Commercial mortgage lenders are working through a heavy year. Roughly $875 billion in commercial mortgages come due in 2026 against originations forecast near $805.5 billion, which means a large volume of refinancing has to be underwritten by teams that were not staffed for it. Pressure like that tends to push technology adoption faster than it otherwise would move.

It has. A market study found 79 percent of financial institutions running AI pilots or operating AI inside lending workflows by July 2026, up from 34 percent in late 2025. That is a steeper adoption curve than online lending, mobile lending, or statistical automated underwriting ever produced.

Where the technology has earned its place is in the document work that occupies the front end of every loan. Converting an offering memorandum, a rent roll, and a trailing twelve into a populated underwriting model has always been the most expensive manual step in a lending shop, and it is exactly the kind of extraction AI handles well. Another study found institutions running purpose-built underwriting platforms reporting 40 to 60 percent reductions in analyst time per commercial loan, with application-to-decision timelines collapsing from the traditional five to ten business days down to 24 to 48 hours. Banks deploying AI underwriting have reported time-to-decision compression of 50 to 75 percent.

Portfolio surveillance has benefited similarly. Covenant monitoring, financial variance analysis, and insurance review are recurring obligations across a loan book that nobody enjoys doing manually, and automating them lets a servicing team watch more loans more closely than headcount would otherwise allow.

Trust is where the progress stops, and the distance between usage and reliance is striking. A CRE Industry Pulse Check found 66 percent of commercial real estate professionals using AI weekly or daily, while only 5 percent said they trusted it enough to inform an actual deal decision. More than half use it for support while explicitly excluding it from final decisions, and another 17 percent use it only with heavy verification.

That caution reflects a specific requirement rather than general skepticism. Institutional credit committees need to know where a number came from, and a well-formatted model that cannot be traced back to a source document fails that test regardless of how fast it was produced. Lenders have increasingly begun treating untraceable data as financial risk, which means deal teams relying on unverified outputs face longer credit reviews, heavier manual audits, and a higher cost of capital rather than a faster process.

Regulation reinforces that position. The Equal Credit Opportunity Act and the Community Reinvestment Act require lending decisions to be explainable and free of discriminatory patterns, and the OCC, FDIC, and CFPB have all signaled that AI-assisted credit decisions remain subject to those standards. A model that produces a defensible answer without a defensible explanation does not satisfy an examiner.

Underneath both problems sits data that was never organized for this. Deloitte’s 2026 Commercial Real Estate Outlook found more than 60 percent of CRE finance organizations citing data fragmentation and manual reconciliation as their primary obstacle to operational efficiency. A loan generates underwriting memos, borrower correspondence, covenant documentation, insurance reviews, and internal credit discussions across its life, and those artifacts typically live in systems that have no relationship to one another.

JLL’s 2026 Global Real Estate Investor AI Survey, released this week, found that readiness determines outcomes more than tooling does. Among investors whose data is not AI-ready, 59 percent struggle to realize the benefits they expected, against 21 percent of those whose data is ready. The survey covered 837 senior decision makers at firms with at least $1 billion in assets under management across twelve countries.

The same survey points to something lenders may find uncomfortable about speed as a strategy. Investors told JLL their top AI priorities were investment outperformance and proprietary analytics in capital deployment, and those were the areas producing the weakest results. The strongest results came from asset and property management, which respondents had ranked as a lower priority. JLL’s explanation has to do with where competition sits. When every bidder underwrites faster using similar tools against similar purchased data, nobody pulls ahead and the seller captures the benefit.

Lending works the same way. Faster time to decision is a real advantage while competitors are slower and stops being one when they catch up, which has already started to happen. What does not get competed away is a lender’s own loan performance history, its relationships, and its accumulated judgment about which borrowers and which markets have worked out.

Getting value from those assets requires more than an enterprise license, and relatively few firms have done that work. JLL found 83 percent of investors with at least one AI activity scaled or fully embedded, and 81 percent of deployments running on enterprise large language models. Only 15 percent had built custom agentic workflows across all their use cases, and one in five had built none anywhere. Broad access through a general-purpose tool does not change how work gets done.

The architecture that appears to be emerging in credit is deliberately mixed. Speaking at Ai4 2026 in Las Vegas, Pranjal Daga, Co-Founder and Chief Executive Officer of Accend, argued that a powerful language model alone is not sufficient for high-stakes credit decisions, and that the next stage of commercial lending depends less on adding individual tools than on connecting the underwriting process through a common intelligence layer. What that produces is a hybrid rather than an autonomous system, with AI agents handling selected tasks, conventional software performing exact calculations, verification systems checking outputs, and people responsible for judgment and exceptions.

Lenders building toward AI led decision making are working on the parts that take longest, which are the data connections and the audit trails rather than the models themselves. The firms that get there will be able to move quickly on a deal and explain every number in the file, which is what a credit committee has always wanted and what the current generation of general-purpose tools has not delivered on its own.

The post Lenders Have Adopted AI Everywhere But Still Don’t Trust Its Decision Making appeared first on Propmodo.

​