By: Scott Anderson and Bernadette Kogler

AI is already disrupting parts of the U.S. labor market. The more contested question for residential mortgage investors is how and where. The dooms-day camp sees broad structural displacement where white-collar knowledge workers are replaced by AI agents. The optimists counter that AI makes workers more productive, that falling labor costs will expand demand leading to net employment gains. Both camps agree on one thing: the impact will not be uniform. Disruption will concentrate in pockets — specific occupations, metro centers, and income bands.

That specificity is precisely what makes this a mortgage credit problem. Local labor shocks have always transmitted to local housing markets through a familiar mechanism: income impairment reduces demand which pressures home prices, and price declines erode the collateral cushion that protects investors. What varies is the speed, the depth, and whether prices recover once the shock stabilizes.

Early indicators are consistent with the thesis. A Wall Street Journal analysis found the unemployment rate for recent college graduates ages 20–29 at 7.1% as of October 2024 — nearly double the 4.1% rate for the general population — with computer science graduates at 6.1% unemployment and computer engineering majors at 7.5%. The Federal Reserve of NY also published data showing labor market conditions worsened for recent college graduates at the end of 2025, with the unemployment rate climbing to about 5.7% in Q4 2025 — and underemployment rising to 42.5%, its highest level since 2020.

The occupations most exposed to AI displacement cluster in precisely the metros where high-balance prime mortgages are most concentrated, changing the traditional metrics that investors rely on for early warning. AI-driven labor disruption is not a macro thesis to be managed with national averages. While mortgage borrower occupation data is not complete or granular enough to drive AI job loss assumptions, property location data is. The risk needs to be identified, stress-tested, and monitored at a granular level — before it shows up in delinquency data.

The good news is that the data, models and historical reference points exist. With the growth of Non-QM lending, RTL and loan products that carry credit risk without an insurance wrapper, treating AI-driven labor disruption as a present-tense portfolio risk is not optional. To move from thesis to numbers, we ran a portfolio-level stress test on the universe of securitized Non-QM mortgages – starting with how AI exposure is measured at the local level.

Key Findings for investor consideration:

1. AI labor disruption is expected to have a higher disproportional impact on white-collar professional jobs that are concentrated in certain “knowledge-work” centers.
2. The largest job markets and the deepest credit pools are systematically the most AI-exposed, suggesting AI risk and portfolio risk overlap rather than diversify.
3. Stress-testing the $196B universe of securitized Non-QM mortgage loans shows the magnitude is meaningful: under an AI-driven stress scenario calibrated to GFC-level shocks but re-allocated to AI-exposed geographies, RiskSpan’s credit model projects universe-wide cumulative defaults to nearly double and losses more than quadruple.
4. Securitized deals benefit from built-in geographic diversification, which dampens the absolute impact across deals. Whole-loan investors, who construct their own geographic mix, face concentration risk that accumulates silently if exposure isn’t an explicit input into selection and surveillance.
5. Act now — put the infrastructure in place to use granular data to monitor unemployment, HPI and loan performance and apply scenarios where relevant.

Measuring AI Exposure at the Local Level

Leveraging research published by the Brookings Institution, RiskSpan used their geographic AI exposure scores as the foundation for a credit stress scenario. The resulting county- and metro-level scores follow a similarly clear geographic pattern to that published by Brookings. Highly educated, high-paying knowledge-work centers — Santa Clara (42.8% exposure), King County WA, New York County, DC, Boston — show the highest exposure. Smaller industrial and rural counties score considerably lower. As a sanity check on the methodology, we aggregated and plotted state-level AI exposure against the share of each state’s adult population with a bachelor’s degree or higher (a well-understood proxy for knowledge-work concentration). The two track closely, which is the result one would expect — it confirms the Brookings measure is capturing what it intends to and isn’t producing arbitrary geographic rankings.

Unlike earlier automation, generative AI concentrates in exactly the metros that house the largest mortgage markets: as the chart below shows, the most AI-exposed MSAs are also among the deepest non-agency RMBS markets.

One important caveat before moving on to how we used this data. As Brookings is careful to note, “‘exposure’ does not speak only to the displacement of workers; it also may involve their ‘augmentation’ through rapidly improving AI tools.” The same coder whose role is highly exposed could be made dramatically more productive rather than displaced; the same financial analyst could spend less time on rote modeling and more on judgment-intensive work. Whether exposure plays out as net displacement, net augmentation, or some mix will vary by occupation, employer, and how quickly the technology evolves. For purposes of this exercise, we focus on the displacement story: the scenario where AI exposure translates into labor market shocks that transmit to housing markets and mortgage credit. That’s not the only possible outcome, but it’s the one investors need to be prepared for.

Translating that exposure into a portfolio-level credit stress test on the Non-QM universe, the largest segment of the non-agency RMBS market, required two steps.

Step 1: Construct an AI-driven stress scenario calibrated to the GFC.

Rather than forecast AI displacement magnitude — a genuinely uncertain exercise — we sized the stress to a known reference point: the total HPI decline and unemployment increase observed during the GFC. The GFC offers a useful calibration episode because the empirical relationship between local unemployment shocks and HPI declines is strong and well-documented, regardless of which direction the causation primarily ran. The chart below shows that relationship across MSAs — a clear negative slope linking the spike in unemployment rates to the depth of the HPI decline.

The critical observation for our purposes is what happened in 2008: the metros most exposed to AI today (DC, Boston, San Francisco, Seattle) cluster toward the middle of that scatter plot. They experienced more moderate impacts than the hardest-hit metros, because the speculative and underwriting excesses that drove the GFC were concentrated elsewhere — Las Vegas, Phoenix, Miami, the Inland Empire. An AI-driven shock would invert that geography.

We took the GFC’s national-level shock magnitudes and redistributed them, concentrating the stress in MSAs with high AI exposure. Starting from Brookings’ occupation-level AI exposure scores rolled up to MSA and state-level, we scaled both the unemployment and HPI shocks to each MSA in proportion to its AI exposure share. The result is a per-MSA stress profile that adds up to GFC-magnitude pain at the national level but lands very differently on the map.

The chart below translates that scaling into implied HPI shock magnitudes across the metros where non-agency RMBS exposure actually sits. The metros at the bottom end of both axes (high portfolio balance and severe implied HPI shock) are where the AI thesis would hit hardest if it materializes.

Step 2: Run the loan universe through two scenarios.

We took the $196B universe of Non-QM mortgage loans (~500K loans across 810 active Non-QM deals as of April 2026) and ran these loans through RiskSpan’s proprietary credit model under two scenarios: a baseline and the combined AI-driven Unemployment + HPI stress scenario described above. Every loan received a projected cumulative default rate and cumulative loss rate under each scenario, isolating the AI-driven impact at the loan level.

Findings

Under the AI-driven combined stress scenario, universe-wide cumulative defaults nearly double (3.6% → 7.4%) and cumulative losses more than quadruple (0.5% → 2.2%). The steeper loss response reflects the familiar non-linearity between defaults and severities under HPI stress. The mechanism behind those aggregate numbers becomes clearer when you look at it geographically.

Finding 1 — AI exposure drives MSA-level credit performance.

This finding is by construction — we built the stress scenario by re-allocating shocks to AI-exposed MSAs, so the model output should track AI exposure. The MSA-level chart confirms it does. MSAs at the high end of AI exposure (DC, San Francisco, Boston, Seattle) see over an 8% percentage point (pp) increase in cumulative defaults compared to MSAs at the low end. Riverside, near the lower end of the AI exposure range at just under 30%, sees only a 2.6% pp increase in cumulative defaults. There’s MSA-specific noise from collateral mix, but the geographic risk concentration the AI thesis predicts shows up cleanly in the model output. This is the core mechanism: re-allocate the stress to AI-exposed geographies, and the credit impact tracks the re-allocation.

Finding 2 — The aggregate impact is shaped primarily by leverage.

Decomposing the universe results by current LTV bucket explains where the headline numbers come from. The first thing worth noting is what the LTV distribution looks like: less than 5% of total non-QM balance sits above 80% LTV, and roughly half is below 60%. That’s a meaningfully larger equity cushion than the borrower base of the mid-2000s, when the proliferation of low- and no-down-payment products left investors with thin protection against any HPI decline. Today’s Non-QM universe starts the AI stress scenario with a lot more skin in the game.

Within that distribution, stress impact scales monotonically with leverage, exactly as one would expect: low-LTV loans (≤40%) see a 1.3pp increase in defaults (from 0.96% to 2.24%), while high-LTV loans (>80%) see a 6.7pp increase (from 5.97% to 12.62%), though the relative impact on cumulative defaults is broadly stable across LTV buckets, with defaults about doubling in each bucket. The same pattern holds for losses, where the highest LTV buckets see losses rise by over 3pp under stress versus 0.16pp for the lowest LTV bucket. FICO declines slightly as LTV increases (from 756 for the lowest LTV bucket to 745 for the highest), so the projected cumulative default differences across LTV buckets are driven, at least partly, by underlying borrower credit as well.

Finding 3 — Deal-level concentration matters, but only in relative terms.

When we rolled up by deal rather than by MSA, the picture got more interesting. The absolute (pp) impact across deals shows essentially no relationship with the deal’s AI exposure share — high-AI-exposure deals don’t show systematically larger pp increases. But the relative impact (stress / base − 1) is clearly correlated with AI exposure.

One plausible explanation is that deal-level base default rates vary widely based on traditional credit characteristics — LTV mix, FICO mix, doc type (Full Doc vs. DSCR vs. Bank Statement), occupancy. That variation in base rates may be dominating the AI signal in absolute terms while leaving it visible on a relative basis. Either way, the relative measure is the cleaner lens for isolating AI exposure’s marginal contribution to deal-level risk.

There’s also a more reassuring story embedded in these charts, at least for securitized mortgage assets. Compare the deal-level scatter to the MSA-level scatter and you’ll notice the deal-level points cluster in a meaningfully narrower range — both on the AI exposure axis (X) and on the impact axis (Y). That tighter clustering is geographic diversification at work: even the most AI-concentrated deal in the universe pulls in loans from enough different MSAs that its blended AI exposure sits closer to the universe average than the most exposed individual MSA does. Diversification within deals is doing what it’s supposed to do — dampening the variance of outcomes — even though the unit of analysis is smaller than an MSA. The magnitude of the differential across deals is bounded by how concentrated any individual issuer is willing to get.

The implication for whole-loan portfolios is sharper. Whole-loan investors don’t get the benefit of pre-built deal diversification — they’re constructing their own geographic mix one loan or pool at a time, and concentration risk accumulates silently if it isn’t an explicit input into selection and surveillance. AI exposure is one such risk, and a relatively new one, but it sits alongside more familiar concentrations — natural disaster exposure, regional employment dependencies, single-industry metros — that warrant the same lens. The toolkit is the same in each case: granular geographic data, stress scenarios calibrated to defensible reference points, and loan-level credit modeling. Whether the goal is initial portfolio construction or ongoing surveillance of a seasoned book, the ability to identify these concentrations early — before they show up in delinquency data — is what separates active risk management from passive observation.

What This Means for Investors

History offers instructive analogies including the energy-patch metros of the mid-2010s when oil prices collapsed — Houston, Oklahoma City, and Calgary experienced home price corrections in the range of 5 to 15 percent. Regional manufacturing declines showed a similar story over a longer time period — the long contraction of auto-sector employment in Detroit, for example, played out in local housing markets over years while national HPA remained positive. In neither case did the national trend warn you that specific MSAs were in distress.

Acting now means two things in practice. The first is tracking — building the early-warning infrastructure before you need it. That means monitoring MSA-level unemployment, not just the national rate; watching labor force participation shifts; and tracking occupational mix changes in the markets where your collateral is heaviest. These indicators move before delinquency does. The second is scenario preparation — running stylized AI displacement shocks against the portfolio today using MSA-level or county-level HPA overrides and occupational exposure overlays. The goal is not a forecast but rather a map of your exposure — which corners of the portfolio are most sensitive to particular labor market stress, and what the loss distribution looks like if the more adverse scenarios begin to materialize. That map is worth building now, while the cost of being wrong about the timing is low.

RiskSpan is releasing a new Non-QM-specific credit model in the coming weeks, estimated directly on Non-QM performance data and designed around the underwriting features (DSCR, bank statement, expanded LTV) that distinguish these segments. Re-running this scenario and other stress scenarios on the new model is on the near-term roadmap and may sharpen some of the findings reported here. For investors who want to evaluate their own exposure, the analytical infrastructure described here can be applied to specific portfolios on request.

Note: The Brookings Institution has been tracking AI’s labor market impact since 2019. A 2024 Brookings piece analyzed data from OpenAI to measure occupational AI exposure as the share of tasks where AI could reduce human completion time by at least 50%. AS one would expect, the cognitive/manual divide is evident in the rankings, with computer and mathematical work atop the list at 75% while construction sits at the bottom at 5.6%.

Their 2025 piece extends that work to geography by multiplying each occupation’s exposure rating by its share of local employment in each county or metro, then aggregating. Counties with large workforces in highly exposed occupations (software developers, financial analysts, lawyers, marketing professionals) score higher than those whose employment skews toward construction, agriculture, or in-person services.

Key References:

Generative AI, the American worker, and the future of work | Brookings

The geography of generative AI’s workforce impacts will likely differ from those of previous technologies | Brookings

There Is Now Clearer Evidence AI Is Wrecking Young Americans’ Job Prospects – WSJ

www.newyorkfed.org/research/college-labor-market

As generative AI becomes a powerful tool in Private asset-backed finance (ABF), the need for precision and transparency is more critical than ever. At RiskSpan, we’re applying Large Language Models (LLMs) to automate and accelerate private ABF deal modeling and surveillance. But speed is only half the battle—accuracy is non-negotiable.

That’s where Human-in-the-Loop (HITL) validation plays a vital role. While the RiskSpan platform incorporates sophisticated AI guardrails, we believe the right blend of automation and expert oversight ensures results that are not just fast—but reliable, auditable, and production-ready.

The AI-Powered Workflow: What’s Automated

Our private ABF modeling and surveillance system uses LLMs to handle several critical tasks:

• Data Extraction: Parsing offering memos, indentures, and loan tapes to extract structural and financial data.
• Deal Code Generation: Producing executable waterfall models based on extracted rules.
• Database Ingestion: Uploading validated deal terms and triggers into the RiskSpan system of record.
• Surveillance Automation: Running periodic deal performance analyses and compliance checks.

But What About Hallucinations?

Generative models are powerful but imperfect. Without the right controls, they can fabricate securitization tranches or fees that are not present in the deal documents. They can also misinterpret waterfall rules or omit critical override conditions or generate semantically incorrect code for cashflow models. To address these challenges, RiskSpan employs a multi-layered safeguard framework, combining asset class based extraction; LLM-as-Judge; Rule-Based Guardrails and Inline Human Review

Humans in the Loop: Three Layers of Oversight

We’ve embedded human validation at three key points in the deal lifecycle:

1. Pre-Modeling Validation: before LLM-generated outputs are finalized, RiskSpan analysts review extracted terms and model prompts—correcting anything misaligned with the source documents.
2. Inline Oversight: during waterfall code generation, humans validate AI-generated logic in context, ensuring correct treatment of subordination, triggers, caps/floors and other.
3. Post-Deployment Monitoring Surveillance: outputs are reviewed both by the RiskSpan team and client-side structuring or credit teams. Feedback is looped back into model tuning and prompt optimization.

Looking Ahead: RAG and Continuous Improvement

We’re actively exploring Retrieval-Augmented Generation (RAG) to reduce hallucinations even further. By grounding AI responses in deal-specific material such as offering documents, trustee reports, and internal risk memos—we aim to: 1) eliminate off-topic responses. 2) increase trust in model-derived outputs and 3) enable deeper customization per issuer or asset class.

The Takeaway

AI is transforming how private ABF deals are modeled and monitored—but it must be grounded and guided by human expertise and built for institutional rigor. At RiskSpan, we’re not just accelerating workflows—we’re raising the bar for accuracy and trust in AI-assisted private structures. Human-in-the-loop is not a fallback—it’s a strategic pillar. Want to see how our AI platform works in action? Reach out to schedule a demo or contact your RiskSpan representative to learn more.