Cap Orbit vs ChatGPT: What a CRE Underwrite Actually Requires

Credit where due

What ChatGPT does well on a deal team.

ChatGPT is the default assistant of the working world, most of your analysts already know it, and the Business plan prices at $20 a seat per month on annual billing with a two-seat minimum, cheap enough that everyone in the firm can have one. For drafts, summaries, quick questions, and the long tail of work that is not the deal in front of you, it is the right tool and the easy buy.

The 2026 product is broader than chat. ChatGPT for Excel builds, updates, and explains multi-tab spreadsheets from plain-language instructions. Deep Research runs long research jobs. Workspace Agents, launched in April 2026, automate recurring work across connected tools. On data, OpenAI commits by contract not to train on Business or Enterprise workspace data, holds the standard certifications, and on Enterprise offers data residency across ten regions and customer-held encryption keys.

If the question is whether your firm should have ChatGPT somewhere, the answer is probably yes. The question this page answers is narrower: what happens when the deal lands on it.

Where it breaks

A confident figure is not a traced figure.

Then broker materials land, and the tool meets a different standard. Third-party evaluators put it plainly: asking ChatGPT to build a real estate model produces something no IC committee would accept. The Excel add-in does not support pivot tables, VBA macros, or Power Query, and it carries no house conventions for tab structure or formula discipline.

Extraction is the same story. Evaluators report low accuracy pulling financial data out of complex CRE documents, rent rolls and T-12s included, with significant manual correction before the numbers can be trusted. And the figures that do come back arrive confident, with no trace from a number in the output to the cell or page it came from. On a deal team an untraced figure is not an answer; it is a task, because someone now has to go find out whether it is true.

There is also no deal in ChatGPT, only conversations. No stages from screening to closing, no committee record, no tracking of actuals against the underwrite after the wire goes out. A team that builds all of that out of Workspace Agents is constructing a deal platform by hand, one prompt at a time.

Where Cap Orbit wins

The model ties out, and the record carries.

Cap Orbit is not a document chat layer; the terminal has the run of the deal file. It reads across every file at once, the offering memo, the rent roll buried in a workbook tab, the T-12, the loan agreement, the firm’s own templates, in whatever format they arrived, and one instruction can read the documents, normalize the statement, build the model, and stage the memo, with the analyst approving each consequential step. The rent roll comes out unit by unit, each figure traced to the exact file, sheet, and row or page it came from and footed to the document’s own stated totals. The model is a real workbook with live formulas, built to the institutional standard for the asset class, recalculated and checked before delivery, with Base, Upside, and Downside priced off one switch, and nothing writes into it until the analyst accepts.

The memo reads the model instead of improvising near it. Point Cap Orbit at the firm’s filed memos and it drafts in that format and voice, every figure pulled from the model’s computed cells or footnoted to a cited document, and for IC and credit memos the outline is approved section by section before a word is written. A genuinely missing number stays a flagged blank for the analyst to fill.

And the work does not end at the signature. Closing reconciles the settlement statement against the contract, the loan, and the underwrite, flags every variance with its cause, and writes the trued-up going-in basis back into the model. Asset management then closes each period against the budget and the original underwrite on a record that only adds and never overwrites, so the deal you manage stays the deal you approved.

Memory and isolation

Starting from zero, or starting from the deal.

ChatGPT starts from zero knowledge of your deal each session unless someone builds and maintains that context by hand. Workspace Agents can carry some of it, but the carrying is your team’s work, and nothing in OpenAI’s public materials describes a boundary between one deal’s documents and another’s inside a workspace.

Cap Orbit treats the deal as the unit of everything. A briefing is generated per deal and carried into every future session, so the second week builds on the first instead of starting over. Sessions are deal-bound and never open under another deal. And each deal runs sealed in its own dedicated space with only its own files attached, so the team working one transaction cannot see another’s materials, inside a firm environment that is itself walled off from every other customer.