The diagnosis everyone is missing
AI pilots in commercial real estate are not failing because the data is messy. They are failing because nobody built valuation expertise into the system doing the valuing.
The numbers are real. JLL’s 2025 Global Real Estate Technology Survey of 1,500+ senior CRE decision-makers found that 88% of investors, owners and landlords — and 92% of corporate occupiers — have started piloting AI. Yet among occupiers, only 5% report achieving all of their program goals.
A popular explanation is making the rounds on LinkedIn: the model is fine, the data underneath it is broken. Build the “intelligence layer” first — a knowledge graph, a data governance layer and a spatial intelligence layer — and you join the firms beating the failure rate.
Clean data matters. I have spent more than four decades as an appraiser and MAI insisting on it. But clean data fed into a system that does not know how to value real estate still produces an unreliable value. You have simply made the wrong answer faster and more confident.
The missing layer is not data. It is valuation domain expertise, vetted by MAI-designated appraisers and governed by best practices valuation methodologies that every user must follow. Neither the AI model nor the AI-wrapped financial-modeling software sitting on top of it has any of these built in.
Don’t take my word for it. Ask the AI yourself. Paste this into ChatGPT, Claude or Gemini:
“Do you contain valuation methodology vetted by MAI-designated appraisers that you are required to follow, or does that expertise have to come from me if I use you to value and underwrite a CRE property?”
When I ran it in Gemini, in incognito mode with no prior context, the answer was unambiguous: “The core commercial real estate (CRE) expertise, guardrails, and ultimate judgment must come from you.”
Now ask the same question a second time and compare the answers. The wording will change. That is the point.
The missing layer: valuation domain expertise
A large language model has read about appraisal. It has not been trained to appraise. Those are different things.
An LLM knows the vocabulary: cap rate, absorption, fee simple, leased fee, entrepreneurial incentive. What it does not have is the discipline that decides which of those concepts applies, in what order, and with what support, for this property type at this stage of development.
The same is true of the financial-modeling platforms now being wrapped in AI. A cash-flow modeling tool is a calculator. A very good one. But it has no opinion about whether the model the user built is the right model. It will discount whatever cash flows it is given at whatever rate it is given.
And the user often doesn’t know what the model is missing. Even experienced valuation experts routinely overlook or misapply finer points like these:
- Expense stops, expense caps and gross-up provisions, and how each changes the income a lease actually produces.
- Probability of tenant renewal. Fee simple uses the market’s typical renewal probability. Leased fee uses the probability of the actual tenants renewing. They are not the same number.
- Lease start dates. When leases begin in a fee simple analysis is not the same as when they begin in a leased fee analysis.
- Concessions. Should free rent and tenant improvements be modeled as given in the actual leases, or compared against the concessions typical in the market?
- Development stage risk. Discount rates and the allocation of total entrepreneurial profit should vary at each stage of the development process, reflecting the different risks at each stage. One flat rate and one lump-sum profit for the whole project misstate the value at every stage.
- Cost of production versus the Cost Approach. Cost of production (COP) covers all costs, including the “as is” value of the property, up to the “as if complete” non-stabilized stage. The Cost Approach goes further and adds the costs of absorption to stabilized occupancy. Treated consistently, the Cost Approach conclusion should land close to the Sales Comparison, Direct Capitalization and Yield Capitalization conclusions.
Many practitioners assume the AI will handle these with best practices. It doesn’t. It models what it is told, and if nobody tells it, the error flows straight into the value. That is another reason the pilots fail.
In the ValueXpose workflow, these points are not left to chance. The guided questionnaire raises each finer point where the assignment calls for it and applies the best-practice treatment automatically.
So who supplies the expertise? The user. And that is the trap:
- An experienced MAI can catch a broken assumption. A first-year analyst cannot.
- Two analysts at the same firm build two different models for the same property and get two different values.
- When the AI agent builds the model for them, nobody can see which methodology it chose, or why.
Ask the users of any AI underwriting agent a simple question: where, exactly, is the valuation expertise in this system? Most cannot point to it, because it is not there. That inability is why so many pilots stall.
Why you can’t prompt your way to valuation expertise
The key to a successful AI valuation pilot is not cleaner data. It is building in valuation domain expertise first. The AI has already told us it was never trained to supply that.
So whoever builds an AI valuation agent has to teach it, step by step. That takes decades of valuation experience, plus the ability to encode what the ValueXpose team has learned over 15 years of trial, error and field testing. Among the things it would have to learn:
- A best-practice financial model for every property scenario, from stabilized through non-stabilized, partially completed and ground-up development.
- Models wired to each other. Change an input for a new scenario and every connected model updates consistently, in under 2 minutes.
- Developments that are infeasible today but expected to become feasible. When a site is not yet ready for development to its highest and best use, the system must model the holding period in between and derive the “as is” land or property residual market value.
- Timeline-driven adjustments. Move a timetable and dozens of subtle changes must cascade through every model: present value factors, rents, operating expenses, holding periods and cost of production (COP), to name a few.
Now suppose someone did teach an AI all of it. The pilot would still fail:
- Time. A widely shared LinkedIn post celebrated an AI agent that valued the Flatiron Building’s office-to-condo conversion in 89 minutes. That is one scenario. Even at 8 minutes a run, five scenarios eat most of an hour. SmartVal produces the first scenario in under 2 minutes, every additional scenario in under 2 minutes, and reopens any saved scenario in under 15 seconds.
- Inconsistency. Revisit that scenario and you wait all over again, and you get a different value.
- Cost. Every run and every re-run burns tokens.
- No guidance. The agent cannot walk a user through the development appraisal process, so only seasoned valuation experts can use it safely.
And experience alone is not the answer either. Much of our profession, MAIs included, was trained on back-of-the-napkin valuation techniques. Current appraisal education still passes many of those shortcuts from one generation to the next. An AI agent taught by an expert inherits the expert’s shortcuts too.
Why AI cannot give you the same value twice
Large language models are probabilistic by design. Every response is generated one word at a time by sampling from a range of likely next words. That is what makes them good at writing. It is also what makes them unfit to be the valuation engine.
In practice, this means:
- Same inputs, different outputs. Run the same rent roll, the same comps and the same prompt twice, and you can get a different cap rate selection, a different reconciliation weighting and a different final value.
- Methodology drifts. One run treats a below-market lease as a leased fee adjustment. The next folds it into the fee simple value. Nothing in the model forces the correct sequence.
- Judgment is improvised, not encoded. The model does not carry a fixed rule that fee simple comes first, or that cost to stabilize must be deducted from a non-stabilized property. It may do it. It may not.
An appraiser who gave a lender three different values for the same property on the same day would lose the client and possibly the license. Yet that is exactly what a probabilistic engine is built to do.
This is the root cause the data-layer argument misses. Perfect data does not make a probabilistic engine deterministic. If it isn’t defensible, it isn’t a value.
The five ways AI underwriting pilots fail
When the valuation engine itself is AI, the same five problems show up in pilot after pilot.
- Slow scenarios. Underwriting is scenario work: change the cap rate, the lease-up period, the market rent, the exit year. When every scenario is a new AI run, analysts wait on generation, then check the output, then re-prompt when it drifts.
- Costly token bills. Every scenario, every re-run and every “try it again” burns tokens. The more thorough the analyst, the higher the bill. Firms end up rationing the very scenario testing that makes underwriting sound.
- Inconsistent results. Probabilistic output means the same property can produce different values on different runs. A credit committee cannot approve a loan on a number that moves when nobody changed the inputs.
- No governance or consistent best practice. The expertise lives in whoever wrote the prompt. A senior MAI gets one answer, a junior analyst another. There is no firm-wide standard the system enforces, so there is nothing to govern.
- No defensibility or transparency. Ask the AI to show how it got the value and you get a plausible explanation written after the fact, not an audit trail. That does not survive a review appraiser, a regulator, opposing counsel or a deposition.
The result is a black-box financial model that is slow to iterate, expensive to run, unusable by inexperienced staff and indefensible under scrutiny. No firm can scale that. So the pilot ends.
How ValueXpose is built: deterministic at the core, AI at the edges
ValueXpose (VE) was designed by MAI-designated appraisers around one rule: AI never touches the valuation methodology.
The core is SmartVal, a deterministic valuation operating system. The valuation expertise is encoded in the engine, not left to the user or the prompt:
- Built-in financial models for every property type and stage, from stabilized assets through non-stabilized, partially completed and ground-up development, subdivision and mixed-use. Users do not build models, so they cannot build the wrong one.
- A guided questionnaire that adapts to the property type and stage of development and asks only the questions that matter for that assignment.
- Fee simple first, always. Leased fee, rent loss, cost to stabilize and development feasibility are layered on in the correct sequence every time.
- All four approaches (Cost, Sales Comparison, Direct Cap and Yield Cap/DCF) reconciled, then cross-checked by reverse-engineering implied going-in and going-out cap rates from the market data.
- Value Sustainability Analysis (patent pending), which flags when a well-supported reconciled market value, initially in fee simple (or its fee simple equivalent if leased), may not be sustainable over a typical holding period. If that value is not sustainable, the other waypoint values along the development timeline will not be sustainable either, if the property is purchased at that stage of development.
AI works at the edges, where it is genuinely good: digesting due diligence documents, formatting comp sale data sheets, writing the report narrative and reviewing the finished report. The appraiser still sources and verifies every comp.
The engine does the math and enforces the methodology. The AI does the reading and the writing. Neither is asked to do the other’s job.
How VE solves each failure, one by one
Because the SmartVal engine runs no AI, it uses no AI tokens. That single design choice changes the economics and the reliability of underwriting.
That re-run test is the proof. A probabilistic system cannot promise you the same answer tomorrow. A deterministic one must. Revisit a scenario next month and you will get the same value you got today, down to the dollar, because nothing in the engine is guessing.
The first-assignment test
Here is the real measure of whether domain expertise is built into a system: hand it to your least experienced analyst.
With an AI underwriting agent or a blank financial-modeling template, a first-year analyst does not know which model to build, which assumptions to question or when the output is wrong. The system will not tell them.
With ValueXpose, that same analyst answers a guided questionnaire written by MAIs. The engine selects the right financial models for the property type and stage, runs every approach in the correct sequence and enforces the checks. An analyst on the first assignment of their career can complete a complex valuation and underwriting assignment, such as a non-stabilized or development property, on the first try.
The expertise is in the system, not in the operator. That is what makes it scalable, governable and defensible.
That goes beyond fixing failed AI pilots. It ushers in the next generation of valuation experts. New analysts become productive immediately, with minimal training from their mentors, and they learn best practice by doing it correctly from the first assignment rather than inheriting the shortcuts of the past.
The bottom line
The firms piloting AI for underwriting are not wrong to want speed and scale. They are wrong about where to put the AI.
Credible valuation and underwriting require five things at once: fast scenarios, low operating cost, identical results on identical inputs, enforced best practice, and a transparent audit trail. A probabilistic engine cannot deliver these five at once, no matter how clean the data underneath it. Each scenario is slower and costlier to run, and it cannot guarantee identical results, enforce best practice or produce a true audit trail.
ValueXpose is, to our knowledge, the only deterministic valuation operating system that delivers all five in one platform, while still using AI where it makes appraisers and analysts more productive.
Deterministic at the core. AI at the edges.
So stop spending thousands of dollars in AI tokens testing AI valuation agents that are structurally unable to deliver credible, repeatable values. More data, better prompts and bigger models will not change how a probabilistic engine behaves. Put the AI where it belongs, at the edges, and put the valuation on a deterministic core.
Instead, pilot-test ValueXpose. The valuation engine runs at no token cost, so you can test as many scenarios as you like. And if your testing finds anything structurally wrong with the methodology, show us your support. Our expert development team can correct it immediately.
If your firm is running an AI underwriting pilot, put it to a simple test. Run the same scenario twice. Then run it on SmartVal. I will gladly set up a live side-by-side on one of your own properties with your least experienced analyst at the keyboard. Message me to arrange it.
Schedule a time at no charge to set up a live demonstration using your AI test subject property an lets compare results
