Picture the meeting. A return number is on the screen, a stabilized yield, say, or an exit value, and a committee member asks the only question that matters: where does that come from? In a healthy process, the analyst clicks, and the chain unspools, the exit cap assumption, the in-place NOI it was applied to, the lease lines that built the NOI, the amendment that reset one of the rents. In an unhealthy process, the analyst says the model produces it, and the room quietly downgrades its confidence in every other number on the page. The figure did not change. What changed was whether anyone could stand behind it.
This is the whole subject. As AI starts to author the analytical work that used to be hand-built in spreadsheets, the live question is not whether the output looks right. Output has always looked right; that was never the hard part. The question is whether it can be traced, line by line, back to the documents and ledgers it claims to summarize. An IC memo is not a piece of writing. It is a load-bearing argument, and every number in it is a joint that has to hold weight. Cite every cell, or do not put it in the room.
The number in the room
Institutional real estate has always run on traceability, even when it was painful to maintain. The reason a spreadsheet model is built the way it is, with an assumptions tab, a rent roll feeding the income lines, and formulas you can follow cell by cell, is that someone, eventually, asks how a number was made. The analyst can answer because the workbook is a record of the derivation: slow and fragile, but auditable. You click into the cell, read the formula, and find the precedent it points to.
The danger when AI enters this workflow is that it can sever the link between the answer and its derivation while making the answer look more polished than ever. A model produces a clean, confident memo in seconds, and if the numbers were generated rather than calculated, the memo is a beautifully formatted guess. The polish is the risk, because polish reads as rigor and the committee has no native way to tell the two apart. So the standard cannot be how the memo reads. It has to be whether each figure carries its provenance with it.
A generated explanation is not a citation
Here is the trap that catches most AI in finance. Ask a general language model for a number and a justification, and it will give you both, fluently. But the justification is generated, not derived. The model writes a sentence that sounds like a reason because sentences that sound like reasons are what it was trained to produce. It is reverse-engineering a plausible story for a figure, not exposing the computation behind it. The explanation and the number are two separate acts of generation, and nothing guarantees they agree.
This is why a chatbot bolted onto a data warehouse feels helpful and is, under audit, useless. It can tell you the DSCR is 1.28x and write a paragraph about why, but the paragraph is prose, not a calculation graph. There is no link from the 1.28x to the debt schedule and the NOI it should have divided, and no guarantee the model used those inputs at all rather than pattern-matching to a number that looked reasonable. When the auditor asks to tie the figure to the loan agreement, there is nothing to click. The explanation evaporates on contact with diligence.
EY and other large audit and advisory firms have been blunt about it: the central obstacle to using AI for anything that touches a financial control or an investment decision is the black box. An output you cannot explain, reproduce, and trace to source is an output you cannot rely on, no matter how often it happens to be right. Confidence without provenance is not a control. It is a liability waiting for the quarter it is wrong.
The point is not that language models are bad. It is that a generated explanation answers a different question than an auditor asks. The auditor does not want to know whether the number is plausible; the auditor wants to know exactly how it was made, from which inputs, under which formula, so the work can be reproduced and signed. A plausible sentence cannot do that. Only a real derivation can.
What a real calculation graph is
The alternative is to make the calculation real, separately from the language. Every figure in the memo is the output of a defined operation on named inputs, and those inputs are themselves either source documents or the outputs of earlier operations. That structure, inputs flowing through operations into outputs all the way down to a clause or a ledger line, is a calculation graph: not a story about how a number was made, but the actual record of how it was made, and you can walk it. Two properties make it trustworthy, and both have to hold.
It runs on a deterministic engine
The arithmetic, the cap-rate math, the NOI build, the waterfall, the DSCR, runs in a deterministic engine, not in a language model's head. The same inputs always produce the same output. Run it today, run it next quarter, run it on the auditor's machine: identical result. Determinism is what makes a figure reproducible, and reproducibility is the floor for anything an auditor will sign. A number you cannot re-derive on demand is one you cannot defend, and a language model, by design, does not re-derive; it re-generates, and the result can drift.
Every leaf links to a source
Follow any branch of the graph down far enough and you reach a leaf, and every leaf is a citation: a clause in the lease, a line in the rent roll, a posting in the general ledger, a row in the debt schedule. The exit value cites the cap assumption and the NOI; the NOI cites the income lines; each income line cites the lease and the specific amendment that set the current rent. There is no point where the trail goes cold and you take the system's word for it. The model's job is to read documents, extract the inputs, and assemble the graph, the part it is genuinely good at, then hand the arithmetic to the engine and the citations to the record.
Exit value = NOI / cap. Each input is itself a node, not an assumption typed by hand.amendment that reset the current rent.GL posting. The trail never goes cold.Defensible to an IC, an auditor, and an LP
Three audiences will eventually interrogate the same number, in three different registers. Cell-level citation answers all three with one structure: you do not maintain three versions of the truth.
The investment committee
The IC is testing the argument under time pressure. When a member challenges the exit assumption, the analyst does not defend it rhetorically; the analyst opens the node, shows the comparable set it was drawn from and the NOI it was applied to, with the lease lines underneath. The debate moves to the assumption itself, where judgment belongs, instead of stalling on whether anyone can even reconstruct the math. A cited memo turns committee time into decision time rather than forensics.
The auditor
The auditor is testing reproducibility and control. A deterministic engine with sourced leaves is exactly the artifact an audit wants: a figure that re-derives identically and ties, step by step, to the underlying documents and ledger. The black box that EY warns about is gone, because there is no box. There is a graph, and the auditor can walk it. What used to be a sampling exercise across a fragile workbook becomes a deterministic re-run.
The limited partner
The LP is testing trust over time. When a reported return or a valuation can be tied back, on demand, to in-place leases and actual ledger entries, the relationship runs on evidence instead of assurance. The LP that can see how a number was made, rather than being asked to believe a PDF, extends a different kind of confidence, and it compounds across every reporting cycle and every raise.
What changes when every figure links back
Adopt cell-level citation on a deterministic engine and the texture of the work changes, not just its speed.
- Diligence stops being archaeology. Tying a memo back to source used to mean reconstructing a colleague's workbook by hand. Now provenance is attached to every figure, so verification is a click, not a project.
- Errors surface at the source. When a number looks wrong, you walk the graph to the leaf and find the bad input, a misread amendment, a mis-keyed ledger line, instead of arguing about the output. The fix lands where the mistake actually lives.
- AI becomes usable for work that matters. Because the arithmetic is deterministic and every input is sourced, the model can author memos that touch real decisions, not just low-stakes drafts.
- Trust scales past the people who built the model. The memo no longer depends on the one analyst who remembers how the workbook fits together. The derivation is in the artifact, so anyone can audit it and the key-person risk is gone.
The discipline is old; only the stakes are new. Real estate has always required that a number be traceable to its source, because someone always, eventually, asks how it was made. AI does not change that requirement; it raises it, because AI can produce far more numbers, far faster, and far more convincingly than any analyst, and convincing is not the same as correct. Built AI is built so that every figure in a memo links back to the clause, line, or GL entry it came from, with the arithmetic on a deterministic engine and the citations attached by construction, so the output is defensible to an IC, an auditor, and an LP without a forensic re-derivation. To see a cited memo built from your own documents, explore how the platform works or book a walkthrough.