Engineering the future of lending: closing the gap between AI and the P&L

<h2><span class="h6"><span class="text-bold">Producing and servicing a mortgage keeps getting more expensive, even as AI and automation spread across the industry. The lenders closing that gap are taking a different approach to engineering.</span></span></h2>

<p>The cost of loan origination and servicing has never been higher. Producing a single loan <a rel="noopener noreferrer" href="https://www.mba.org/news-and-research/newsroom/news/2026/05/15/imbs-production-profits-remain-flat-in-first-quarter-of-2026" target="_blank">now costs roughly $11,898</a>—close to $4,000 more than the average since 2008—according to the MBA’s latest quarterly data. Keeping a performing loan on the books <a rel="noopener noreferrer" href="https://newslink.mba.org/servicing-newslink/2025/july/mba-servicing-newslink-tuesday-july-29-2025/chart-of-the-week-cost-of-servicing-performing-and-non-performing-loans/" target="_blank">runs about $176 a year</a>, roughly three times what it cost in 2008, despite two decades of servicing-technology investment.</p>
<p>Set those numbers against the trajectory of AI. <a rel="noopener noreferrer" href="https://www.cognizant.com/en_us/industries/documents/cognizant-hfs-research-reinventing-non-bank-mortgage-lending-journey.pdf" target="_blank">Our research with HFS</a> across nonbank lenders shows that the share of mortgage operations running through automation will climb from 44% in 2025 to 68% by 2027. Despite all this automation, <a rel="noopener noreferrer" href="https://www.cognizant.com/us/en/insights/insights-blog/closing-the-enterprise-ai-gap" target="_blank">our research also shows</a> that just 32% of Global 2000 executives can attribute measurable productivity gains to AI, and one-quarter have paused or shut down deployments they had already funded and staffed.</p>
<p>If AI and automation are everywhere, why hasn’t the cost of lending come down? That gap—between AI’s promise and payoff—is one that loan originators and servicers need to close. And closing it will require a large dose of engineering discipline.</p>
<h3><span class="h4" style="font-weight: normal;">Why the AI value gap persists in lending</span></h3>
<p>There are specific—and addressable—reasons why AI, so far, has not reduced loan management costs.</p>
<ul>
<li><b>Ineffective sequencing. </b>Many organizations focus their AI investments on areas like model selection and talent before building the data and platform foundations those investments depend on. Businesses that get this order wrong are 60% more likely to abandon a deployment than those that lay the groundwork first, <a rel="noopener noreferrer" href="https://www.cognizant.com/us/en/insights/insights-blog/closing-the-enterprise-ai-gap" target="_blank">according to our research.</a><br>
<br>
</li>
<li><b>Overlooked costs. </b>Despite dominating budget conversations, <a rel="noopener noreferrer" href="https://www.cognizant.com/en_us/ocm/bfsi/4486200_IOA-confronting-the-AI-velocity-gap.pdf" target="_blank">token costs account for just 5% to 10% of total agentic AI costs</a>. The remaining 90%-plus goes to process redesign, systems integration, human-in-the-loop operations, governance and infrastructure. Businesses that scope their programs around API pricing discover their real cost structure months in, once the budget is already set.<br>
<br>
</li>
<li><b>Linear, incremental uses of AI. </b>Most AI implementations are human-led and AI-enabled: copilots, automated summaries, proposal generation. These tools are useful, but their gains are linear because the deployments accelerate existing work without redesigning it. AI gains are more likely when AI is the primary actor and processes are reimagined.<br>
<br>
</li>
<li><b>Insufficient skilling. </b>The average Global 2000 company <a rel="noopener noreferrer" href="https://www.cognizant.com/us/en/insights/insights-blog/workforce-upskilling-for-ai-value" target="_blank">spends just over 0.2% of annual revenue</a> on AI skilling, even though trained workers report productivity gains at a rate that’s 28 points above their untrained colleagues. Without skilling, businesses will fail to realize full employee adoption or AI gains.<br>
<br>
</li>
<li><b>Retrofitted compliance. </b>Every time an agent touches a borrower record, a credit decision or a payment, the controls that govern it (audit trails, fair-lending checks, model risk management) have to be incorporated at the outset, not as a final step.</li>
</ul>
<h3><span style="font-weight: normal;" class="h4">What lending leaders do differently with AI</span></h3>
<p>The institutions closing the gap are taking a different engineering approach. For one, they’re designing for straight-through processing as the ultimate end goal. They may start with a deployment where AI acts as an assistant that always requires a person to act, but stopping here would result in an initiative that doesn’t scale.</p>
<p>From the beginning, <a rel="noopener noreferrer" href="https://www.cognizant.com/us/en/insights/insights-blog/how-agentic-ai-process-automation-is-changing-work" target="_blank">leaders plan to evolve toward “human-over-the-loop” operations</a>, where humans act in a supervisory role. With this approach, AI operates autonomously on individual transactions, and humans intervene only when predefined thresholds are breached or for governance purposes.</p>
<p><a rel="noopener noreferrer" href="https://www.cognizant.com/en_us/industries/documents/cognizant-from-ai-to-outcomes-closing-the-value-gap-in-non-bank-lending.pdf" target="_blank">The results bear this out</a>. We’ve heard about one chief operating officer who used an agentic platform to drive origination and underwriting labor from roughly $10,000 a loan toward $3,800, with an eventual target of $1,500. Processing time fell by more than a quarter, conversion improved 12% to 15%, and the program returned around $10 million.</p>
<h3><span class="h4" style="font-weight: normal;">Foundational engineering elements for AI in lending</span></h3>
<p>Leaders also engineer the following foundational elements in parallel with their first use cases, rather than as a prerequisite to them:</p>
<ul>
<li><b>Turn data into products. </b>Borrower, income, collateral, escrow and investor data needs to be governed, documented and structured so the agent can pull a complete, trustworthy view of a customer or loan from one place rather than reconstructing it from several screens.<br>
<br>
</li>
<li><b>Decompose the work. </b>Processes like underwriting and loan servicing need<b> </b>to be broken into discrete units that an agent can reliably complete and be measured against, such as clearing a specific condition, calculating income, analyzing an escrow account or quoting a payoff amount.<br>
<br>
</li>
<li><b>Make governance part of the pipeline. </b>Governance has to be built into the work agents do from the beginning. Personal data should be masked automatically; agent decisions should be logged in enough detail that they can be replayed and explained; and confidence thresholds should determine whether outputs need to be routed to a person. When an examiner makes a request, it should be answerable with a query, not a weeks-long scramble to reconstruct what happened.<br>
<br>
</li>
<li><b>Establish a measurement baseline. </b>Before the first workflow goes live, lenders need to record metrics like cost-per-loan and time-to-close on the origination side, cost-to-serve and retention on the servicing side. Any project whose savings can’t be proved with a before-after comparison risk being cut when budgets tighten.</li>
</ul>
<p>Additionally, businesses need to focus on workforce enablement. As loan processors and underwriters become exception specialists and AI system supervisors, institutions need a workforce that can improve the system, not merely operate it.</p>
<h3><span class="h4" style="font-weight: normal;">Moving from AI pilots to profits</span></h3>
<p>To make the shift to an engineering mindset, financial services firms should make the following commitments:</p>
<ul>
<li><b>Measure first. </b>Establish the unit economics that matter most: cost-per-loan and time-to-close, or cost-to-serve and retention. A program can’t be considered strategic without metrics to measure against.<br>
<br>
</li>
<li><b>Design a workflow. </b>Choose one workflow where the data is clean and the risk is contained. For originators, this could be document intelligence and condition clearing; for servicers, it could be the call center. Design it to run straight through, route exceptions by confidence, and log every decision to the standard an examiner would expect.<br>
<br>
</li>
<li><b>Enable the people who will run it. </b>Bringing even one-quarter of the affected workforce along is enough to move the numbers. Underinvest, and the software sits idle.</li>
</ul>
<p>The build-versus-buy question also needs to be resolved. Build where the process itself is the source of advantage, buy platforms trained on the realities of mortgage rather than generic AI, and partner where the goal is shared accountability for the outcome.</p>
<h3><span class="h4" style="font-weight: normal;">The work ahead with AI in lending</span></h3>
<p>Institutions should treat the next 18 months as a period of building the foundation, discipline and skill required for agentic AI. In the end, an engineering mindset is what closes the distance between what AI promises and what it returns.</p>