How CRE Firms Are Measuring the Return on AI Investment

Nearly every commercial real estate firm has run an AI pilot. Almost none can say what it returned in dollars. JLL's 2025 Global Real Estate Technology Survey found that 92% of CRE firms have piloted AI in some form, yet only 5% report hitting all their stated AI goals. That gap is the story, and most of the industry keeps telling the wrong version of it.

Deloitte's 2026 CRE Outlook adds a strange wrinkle: the share of executives calling AI's impact "transformative" fell from roughly 12% to just 1% year over year. Read that as a verdict on the technology and you'll conclude AI is failing commercial real estate. That reading is backwards. Most firms pointed AI at the wrong layer of the business: drafting emails, summarizing meetings, writing marketing copy. None of that touches the document-heavy, model-intensive work where financial impact can actually be counted. The technology isn't underperforming. It's aimed at the wrong target, and the firms treating this as a tooling problem instead of a workflow problem are the ones dragging Deloitte's numbers down.

IBM's Think Circle report found that only 29% of executives feel sure they can measure AI ROI, even though 79% report productivity gains. Value shows up. It just doesn't land anywhere a CFO can point to on a spreadsheet. Firms know pilots launched, tools got bought, hours got logged against some new system. What they can't say is whether deal velocity moved, or cost per loan dropped, or revenue per analyst changed. The firms that can answer those questions do a handful of specific things differently from the ones that can't, and the difference has almost nothing to do with which vendor they picked.

How large the AI investment is becoming, and why the stakes of measuring it are rising

Annual AI spend for a typical CRE firm runs $25,000 to $75,000 across tools, training, and integration. Smaller CRE firms pulling $100,000 to $400,000 in measurable value do so out of that same range of spending. Same budget, wildly different outcome. The gap isn't about money. It's about method, full stop.

The downside isn't hypothetical either. Translate wasted spend into CRE terms and it turns concrete fast: $50,000 of wasted annual AI spend, capitalized at a 6.0% cap rate, erases roughly $833,000 of asset value. That's not an abstraction sitting in a slide deck. That's a number a lender sees on a balance sheet, at exactly the wrong moment in a deal.

Gartner reports that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before. The firms doing the abandoning are, overwhelmingly, the ones that never built a way to check whether the initiative was working in the first place. They're pulling back exactly when the technology is turning into table stakes rather than a novelty. Firms quitting now are quitting right before the phase where structured programs tend to report their clearest returns.

What separates the firms reporting 300% to 500% returns from those reporting nothing

A February 2026 analysis of CRE firms with structured AI implementation programs found returns of 300% to 500% within the first 12 months. Firms that bought tools without a plan around them reported minimal or negative returns on the same category of spend. The variable was never which vendor got the purchase order. It came down to how the deployment got designed before anyone even logged in, and that design work is the part most firms skip.

Three things separate the two groups, consistently. High-return firms tied AI to named workflows that already had time-and-cost benchmarks attached, so there was something concrete to measure improvement against. They tracked outcomes: deal throughput, error rates, time-to-close, instead of tracking activity like logins or query counts. And they sequenced the rollout by workflow density and error cost, starting wherever volume ran highest and mistakes cost the most to fix.

An IBM CEO study found that only about a quarter of AI initiatives deliver the return their sponsors expected, and only a small fraction ever move beyond experimental status. Most sit stuck in experimental mode indefinitely. PwC's 2026 CEO Survey sharpens the point: CEOs reporting gains on both cost and revenue run two to three times more likely to have built AI into products, demand generation, and strategic decisions, not just tested it in one corner of the business. For CRE, the operative question was never "did we buy AI?" It's whether a workflow got rebuilt around it, and whether anyone wrote down what changed before and after. Firms that skip that second step are the ones showing up with nothing to report a year later.

Why underwriting and financial spreading produce the most measurable AI returns in CRE

Underwriting automation keeps showing up as the highest-return AI use case in commercial real estate, and the reason is almost mechanical: it has a clean before-and-after baseline. Hours per deal, errors per analysis, time-to-decision, all three existed as tracked numbers long before AI entered the picture, so any improvement is easy to isolate and hard to fake.

Picture the workflow without it. An analyst pulls data by hand from operating statements, rent rolls, T12s, appraisals, and offering memoranda, then keys it line by line into an internal underwriting template. It's slow, and worse, it introduces inconsistency right when lenders need clean, reliable inputs the most. Under a traditional process, underwriting a single deal can stretch across multiple weeks.

Freddie Mac's May 2025 analysis offers the clearest validation available: lenders who maximize AI-enabled underwriting achieve up to 40% cost savings in loan processing, with a materially shorter production cycle per loan. Deployments of multiagent AI systems in banking have reported analyst productivity gains and faster credit decisions, though reported ranges vary widely across implementations. The analyst's job doesn't disappear, it shifts: less time keying numbers, more time handling exceptions and checking what the system flagged. Document tools that pull data from borrower financials and spit out standardized spreadsheets in minutes instead of days are showing meaningful cuts in analyst time per deal among lenders who've actually deployed them.

Underwriting works so well as an ROI case study for a plain reason: volume is high, every deal needs it done, and there's no dodging that. Accuracy benchmarks are clear too, since structured data extraction has a known error rate to beat. And the financial link is direct: time saved per deal, times deal volume, times analyst cost, is a number anyone can work out on the back of an envelope. Fewer errors carry their own downstream value: covenant breaches caught earlier, fewer re-underwriting cycles down the road.

How AI agents extend measurable ROI beyond underwriting into deal sourcing and document review

An AI agent isn't the same thing as a single-purpose tool. Agents reason across steps, plan a sequence of actions, and run a multi-step workflow without needing new instructions at every turn, something closer to a diligent junior analyst who moves between data sources, templates, and decision criteria on their own.

Three workflow categories are producing documented returns right now. In deal sourcing, agents scan listing platforms and public records, apply a firm's investment criteria, run preliminary underwriting, and surface top opportunities with summary reports attached, turning research that used to eat up days into a task measured in hours. In document review, every transaction throws off hundreds of pages: purchase agreements, ground leases, construction contracts, title reports. AI pulls out key terms, flags clauses that don't match the standard form, cross-references dates across documents, and populates a deal tracking system on its own. In cash flow modeling, agents read rent rolls, leases, and operating statements, pull the relevant inputs, and build a full cash flow model directly inside a firm's existing template.

The gap between experimenting with agents and actually running them at scale is wide, and that gap is where most of the industry sits right now. McKinsey found 62% of organizations are at least experimenting with AI agents, but only 23% are scaling one anywhere in the business. Gartner forecasts that 40% of enterprise applications will embed task-specific AI agents by the end of 2026. That's a narrow window, measured in months rather than years, for firms trying to move first.

The firms getting this right don't automate everything at once, and that restraint is the point, not a limitation. They start with document extraction, high volume, clear accuracy benchmarks, savings that show up right away, then add deal screening, then layer in market analysis and pro forma automation once the team has checked the AI's output against its own judgment. Each workflow added is a new line item to measure. Firms tracking returns systematically across multiple workflows are the ones most likely to land in that 300% to 500% range from earlier. And the training data matters here too: an agent built specifically around CRE document structures, rent rolls, estoppels, DSCR models, throws off fewer exceptions and fewer re-review cycles than a generic large language model pointed at the same paperwork.

The capital markets context that makes faster, more accurate deal processing directly valuable right now

The Mortgage Bankers Association projects total commercial mortgage origination will reach $806 billion in 2026, up from $633.7 billion in 2025. Volume is climbing at the same time a significant wave of loan maturities in 2024 rolls into a continued wave of maturities in 2025. That's a lot of refinancing hitting underwriting desks that were never built to absorb it.

For lenders staring down that pipeline, the AI case makes itself: whoever processes loans faster captures a disproportionate share of the origination volume moving through the market. The opening is largest for community and regional lenders competing for the same borrowers as the largest banks but without the in-house engineering teams to build custom AI tooling from scratch. That's the segment most likely to get squeezed if it waits.

Sponsors and brokers run a parallel version of the same argument. Faster lender decisions mean better deal execution, so a loan package built to cut a lender's review time shortens the whole path to closing. Adoption isn't even across asset classes, though. Industry research has found student housing leading in enterprise-wide AI deployment, while multifamily, the largest and most operationally intensive asset class in commercial real estate, lags in enterprise-wide AI adoption. That's a lot of unrealized upside sitting in plain sight, in the one asset class with the most units to underwrite.

Industry research has found that only a small share of North American banks have deployed any generative AI use case at all. Whoever moves first in underwriting automation is competing against a market that, on the whole, is still doing this by hand, and that's an advantage that won't stay open forever.

How to build an ROI measurement framework that CRE professionals can actually use

The formula itself isn't complicated: ROI equals total value generated minus total AI investment, divided by total AI investment, times 100. The hard part is defining "total value generated" honestly in a CRE context, and that's where most measurement efforts fall apart before they start.

Three categories matter here. Direct labor savings: hours saved per deal, times the loaded hourly cost of the professional doing the work, times deal volume. This is the easiest number to work out, and the one most firms start with. Error reduction value: the financial hit of fewer mistakes in pro formas, rent roll analyses, and market comps, calculable once error rates before and after deployment are actually being tracked. Revenue acceleration: deals closed faster, or deals reviewed that would've been passed over otherwise due to capacity limits. This one needs pipeline data to work out, but it usually captures the largest dollar impact of the three, and it's also the one firms most often skip measuring.

The metrics gap between the two groups is stark. Underperforming firms track tool adoption rates, queries per week, hours of AI use: numbers that describe activity but say nothing about deal economics. Outperforming firms track metrics that tie directly to revenue or cost — things like time-to-decision per loan, analysis hours per deal, and deals reviewed per analyst per quarter — rather than activity counts alone. Firms that skip setting a baseline before deployment can't show ROI later, even if the AI is genuinely working, simply because there's no pre-AI number left to compare against.

A useful line can be drawn between hard ROI and soft ROI. Hard ROI covers labor cost cuts and rework savings from fewer errors, numbers a CFO can put in a budget line. Soft ROI covers gains like decision quality and team capacity: both real, but harder to put a dollar figure on, and far less persuasive when the time comes to justify next year's spend to leadership. Timing matters too. Efficiency gains tend to show up within 6 to 18 months, but the bigger financial impact, actual deal volume expansion, tends to take 18 to 36 months to show, and competitive effects can take 3 to 5 years to fully play out. Firms that pull the plug early miss the exact phase where the returns compound. One implementation detail deserves its own mention here: firms that skip structured onboarding tend to see adoption rates fall off well before returns can materialize, which undermines ROI no matter how good the underlying tool actually is.

What a CRE-specific AI platform offers that generic tools cannot measure or deliver

None of this framework works if the AI itself spits out unreliable answers, and reliability in CRE workflows comes down to whether the underlying model was trained on the document types and data structures this industry actually uses. Point a generic AI model at a rent roll or an estoppel certificate and it produces more exceptions, more re-review cycles, and more time spent by a human fixing its work. Every one of those eats into the labor savings the whole ROI calculation depends on.

Purpose-built CRE platforms change that picture in three specific ways. They ground their outputs in real-time, CRE-specific data, market comps, lender terms, deal structures, instead of generic web data that still needs a human to check it line by line, platforms such as Lev, a CRE-specific AI and data platform that gives sponsors and brokers purpose-built agents and real-time deal data in a single unified environment, are built around exactly this kind of domain grounding. They run workflow-specific agents trained on the documents that actually show up in CRE transactions: rent rolls, T12s, estoppels, offering memoranda, loan packages. And they connect data, agents, and deal workflows on one platform, so returns compound across the life of a deal instead of getting stranded inside some standalone tool that never talks to the rest of the stack.

Judging any platform against these criteria comes down to three questions. Does its output hold up against CRE-specific benchmarks? Does it plug into the deal workflows a firm already runs? And does it generate data that feeds directly into the ROI metrics that firm is already tracking? Ninety-two percent of CRE firms have piloted AI. Five percent report hitting every goal they set for it. The firms closing that gap aren't the ones with the biggest technology budgets. They're the ones that pointed a purpose-built tool at a specific, high-value workflow and measured, honestly and consistently, what came out the other side.