PhoenixTeam has engaged with about 150 different lenders, servicers, mortgage vendors, government sponsored enterprises (GSEs), and all three federal housing agencies about genAI in mortgage. Across more than 300 individual interactions, we have a broad perspective on how the industry is using genAI today.

1. Routine adoption of assisted code generation. Various tools are in use, with extensive use of GitHub Copilot. With genAI-native applications, as much as 90% of code created is generated by genAI. In mortgage broadly, we see a modest 20% efficiency gain with this use, primarily due to the heritage code bases and the significant specialization required to engage with these code bases.
2. Routine adoption of internal knowledge augmentation supportbots, using retrieval augmented generation (RAG). This is a technical approach where companies take confidential (but generally not consumer) data and enrich large language model (LLM) requests with this context, improving the quality and relevancy of results. This significantly eliminates the need to “hunt and peck” or perform traditional search functions on internal documentation. Instead, users can engage conversationally with these materials.
3. Routine adoption of Microsoft Copilot, with lackluster results. Enterprises are leaning into this strategy and are generally disappointed by adoption and results. Even Gartner has been really tough on this one. It costs about $350 a year per person for Copilot in return for about $350 of value. This is classified as a “return on employee” (ROE) rather than a return on investment (ROI) as this is typically a cost neutral use case and not at all differentiating.
4. Routine failure of genAI pilots as organizations attempt to move from proof of value (POV) to scaling across the enterprise. Many genAI pilots and implementations are failing, primarily due to a lack of focus on error analysis and the expense associated with achieving and evidencing accurate results. “Evals” – short for genAI evaluations – will soon be the new service level agreement (SLA) standard for genAI solutions, and pricing will ultimately move to outcome-based rather than seat, enterprise, or widget-based.
5. Increasing adoption of call summarization in the call center. This will be routine within the next six to 12 months. GenAI-based solutions will replace prevailing approaches like those offered by typical contact center technology companies. These companies tend to move too slow, and their “add-ons” are clunky and expensive. They will become obsolete as genAI native solutions penetrate the market.
6. Increasing adoption of genAI-based sentiment analysis in the call center. This will ultimately replace traditional technologies, which are notoriously unreliable, lag behind real-time availability, and are very expensive.
7. Emerging use of voice agents in the call center. This is getting a lot of traction, with some voice agent companies in mortgage servicing gaining more than 100 customers in the past year. These are very good technologies, with voices that are indistinguishable from humans and capable of a range of conversational and dynamic strategies.
8. Nascent use of genAI based solutions for optical character recognition (OCR). These technologies use vision models instead of traditional machine learning to classify, understand, and extract data from documents. These are virtually always more accurate but are currently too expensive to compete with offshore error resolution of traditional machine learning results. The problem of reliably getting data off documents at low cost and deriving meaning from large image and portable document format (PDF) documents continues to be a very challenging problem without consistently “easy” or reliable solutions.
9. Nascent use of genAI agents in underwriting support, automating tasks in loan origination and servicing systems. This is a very high-risk area, with many different vendors attempting to automate the underwriting process generatively. We are not aware of any company or vendor that has succeeded with any real enterprise scalable solution in this area. Many are trying. The end state of this, for now, will be mortgage loan originator support, not customer facing unassisted decisions. The same goes for underwriting loan modifications.
10. Use of truly agentic genAI solutions is extremely limited. The deployments are fragile and prone to failure. The most mature places for agentic AI are in the call center and in development assistance. This is what we see in Silicon Valley as well. There is significant investment in agentic AI. For now, however, the places where it can be completely unassisted are limited. This will be an enormous focus in the next one to two years.

Zooming out from the technical adoption, we see that many organizations have “use case lists”, an important first step. Unfortunately, the truly transformational uses of genAI are not found on these lists. They require deep mortgage subject matter expertise, a tempered approach to innovation, and significant investment by the organization. The industry continues to be very confused about the difference between AI and genAI, which is a major barrier to actually adopting these technologies in a way that is valuable. We continue to see significant implementations of point-based “fear-of-missing-out” (FOMO) solutions that fail to produce a real return.

By Tela Mathias, Chief Nerd and Mad Scientist, PhoenixTeam and CEO of Phoenix Burst