What AI Agents Actually Do Inside a Commercial Real Estate Deal

A JLL global survey found that 92% of commercial real estate teams have started piloting AI. Only 5% say they've gotten most of what they wanted out of it. That gap should bother anyone paying attention.

Deloitte's 2026 CRE outlook makes it worse, not better. The share of executives reporting a transformative impact from AI dropped to roughly 1%, down from about 12% the year before, even as pilot activity kept climbing. Enthusiasm went up. Results went down. When those two lines move in opposite directions, the industry usually tested the wrong layer of the technology, and I think that's exactly what happened here.

Most firms piloted assistive tools: something that drafts a listing description, cleans up an email, summarizes a call. Fine, as far as it goes. None of that touches the operational chain that decides how fast a deal actually closes, or how many loans a team can push through underwriting in a quarter.

The growth numbers make the stakes concrete. Firms that have adopted AI are projecting 31% portfolio growth for 2026, a margin that no drafting tool can explain on its own. No drafting tool explains a gap that wide. Friction across the whole pipeline does, from sourcing through closing through whatever happens to a loan after it's booked.

How a CRE deal actually breaks down into discrete, agent-addressable workflows

A deal moves through a chain of handoffs, and each one has its own inputs, its own decision point, its own document coming out the other end. Sourcing doesn't resemble underwriting. Underwriting doesn't resemble closing. Treating "AI for CRE" as one undifferentiated thing misses this basic fact about how the work is structured.

Four stages carry most of the weight: deal and lender sourcing, underwriting, document production, and closing through post-close monitoring. Each one used to eat analyst hours on reconciliation work that taught nobody anything. The most time-consuming part of the job was rarely analytical: it was manual transcription of figures from documents into spreadsheets, repeated until the totals tied out.

An agent built to source lenders reasons over completely different inputs than one built to spread financials, and it produces a different kind of output entirely. Judging them by the same yardstick is a bit like judging a title company and a debt broker as the same business because they both show up at the closing table.

Scale is what makes this urgent instead of academic. MBA data puts total CRE lending at $498 billion in 2024, up 16% year-over-year, with $583 billion forecast for 2025 and $806 billion projected for 2026. That curve runs straight into the manual capacity limits of origination desks. Something absorbs it, either headcount or workflow, and headcount doesn't scale nearly that fast.

What a lender-sourcing agent actually does when a broker uploads a deal

The old way of sourcing a lender runs on memory and relationships. A broker keeps a mental model, or a battered spreadsheet, of which lenders are active in which asset classes and geographies, then works the phones to test appetite deal by deal. That model decays fast. Lender credit boxes shift with rates, with balance sheet capacity, with whatever happened in last week's committee meeting. A spreadsheet updated in March is fiction by June.

An agent built for this reads lender credit boxes close to real time, matching deal parameters against current criteria through live inference rather than a lookup against a database that was accurate three months ago and has been quietly going stale since.

CommLoan's AI quote generator, launched in May 2026, shows what that looks like in practice. A broker uploads an offering memorandum or types in deal details, and the tool returns a ranked list of top lender matches in under a minute, drawing from more than 600,000 active loan programs across over 1,000 lenders, spanning 75 property types and every county in the country.

That kills the cold-start problem on a new deal. It also surfaces lenders a broker's personal network would never have reached, sitting right next to the relationships that broker already has. What it can't do is outrun bad data underneath it, because a lender-matching agent is only as sharp as the lender data feeding it. The real ceiling on performance is the quality of that feed, not the cleverness of the matching logic on top.

What underwriting agents do with financial documents — and where human judgment still sets the ceiling

Financial spreading, pulling structured numbers out of rent rolls, operating statements, and borrower financials, is consistently the most time-consuming manual task in underwriting. It's also the most mechanical, which is exactly why it was the first thing handed off to a machine.

An underwriting agent ingests documents in whatever shape they arrive: PDFs, scanned statements, handwritten margin notes off some borrower's accountant's desk, all read through OCR and computer vision. It extracts line items into a standardized schema and flags inconsistencies, a rent roll that doesn't match the operating statement, borrower-reported income that doesn't line up with market comps. Every figure links back to its source page, so the audit trail comes out of the extraction itself instead of getting bolted on afterward as a compliance chore.

The accuracy numbers hold up under scrutiny. Moody's QUIQspread reports accuracy above 95% on completed spreads, and Nedbank now runs more than 90% of borrower financials through automated spreading. Freddie Mac's May 2025 analysis puts a dollar figure on what that means: lenders getting the most from AI-enabled underwriting see up to 40% cost savings in loan processing, around $1,500 per loan, a 14% reduction, with a production cycle five days shorter. Separate case studies show time-to-decision improving 50% to 75%, with approval cycles going from weeks to days.

Still, the underwriter doesn't leave the room. An agent can flag that a rent roll doesn't match a financial statement, but whether that's fraud, a data entry error, or a legitimate quirk in how the borrower reports things still needs a person to sort out. Judging whether an assumption is aggressive for current submarket conditions takes a kind of judgment no agent has, not yet. Credit decisions with material exceptions still need someone with actual signing authority.

What the agent does is compress the gap between a document landing and a qualified person acting on it. That gap used to be measured in days of manual spreading; now it's roughly the time it takes to review a flagged output. CBRE's spread data shows why this matters beyond convenience: average commercial mortgage spreads tightened 49 basis points year-over-year in Q4 2024, down to 184 bps. When margins compress that far, faster underwriting is the difference between a loan that pencils and one that doesn't.

How agents handle the cross-document work that humans find hardest to do at scale

There's a failure mode in manual underwriting that only shows up at volume. A reviewer can check one document carefully, or scan a whole portfolio quickly, but doing both at once, with the same rigor, isn't something people are built for. Nobody reads their fortieth rent roll of the week with the same eye they brought to the first. It's a well-documented pattern: review quality degrades well before the end of a long stack.

Agents don't carry that tradeoff. They compare rent rolls against financial statements simultaneously, check underwriting assumptions against appraisal projections, weigh borrower-reported performance against market data, and hunt for patterns across an entire portfolio that would stay invisible in any single-file review. That's the part of the job that scales worst for a human team and best for a machine.

The result is a connected knowledge base instead of a stack of static files. A lender's portfolio becomes something searchable in real time, rather than an archive that only gets pulled back out at the next periodic review.

Fraud detection benefits the most, honestly. Inconsistencies that used to take hours of manual cross-referencing to catch now surface before a loan closes, not after it's funded and the problem belongs to somebody else. There's a regulatory dimension too. Source-page traceability, being able to point to the exact document and page a number came from, increasingly matters under evolving interagency guidance on AI-assisted outputs. It stopped being a nice-to-have for internal audit a while back.

For lenders, this cross-document reasoning functions as risk management as much as productivity. A firm running it end to end is operating a structurally different process from the old checkpoint-based workflow, one where cross-referencing happens continuously instead of a few times a quarter.

What document-production agents generate — and what makes a produced document trustworthy

Document production is where most CRE professionals have already brushed against generative AI, usually in its least demanding form: a content assistant that drafts prose somebody then has to go populate with real numbers by hand. An agent that pulls verified deal data into a finished, structured document is doing something else entirely.

The difference sits underneath the output. A content tool produces text. A document-production agent pulls structured data straight from the deal's underwriting layer and assembles a complete document with citations built in from the start. Aloan's credit memo product is a concrete example: it generates examiner-ready commercial loan credit memos in under 30 minutes from tax returns, financial statements, and supporting documents, with every number and ratio cited back to its exact source page.

That traceability is the actual bar for quality, not some nice-to-have feature. Lenders and equity partners verify key figures before committing capital, and a document with unsourced numbers creates friction at precisely the moment friction costs the most. Regulatory expectations increasingly require that AI-assisted outputs be auditable. An agent that can't link its numbers back to source material has produced a draft, however polished it looks on screen.

Research on AI-assisted credit memo workflows has found analyst productivity gains of 20% to 60%, and roughly 30% faster credit decision turnaround, in workflows where AI handled the credit memo. The analyst's job shifted toward review and judgment and away from manual assembly. The real test for anyone evaluating these tools is simple: can the finished document go straight to a counterparty, or does someone quietly rebuild it by hand first?

What agents do after closing — monitoring, compliance, and portfolio-level visibility

Post-close is where manual workflows break down most visibly once a portfolio gets big enough. Hundreds of rent rolls, lease expirations, covenant tests, and operating statements arrive on different schedules from different sources, and no spreadsheet handles that gracefully no matter how many tabs somebody builds into it.

An asset management agent ingests data continuously from property management, leasing, and financial systems. It flags expense anomalies, lease gaps, or underperforming tenants before they show up as a drop in NOI, generating forecasts and alerts well ahead of the quarterly report that would otherwise surface a problem that's been building quietly for months. Teams using AI-driven asset management tools have reported operating cost reductions of up to 30%, mostly from catching issues early instead of reacting once they've already gotten expensive.

For lenders, the same logic applies to covenant compliance, insurance certificates, and borrower reporting, document types an agent flags continuously rather than at scheduled review intervals. Covenant breaches don't wait politely for the next quarterly cycle to announce themselves.

The portfolio-level view is the hardest thing to replicate by hand. An agent surfaces a pattern across 200 properties at once; a human team works through those properties one at a time, and by the time they reach property 150, the pattern that mattered at property 12 has already turned into a problem nobody caught in time. Firms that run agents across the full deal lifecycle, not just at underwriting, build a knowledge base that sharpens with every deal passing through it. Firms that automate one stage get a point solution: faster there, unchanged everywhere else.

What determines whether an AI agent actually performs in a CRE context

An agent is only as good as the data it reasons over. In CRE that means real-time, verified, industry-specific data, full stop. A generic financial dataset, or a stale export nobody bothered to refresh, won't hold up under any real pressure.

That data landscape is fragmented by design, and this matters more than people give it credit for. Trepp covers CMBS and debt. CompStak covers lease comps. Reonomy covers off-market prospecting. Placer.ai covers retail foot traffic. Each vendor owns a slice, and an agent that can only see one slice is reasoning over an incomplete picture of the deal in front of it, however confident its output sounds.

There's a plumbing problem underneath the data problem, too. An agent that can't connect to a firm's existing document management, property management, or CRM systems just creates a second workflow running alongside the first one. That's often worse than no agent at all, because now somebody has to reconcile two systems instead of running one.

General-purpose AI models struggle here for a structural reason: CRE documents carry their own grammar. Rent rolls, DSCR calculations, and NOI definitions don't mean the same thing from one lender's underwriting guidelines to the next, and a model trained on generic financial text has no way of knowing that going in. The agents that actually perform in this industry are built against CRE-specific data, wired into the systems firms already run, and judged, in the end, by whether the numbers hold up when somebody actually checks the source page.