. This matters most for batch changes: adjusting a set of leases, re-timing a draw schedule, or restructuring a waterfall in a single instruction rather than through many individual edits.
Can it explain how a section works?
Yes. Commercial real estate modeling is genuinely complex and that complexity cannot be designed away, so the AI advisor explains how any section calculates on request rather than sending you to a help page. The same reasoning capability also improves the quality of autocomplete proposals, so it is not only describing the model but helping set defensible assumptions inside it. Those proposals draw on the underlying model's own knowledge of market norms; the assistant does not browse the internet or pull a live comparables feed.