Why Generic AI Tools Fall Short for Commercial Real Estate Deals

Commercial real estate firms are pouring money into AI pilots and getting almost nothing back. Dig into why, and a clear pattern turns up: most of the tools being deployed were never built to read a rent roll, a capital stack, or a lender's leverage criteria, and no amount of pilot budget fixes that gap. Generic AI can summarize, draft, and explain, but it has no way to know the specific property, the specific lender relationship, or the specific deal history that makes an answer correct rather than just plausible sounding.

Why so many CRE firms are running AI pilots but getting little from them

JLL's 2025 Global Real Estate Technology Survey found that 92% of commercial real estate occupiers have started some kind of AI pilot, while only 5% say they've achieved most of what they set out to achieve. That gap deserves more than a passing mention, because most coverage stops at the number instead of asking what separates the 5% from everyone else. Sitting with that question for a while, turning over what these firms might have actually done differently, the answer starts to come into focus.

The firms in that 5% share mature data infrastructure, established workflows, and teams experienced enough to know exactly where in a process AI belongs. Everyone else bought a general-purpose tool for a specialized problem and hoped it would figure out the context on its own. A tool with no access to a firm's lease data, comp set, or deal history was never going to produce reliable underwriting, no matter how the demo looked.

Industry observers have framed the competitive question consistently: firms that stay ahead find where automation belongs in a workflow, starting with the manual, repetitive tasks that eat hours without adding proportional value. Spraying AI across a business and hoping something sticks underperforms, consistently, and the gap between targeted deployment and broad deployment explains most of the variance in the JLL numbers.

The dominant failure mode inside firms is almost comic once pictured in practice. An analyst opens a chat window, then spends fifteen minutes copying lease details, tenant names, and rent roll figures into the prompt box before asking a single question. The manual data entry the tool was supposed to eliminate becomes a prerequisite for using the tool at all, adding a layer of work on top of what already existed.

The type of AI matters far more than the fact of adopting it. That's the real lesson buried in the JLL numbers, and most firms are still missing it.

The specific ways generic tools break down across core deal workflows

Start with sourcing. A generic assistant can't scan a listing platform, apply a firm's investment criteria, or run preliminary underwriting on a property it's never seen; it only processes what gets pasted into it. The human still does the sourcing while the AI just formats the notes afterward.

Underwriting is where the gap turns expensive. Analysts spend weeks building pro forma models: combing offering memoranda, reconciling lease abstracts, compiling comps, all under deadline pressure with real error risk if a rent bump or expense ratio gets fat-fingered. A generic tool can explain what a debt service coverage ratio is in plain English, but reading an actual rent roll, pulling the right comps for that submarket, and dropping the numbers into a firm's Argus template are different tasks entirely, and a generic tool has no real path to any of them, platforms like Lev, a CRE-specific AI and data platform that connects sponsors and brokers to real-time lender data, deal history, and purpose-built agents for underwriting and document workflows, are built precisely because generic tools have no path to any of them. Financial statement processing shows the bottleneck most clearly: without extraction built for this specific task, analysts still spend 30 to 40 minutes on a single operating statement, according to JLL.

Capital markets work compounds the problem. Capital stacks today carry more moving parts than ever: senior debt, mezzanine, preferred equity, CMBS, agency execution, each with its own eligibility rules. A generic tool has no model of which lenders are active in a given market this quarter, at what leverage, for what property type, because that information changes constantly and lives nowhere near the model's training data. Ask it to recommend a capital stack structure and it hands back something that sounds informed but amounts, in practical terms, to a guess wearing finance vocabulary.

Document generation follows the same pattern. Boilerplate is easy, but a real OM or loan package only becomes useful once it's filled in with accurate property data, verified comps, and deal-specific financials, and that's exactly where generic tools stall out. At major brokerages, generic AI produces a rough first draft that still needs heavy manual correction, because it can't pull from a live CRM, deal history, or proprietary comp set. That's the current ceiling: faster first drafts, with the deal itself still moving at the old pace.

CRM and relationship work expose the same blind spot from a different angle. A generic tool doesn't know a firm's client history, its active pipeline, or which tenant's lease expires in four months. Without that context, it can't surface the right contact, flag the expiration, or suggest a lender based on how a similar deal closed last year.

What purpose-built AI can actually do differently in these same workflows

The architectural difference is straightforward once traced through each workflow above. Purpose-built AI starts connected to property data, lender data, deal history, and workflow logic, which changes what the tool can do across every workflow described earlier. A generic tool starts as a blank text box waiting for a human to feed it context by hand.

Financial spreading makes the before-and-after concrete. Per JLL, financial statement processing time dropped from 30 to 40 minutes per statement to 1 to 3 minutes once purpose-built extraction was deployed. That's roughly a 30x gain on one of underwriting's slowest, least value-added steps, and it changes how many deals a single analyst can carry at once.

Underwriting throughput tells a similar story at the institutional level. Banks using AI underwriting systems built specifically for commercial lending report time-to-decision reductions of 50% to 75%, according to research from V7 Labs. That compression comes from a system that already knows the loan criteria and checks a deal against them automatically, a different mechanism entirely from a chatbot explaining cap rates fluently.

Sourcing and due diligence benefit the same way. Purpose-built agents scan new listings daily, filter them against a firm's specific investment box, run a preliminary underwriting pass, and flag problems before a human analyst opens the file. Document generation shifts from filling in a template to producing something genuinely deal-specific: credit memos and term sheets built from the actual numbers in the actual file, with far less generic filler left for a human to clean up.

Auditability might be the most underrated gap of all, easy to miss unless both types of tools get watched handling the same request side by side. Purpose-built tools trace every figure in a credit memo back to the source document and page it came from, offering a kind of verification that generic tools have little mechanism to provide. A related distinction separates an AI assistant, which waits to be asked something, from an AI agent, which runs sourcing, underwriting, modeling, and documentation as connected steps under supervision, rather than as isolated requests typed one at a time.

Why data quality and integration determine whether any AI delivers value in CRE

Garbage in, garbage out is an old idea, but it lands harder in CRE than in most industries, because the financial stakes of a bad output are large and the underlying data is genuinely a mess. Rent rolls arrive in a dozen formats, leases carry inconsistent terms property to property, and operating statements follow different accounting conventions depending on who prepared them, while market comps come from competing sources that don't always agree.

Across CRE finance, data fragmentation and manual reconciliation are widely cited as leading barriers to operational efficiency. On its face, that's a plumbing problem. But sit with it a little longer and it becomes an AI problem: the moment a firm expects a model to produce reliable output from data scattered across five systems that don't talk to each other, the plumbing problem and the AI problem turn out to be the same problem wearing two names.

Generic AI has no real path through that mess. It can't reach the systems where the fragmented data actually lives, so it can't reconcile much of anything; it works with whatever a person happens to paste in front of it. Purpose-built AI carries the opposite risk: it's only as valuable as the data it connects to, current property data, current lender activity, current market comps, real deal history, and that data has to actually stay current.

The word "current" is doing real work in that sentence. A market comp from six months ago, or a lender's last known rate sheet, can point an underwriting decision in the wrong direction entirely once the rate environment has shifted, and rate environments shift often. The right first question to ask any AI tool is what data the tool actually knows, and how fresh that data is.

How leading CRE firms are separating AI adoption from AI results

CBRE has made AI infrastructure a core business priority, with leadership signaling that deeper integration is central to its operational strategy going forward. That kind of spending treats AI infrastructure as core to the business, well beyond a pilot line item buried in an IT budget.

Senior CRE executives have broadly said versions of the same thing: AI takes the routine and repetitive work, while the complex transaction, the client relationship, and the strategic judgment calls stay with people.

A pattern shows up across the firms actually getting results, and it takes comparing several of them side by side to see it clearly. They find specific bottlenecks first, financial spreading, document prep, lender sourcing, instead of deploying AI across the business all at once. They build or buy real data infrastructure before layering AI on top of it, because a model sitting on fragmented data just produces fragmented answers faster. And they treat AI as something that supports professional judgment.

There's a talent problem underneath this too. Plenty of firms simply lack people who know how to evaluate an AI tool against a workflow's real requirements. That's exactly why tools built for CRE from the outset lower the barrier to adoption: the domain knowledge sits inside the product already, instead of something the buyer has to supply from scratch. Firms not getting results are usually running general tools against specialized problems and quietly absorbing the cost of manual context-loading, a cost that traces straight back to buying the wrong tool in the first place.

Platforms built specifically for CRE deal workflows, covering lender search, deal matching, memo generation, and document intelligence inside one connected system rather than separate point tools bolted together, are where this results gap tends to close. That matters most for sponsors and brokers who don't have enterprise IT departments behind them.

What to look for when evaluating whether an AI tool is actually built for CRE

The evaluation question underneath all of this is simple: does the tool know commercial real estate, or does it just know language? The answer usually shows up fast, in the first deal-specific request a person actually throws at it.

Four questions surface it quickly. What data does the tool actually connect to, and is that data current or static? Can it work directly on a specific deal's documents, or does someone have to type every piece of context in by hand first? Does it understand DSCR, NOI, cap rates, debt structures, and lease terms as real financial constructs, or does it treat them as ordinary text to summarize? Does it produce output traceable back to a source document, or output that still needs a full manual check before anyone can trust it?

Breadth is usually a trap in this evaluation. A tool that claims to draft anything is almost always less useful in practice than one that does five CRE-specific tasks reliably, backed by real, current data. A set of disconnected point tools, one for lender data, another for documents, a third for CRM, just recreates the same silo problem that made generic AI unreliable in the first place. A platform that keeps lender data, deal data, documents, and CRM context in one place removes the manual handoffs that quietly erode the productivity gains everyone is chasing.

The stakes aren't abstract. The Mortgage Bankers Association projects total commercial mortgage origination volume of $806 billion for 2026, up from $633.7 billion in 2025. At that scale, firms executing deals faster and with fewer errors hold a structural advantage over competitors still copying lease terms into a chat window by hand.

The most powerful general-purpose model on the market carries real limits as a tool for commercial real estate, no matter how fluent its output sounds. Weighing that against everything above, the conclusion holds firm: the tool that knows the most about the deal sitting in front of a professional right now, with data current enough to trust, holds the stronger position. Everything else is a demo.