AI Agents for Mortgage Lenders and Servicers: What a New Industry Survey Reveals About the Scaling Gap

Mortgage lending has never been short on paperwork, and it has never been short on regulators watching how that paperwork gets handled. So when a joint survey from the American Association of Residential Mortgage Regulators (AARMR), the Mortgage Bankers Association (MBA), and Boston Consulting Group asked 31 lenders and servicers — representing about 40% of the entire U.S. mortgage market — how their AI programs were actually going, the answer was blunt: adoption is everywhere, but scale is rare.

The survey, conducted between April and July 2026 and reported by HousingWire, assessed 38 distinct AI use cases spanning marketing and sales, origination, secondary and capital markets, servicing, and corporate functions. Nearly every respondent had at least one AI use case in production. Only about a quarter had fully scaled even one.

Where the AI actually is

The pattern in the data is consistent: AI is doing real work, but mostly work that helps a human go faster, not work that changes the underlying process.

• 87% of respondents had employee productivity tools (writing and summarization) in production — the single most common use case by far.
• 65% had code generation and developer support running.
• 61% were using AI for data extraction from documents.
• 54% had agent-assistance and knowledge-search tools live.
• 52% were using AI for investor guideline and eligibility extraction, and document classification/summarization.

Within origination specifically, data extraction and investor guideline extraction were the most mature use cases — but even there, only 42% of respondents had reached scaled production on either one. Higher-stakes applications like underwriting decision support, fraud detection, and credit risk analytics lagged well behind. Secondary and capital markets were described in the survey as "largely untapped," and sales/distribution analytics hadn't reached production at any respondent.

Servicing — an area with obvious agentic upside, given how much of the job is monitoring, triaging, and following up — was flagged as having "significant room for expansion," particularly in early default warnings, collections prioritization, and loss-mitigation decision support.

Why lenders are stuck at "in production," not "at scale"

The survey asked directly what's blocking broader rollout, and the answers point less at the technology and more at governance and proof:

• 59% cited regulatory and compliance uncertainty as the top barrier to scaling.
• 45% cited unclear return on investment.
• 24% each cited data quality and the availability of suitable solutions.
• 21% cited concerns about AI reliability and hallucinations.

That regulatory number matters more in mortgage than in almost any other vertical this blog has covered. The survey specifically points to Federal Housing Finance Agency guidance for Fannie Mae and Freddie Mac around AI risk management and transparency, plus voluntary NIST frameworks — a compliance surface that gets more complicated with every automated decision that touches a borrower's file.

Governance itself is also uneven. 87% of respondents have written AI policies and 84% have privacy controls, but only 58% do ongoing monitoring for model drift and accuracy, and just 45% report AI performance to their boards regularly. Perhaps most telling: more than a quarter of respondents admitted employees are using AI outside company-approved tools and processes — 20% said this happens occasionally, 7% said frequently.

What this means for lending and servicing teams going forward

The AARMR/MBA/BCG data lines up with a pattern this blog keeps finding across regulated industries: the jump from "we have an AI pilot" to "we trust this enough to run it unsupervised at scale" is where almost everyone stalls, and the stall point is rarely the model itself. It's whether the organization can prove — to a regulator, a board, or its own risk team — exactly what the AI did, why, and under what boundaries.

For mortgage specifically, roughly 80% of respondents still plan to increase AI investment over the next year, so the appetite isn't the problem. The gap is between deploying a tool and being able to answer, with an audit trail, "what did this do, and can we show our work." Lenders and servicers that close that gap first — with clear approval boundaries, logging, and human review built in from day one rather than bolted on after a regulator asks — are the ones positioned to move past productivity gains and into the underwriting, servicing, and risk workflows where the real return sits.