AI in CRE lending, what is real.

Where AI is real, and where it is noise

AI is changing commercial real estate lending in three specific places, and being oversold in most of the rest. The real changes are unglamorous: reading documents and extracting their data, drafting and summarizing routine text, and answering questions about a deal's status without a human assembling the answer. Everything downstream of those, faster reviews, fewer status emails, is a consequence, not a separate miracle.

The noise is the promise that AI will make credit decisions or replace judgment. In institutional lending, where the decision is scrutinized by examiners and the file has to prove independent analysis, that is neither imminent nor wanted. The useful frame is narrow: AI is very good at handling the document and status labor that surrounds the decision, and that labor is most of the work.

The three changes that are actually landing

Each of these is in production somewhere today, and each targets a real cost, not a demo:

• Document reading and extraction

  Pulling the numbers off a rent roll or an operating statement, and flagging what is missing or mismatched, instead of a person keying it.
• Drafting and summarizing

  Turning a deal's state into a status note, a summary, or a first-draft communication, so the analyst edits instead of composes.
• Status and retrieval

  Answering where does this deal stand and which document is the current one directly from the record, rather than by interrupting the person who knows.

The precondition nobody markets: structured data

Here is the part left out of most AI pitches. All three of those uses require something to read, and in most lending operations the something does not exist in usable form. The deal lives across inboxes, a shared drive, and a spreadsheet, and an assistant pointed at that mess produces confident nonsense, because it is summarizing chaos.

AI is only as good as the record it reads. A model that can answer where a deal stands needs a deal whose status is actually recorded, against a checklist, with owners and timestamps, in one place. This inverts the usual advice. The first move toward AI-forward lending is not buying an AI tool; it is putting the deal on a structured, current record, because that record is the substrate the AI needs. Lenders who did that for operational reasons find they are the ones who can adopt AI at all.

That is also why the connection matters more than the model. Prodeal exposes a loan's documents and status through the Model Context Protocol, the open standard, so an assistant like Claude works from the organized record rather than a summary someone pasted in. The AI is a layer on a clean substrate. The substrate is the part that takes work, and the part that determines whether any of the AI is real for you.