Excel Stays the System of Record

Agents should drive the approved workbook — not replace its math with AI arithmetic.

Pramod Thaz · Grafolio

On most commercial real estate and private credit desks, the underwriting model is not a draft. It is the system of record.

The cap rate lives in a cell. NOI is pulled from a named range or a lookup. Value is a formula the investment committee has seen before. Sensitivities sit in a table the acquisitions lead can explain in two minutes. When the deal moves, that workbook — not a chat transcript — is what gets filed, versioned and revisited at asset management.

So when an AI pilot offers to "calculate NOI and value for you," the right question is not whether the model can do arithmetic. It is whether anyone on the desk will trust a number that did not come from the approved workbook. Most will not. That is why replacing underwriting and asset-management workbooks with AI math fails, and why production agents treat Excel as the place where the calculation still happens.

Why the workbook won — and keeps winning

Excel is not a legacy habit. It is where three things the institution cares about already live:

Reviewed formulas. The logic that turns rent-roll inputs and a T-12 into NOI, value, DSCR or a refinance screen has been checked by people who own the outcome. Changing that logic is a model change, not a chat preference.

Version history and sign-off. Workbooks travel with the deal — IC pack, asset-management file, often the auditor's request list. A figure that appears only in an AI session has no equivalent trail.

Firm-specific conventions. One fund annualizes incomplete months; another does not. One template folds pet fees into effective rent; another keeps them separate. Those conventions already live in the sheet. An agent that invents a parallel calculation invents a parallel set of conventions.

Rebuilding all of that so the AI can "own" the math is a multi-year project dressed up as a pilot. Driving the workbook you already trust is a workflow change you can ship.

What goes wrong when AI does the math

Consider an illustrative multifamily asset — Beacon Wharf in a demo portfolio. Trailing-twelve-month NOI on the operating statement is $1,310,000. The approved underwriting workbook applies a 5.74% cap rate and produces a value of about $22.8 million, using the firm's own formula in the sheet.

An agent that recomputes value in a chat window might match today. Tomorrow it might round differently, drop a charge category, or apply a cap rate from another tab. The reviewer has no cell to open, no formula to inspect, and no way to run the sensitivity the committee expects — rent down 5%, cap rate to 6.75% — inside the model they already signed off.

That gap is why AI-math pilots stall:

Two sets of numbers. Workbook and agent disagree; committee time goes to reconciling tools instead of judging the deal.
No cell-level audit. "The model said so" is not an answer when the model is a prompt.
Silent formula drift. Workbook logic changes under change control; chat arithmetic changes with the prompt or session.
Broken handoff. The asset manager inherits the workbook, not the chat. Underwrite outside the sheet and the living model starts wrong.

The failure is not that language models are bad at division. Institutional math needs a durable, inspectable home — and on these desks, that home is still Excel.

What agents should do instead

The production pattern is straightforward in principle:

Use the approved workbook as the calculator. The agent opens the model your firm already uses. It does not rebuild NOI, value or debt screens in free text.
Enter inputs; leave formulas as formulas. Rent-roll figures, statement lines and assumption cells can be populated from documents the agent has read. The sheet's own formulas produce the outputs.
Cite cells the way you cite document lines. When the agent reports value, it should point to the cell (or named output) that calculated it — not only to a prose summary.
Run sensitivities in the workbook. Cap-rate and rent scenarios belong in the model the committee already knows how to read.
Keep model changes under human review. If a formula, structure or assumption set needs to change, a person approves it. The agent proposes; it does not silently rewrite the system of record.

That is the difference between an agent that assists underwriting and an agent that competes with it.

CRE examples that stay on the desk

NOI and cap rate. The agent reads the operating statement, places NOI in the input cell the template expects, and lets the workbook compute value. If the statement has no NOI line — as in a demo property with a missing line — the agent flags the gap. It does not estimate NOI so the sheet can still "run."

Rent-roll inputs. Unit counts, occupancy, contract rent, vacancy and ancillary charges feed the model through the cells your template already defines. Occupied units use contract rent; vacant units use market rent — or whatever your approved definition says. The agent applies the definition; the workbook does the annualization and the roll-up.

Budget versus actual. Month-end variance packs stay tied to the asset-management workbook. The agent can draft commentary and pull cited figures; it should not invent a second variance engine that disagrees with the file the portfolio manager opens every Monday.

Private credit. Borrower packs and covenant tests often live in servicing workbooks for the same reason. An agent that monitors covenants is useful; one that recalculates DSCR outside the approved model creates the same dual-number problem.

Human review of model changes

"Excel stays the system of record" does not mean the workbook is frozen forever. It means change control stays with people.

When a new fund definition of NOI is approved, the template updates and the agent follows. When a sensitivity tab is redesigned, a reviewer signs off before agents use it. When a deal needs a one-off structure, a person decides whether it joins the approved model or stays a deal-specific exception.

Agents accelerate population, citation and first-pass narrative. They should not become silent editors of institutional models.

What this means for AI pilots

If your pilot plan includes "replace the underwriting model with AI," expect resistance from acquisitions, asset management and compliance — and expect that resistance to be correct.

If the plan is "read the rent roll and T-12, drive our approved workbook, cite every input and output cell, and queue anything that does not reconcile," you are building something the desk can adopt.

The test is simple. Can a reviewer open the same file, see the same formulas, and reproduce the same number without trusting the chat? If yes, talk production. If no, you still have a demo.

Keep the sheet. Give the agent a job.

Analyst time should go to judgment: which cap rate, which rent story, which exception matters — not to rekeying rent rolls into models that already know how to calculate.

Agents earn their place when they respect the workbook as the system of record — and make it faster to populate, easier to cite and harder to contradict.

Grafolio builds agent workflows that run against the documents and Excel models investment teams already trust. Meridian by Grafolio, our AI-first CRE portfolio platform, is the reference experience: it runs existing underwriting and asset-management workbooks with formulas intact, rather than replacing them with AI arithmetic.

grafolio.ai · Meridian

Technology and process advisory. Not investment, legal or regulatory advice. Illustrative demo examples only; no client results are implied.