A widely shared post this week ran the Flatiron Building through an AI underwriting agent. From the address alone, in 89 minutes, it concluded: $161 million in, $739 million of condo sales out, roughly $350 million of profit, and a value of $530 million “today.”
I ran the same property through ValueXpose. Loading the due diligence took about 10 minutes. The complete analysis, with every waypoint value along the development timeline and the financial model behind each one, took less than two minutes.
The two results tell very different stories. The difference isn’t speed. It’s methodology.
What a mixed-use conversion appraisal requires
The Flatiron conversion is mixed use: a condominium subdivision above an income-producing retail property on the ground floor. After 43 years of valuing development projects, including in litigation, here is what I’d expect a complete analysis to include:
- Two value models. The “as if complete” bulk value of the condos, and the “as if complete” stabilized and non-stabilized market values of the retail.
- A financial feasibility test for the whole development. If the condo bulk value plus the retail’s non-stabilized value exceeds the cost of production, the project is feasible.
- A property residual of the “as is” market value, rather than accepting the purchase price as given.
- An allocation of the “as is” value between the land as if vacant and the contributory value of the existing building improvements.
- Stage-specific risk rates for each phase of development, quantifying the waypoint value at each step.
- Entrepreneurial profit allocated across the phases, not taken as one lump sum.
What the AI agent produced
The agent started from the gross sell-out and worked down to a single “today” value in one step.
Step
Amount
36 condos at asking price
$739M
Units close ~5% below asking
−$37M
Selling costs, 6% of closing prices
−$42M
Remaining construction
−$62M
Ground-floor retail
+$22M
Up to two years of absorption at 12%
−$90M
Agent’s value “today”
$530M
Separately, it estimated profit as $739M less $44M of transaction costs, the $161M 2023 purchase price, and ~$183M of conversion cost at $1,000 per square foot. That leaves ~$350M.
The arithmetic is fine. Measured against the list above, here is what’s missing:
- The retail was valued as stabilized. At completion, the retail isn’t leased up yet. The feasibility test calls for its “as if complete” non-stabilized market value.
- No feasibility test. The only check was a back-of-the-napkin sum of costs. Value was never compared with a full cost of production.
- No property residual. The agent simply assumed the $161M purchase price was the “as is” market value.
- No allocation between the land as if vacant and the contributory value of the existing building.
- One flat 12% rate for every risk, instead of a rate matched to each phase. As a result, it never quantified the waypoint values along the way.
- Entrepreneurial profit as one lump sum, rather than allocated to the entitlement, construction and sell-out phases.
What ValueXpose produced
ValueXpose builds the development timeline and values every waypoint on it: a 24-month construction phase from June 2023 and a 24-month sell-out phase to June 2027.
The land residual the AI never ran
ValueXpose automatically performs the land residual and allocates the “as is” value between land and building:
- Land value as if vacant: $82,374,653
- Contributory value of the existing building improvements: $85,795,925, the spread between the cost of production and the condo bulk value plus the retail
- Maximum “as is” value: $168,170,578
That $168.2M is what a developer could pay for the property “as is” and still meet every required return at each waypoint. The developers actually paid $161M, within that range.
The waypoint returns hold up along the way. On these inputs, an owner who sold at the bulk value would earn its share of entrepreneurial profit plus a 15% unleveraged return. A bulk buyer who then sold off the condos and the retail over the 24-month sell-out would earn its share plus a 12% unleveraged return, provided the bulk price is fixed as of the appraisal date and paid at delivery.
The AI agent never asked whether $161M was the right price. It simply started from it.
I asked an AI whether it would have caught these mistakes
While working through this analysis with Claude, I put the question to it directly:
If a Claude user with no extensive valuation domain experience builds an AI CRE valuation model, will the AI agent have flaws, or does AI guide him in the best valuation practices?
Here is its answer, unedited:
Both, and the flaws mostly show up when the user can’t tell the difference.
Five weaknesses of AI valuation agents
The Flatiron example shows the methodology gap. Behind it are five structural weaknesses that show up whenever a general-purpose AI agent is asked to value complex real estate:
- Time. This agent took 89 minutes for one scenario. Depending on how an agent is built, every alternative scenario can mean another lengthy run.
- No saved scenarios. Revisiting an earlier scenario means going back and forth with the agent, often rebuilding it from scratch.
- Results that change. AI is probabilistic. Run the same scenario again and it will likely give you a different answer than it did the first time.
- It inherits the user’s experience gaps. If the user doesn’t have deep valuation domain experience, the AI reflects that weakness, as its own answer above explains.
- Token costs. Every valuation and every scenario consumes paid AI compute, and complex properties consume a lot of it.
Why we built ValueXpose
We built ValueXpose to solve these weaknesses, along with the back-of-the-napkin shortcuts that appraisers sometimes take on complex properties and that now find their way into AI valuation agents.
ValueXpose is a deterministic valuation operating system:
- The valuation experience is built in. The system guides the user step by step through the initial valuation and each alternative scenario, so the operator doesn’t need decades of experience to apply the right methodology.
- Every waypoint value has its own financial model, from the “as is” land value through the bulk value, the retail and the aggregate sell-out.
- Every scenario is saved and returns exactly the same results whenever it’s revisited, now or five years from now.
- No AI tokens are used to perform any valuation.
- The land residual runs automatically, with the “as is” value allocated between the land as if vacant and the contributory value of the existing improvements
The result is a system that outperforms AI valuation agents on time, cost, scope of analysis, transparency and defensibility. A trainee without years of valuation education and experience can complete an analysis as complex as the Flatiron development, because the experience is built into every step.
The money in development sits in the gap between what a property is worth today and what it will be worth tomorrow. You only find it if you value every waypoint in between.
See it on one of your own assignments
Don’t take our word for it. Pick a live appraisal from your queue, the more complex the better: a conversion, a subdivision, a mixed-use development, a property headed for litigation. We’ll run it through ValueXpose with you, live, and you’ll see every waypoint value, the feasibility test and the land residual built in minutes, then saved so you get the same answer every time you come back to it.
Schedule a time> https://calendly.com/ray-nicolas/valuexpose-demo-team
“Deterministic at the core, AI at the edges”