What Is an AI Agent in CRE Underwriting?
Three different things are being sold as AI agents right now, and only one of them is one.
An AI agent in commercial real estate underwriting is a language model working toward a goal with tools, choosing its next step inside limits you set, and stopping at a gate where a person approves. It is not a tool, which is code that does one job the same way every time with no model involved. It is not a workflow, which is a fixed sequence of steps in a fixed order. And it is not a chat window, which answers what you paste into it, has no deal record, no gates and no audit, and forgets the last deal when you start the next one. Those four things get sold under one word, and the word that gets used is agent.
What is the difference between a tool, a workflow and an agent?
A tool does one job the same way every time, a workflow runs several steps in a fixed order, and an agent chooses its own next step toward a goal inside limits you set. Only the third one involves a model deciding anything.
One word has to be clarified first, because in commercial real estate it already means something else. A model, everywhere in this post, is a large language model: software trained on an enormous amount of text, which produces language in response to language, and which can be handed tools and asked to use them. It is not your underwriting model. Below, the spreadsheet is always called the underwriting model, and the word model on its own always means the software.
| Term | What it is | Who decides the next step | On a deal |
|---|---|---|---|
| Tool | Code that does one job the same way every time, no model involved | The code, always | The rent roll normalizer, the folder build, the demographics pull |
| Workflow | A fixed sequence of steps in a fixed order, some of which call a model | The sequence, written in advance | Set up, screen, underwrite, memo, in that order |
| Agent | A model working toward a goal with tools, stopping at a gate where a person approves | The model, inside limits you set | Reading a package and producing extractions, each with the page it came from |
| Chat window | A general assistant answering what you paste into it | You, by hand, every time | Whatever you remembered to paste |
This is not made up vocabulary. The same split runs through the engineering world: Anthropic's Building Effective Agents (December 19, 2024) defines workflows as systems where models and tools are orchestrated through predefined code paths, and agents as systems where models dynamically direct their own processes and tool usage.
The distinction is worth defending because the word is doing sales work. A large share of what gets marketed as an agent is a tool with a chat box on it, and a rent roll processor is not made smarter by being called one. Most of the work in underwriting does not need a model at all, which is a separate argument I have already made: deterministic work belongs in code, where it is cheaper, faster to check, and wrong in ways you can see.
Why is an AI agent different from a chat window?
An agent runs inside a system that has a deal record, gates and an audit trail; a chat window on its own has none of the three. That is the whole difference, and it is structural rather than a matter of which model is behind the glass.
Four things the chat window does not have:
- A deal record. Out of the box it does not know which deal this is, and it will not know tomorrow. Every session starts from nothing, so every session starts from whatever you remembered to paste. This is the one you can buy your way out of, and the next section is about what happens when you do.
- A gate. Nothing stops output from becoming an input to the next step. There is no point in the process where a person is required to approve before the work moves.
- An audit. You get a number in a sentence. You do not get the page it came from, so verifying it means doing the work again.
- A destination. The answer lands in a scroll buffer. It does not land in a database, which means the second deal learns nothing from the first.
A chat window can sit inside a system. It is not the system.
What if our chat window is connected to the deal database?
Then you have started building the system, and connecting the record is the first of those four things rather than all of them. This is the strongest version of the objection and it deserves a straight answer: a chat window with your deals behind it genuinely does have a deal record, and it is better than one without.
What it still does not have is the other three. Nothing decides on its own to read the next document, so a person is the scheduler. Nothing stops an answer from becoming the input to the next step, so there is no gate, only good intentions. And unless somebody built the citation trail, you are still handed a number in a sentence with no page behind it.
So the dividing line is not the chat box, and never was. An interface is not a system. Ask what happens between the question and the answer: who chose the steps, what could not proceed without a person, and where the output went when the window closed. A product that answers all three is a system with a chat window on the front, which is a good thing to own. A product that answers none of them is a chat window, whatever the pricing page calls it.
What does an AI agent actually do on a deal?
On a live file the work splits into four workflows, and inside them some steps are tools and some are agents. The labels below are the ones used on each step, not a general claim about the category.
| Workflow | What it produces | Agents in it | Tools in it |
|---|---|---|---|
| Deal Set Up | A provisioned deal room, a checklist and one email address per deal | The intake filer, which reads a document and proposes its type | The filing cabinet, the deal mailroom |
| Diligence Screen | A screen memo with the red flags on top | Package reader, key parties, market read | The rent roll and financial normalizer |
| Underwriting Draft | A populated underwriting model on your template, every output traceable | Assumption basis, model populator, model audit | The model wiring |
| Memo Draft | An IC memo or OM that ties out to a locked underwriting model | Memo builder, memo audit | Delivery of the finished PDF |
Read the tool column. Half of the useful work in that table is code with no model in it, and calling any of it an agent would be a lie to the one reader who can check.
Each workflow ends at a person. The filing is confirmed before anything moves. Every extraction is approved before it counts. The team locks the underwriting model version. The memo is a draft until someone says it is not. An agent that cannot be stopped at a gate is not a system, it is an exposure.
How do you check an AI agent's work?
You check an agent's work by being able to state what the output should be before you see it. If you cannot, you are not reviewing the work, you are accepting it.
I underwrote 150 or more deals on a multifamily acquisitions team at a public REIT between June 2021 and February 2025, across ground-up, lease-up, distressed, value-add and recapitalization. That is where the test comes from. When a rent roll comes back parsed, I already know roughly what the tenant count should be, roughly where in-place rent should land, and which units will be missing a lease end date. The output gets checked against something that existed before it.
The failure mode is quiet. An agent does not return an error when it misunderstands your data. It returns a confident, well formatted, plausible number. On a deal, plausible is the most expensive kind of wrong, because it survives the review that would have caught something obviously broken. This is why every figure carries the page it came from: a citation is not a courtesy, it is what makes a wrong number findable.
What makes an AI system worth keeping?
A system is worth keeping when someone other than the person who prompted it can open it, understand it, and change it later. That is a structural property, and it is decided before any code exists.
Three things carry it. The data lives in one place with one definition per field, so two reports cannot disagree about what occupancy means. The logic sits where it can be inspected, not buried inside a prompt nobody saved. And every output ties back to a source you can name out loud in an investment committee.
None of those are AI questions. They are the same questions I was answering when I migrated a team's underwriting records into SQL and wired an Excel module so anyone could push an underwriting model straight into the database. The work that decides whether a system survives happens before the building starts, and it is still done by the person who understands the deal.
What is built today, and what is not?
Rent roll normalization runs in code today. T-12 normalization and offering memorandum extraction are in development, and saying so is cheaper than being caught claiming them. No agent produces an underwrite on its own: the assumptions are drafted with a basis and a source, a person signs off, and only then is the underwriting model populated through wired named ranges.
The four workflows above, and the deal database underneath them, are what the AI Deal Workspace is. If the definitions in this post are the ones you have been looking for, that page is where the rest of it lives.