Tag Archives: Harvey

Offloading, not Surrendering, to AI

Posted on by
Reply

 

Robert MacKenzie and I published a column in Corporate Compliance Insights based on a recent eCornell workshop we taught. It reads,

We have been teaching lawyers how to use generative AI in their actual work — the drafting, reviewing and decision-making that fills their days. But when we designed our workshop, “Generative AI for Business Transactions,” we built it for a broader range of professionals: the healthcare compliance officer who had never opened ChatGPT, the corporate counsel whose legal department had recently deployed Harvey, the financial analyst running queries through Gemini and the operations manager who had heard the buzz but didn’t know where to begin. What we found confirmed what we suspected: the gap between professionals experimenting with AI and those waiting on the sidelines is widening fast. The ones who will thrive are not those using AI most aggressively but those using it most deliberately.

AI is transforming professional work

Generative AI is reshaping professional workflows in every industry we have encountered. In our workshop, we organize its everyday applications into four areas: communication, such as turning bullet points into polished emails and summarizing meeting transcripts; ideas and content, such as brainstorming and adapting material for different audiences; people and careers, such as preparing for interviews and difficult conversations; and money and numbers, such as building budgets, comparing costs and translating dense financial or legal language into plain English.

The best use cases are for time-intensive tasks. A transactional lawyer compares indemnification clauses across a dozen precedent agreements. A healthcare administrator turns regulatory guidance into a compliance checklist. A finance team compares top holdings across multiple fund prospectuses. The common thread: AI tools excel at quickly doing first-pass, high-volume work that used to consume hours.

A practical framework for responsible use

Every industry carries confidentiality obligations. Privilege in law, HIPAA in healthcare, fiduciary duties in finance, trade-secret protections in business. AI introduces a new exposure vector for professionals who are not careful about which tools they use. A key distinction we identify is the level of control and protections granted by enterprise AI tools versus consumer or free-tier tools. Enterprise tools are provided under negotiated contracts that typically commit the vendor not to train on your inputs and to keep your data confidential Consumer or free-tier tools often are packaged with settings permitting the provider to train on whatever information you input into the tool, undercutting confidentiality obligations you may be subject to. Vendor policies and features change, so verify that your expected protections are in place rather than assume.

We summarize this verification discipline in three words: pause, read, protect. Pause before entering data and ask whether it is safe to share and whether your workplace policies or professional obligations permit use of the tool for the intended purpose. Read the tool’s terms, and your workplace policies or guidance regarding the tool, to understand how your information will be treated. Protect by changing default settings, anonymizing confidential details and ensuring your cybersecurity and IT teams are in the loop when seeking to use new tools or approving use of updated features.

For task-level decisions, we recommend users adopt a red/yellow/green triage system. Red tasks are high importance and high risk and never get delegated to AI (e.g., strategy, high-stakes judgment calls and final approvals). Yellow tasks are lower importance and lower risk and may be delegated because they benefit from AI’s speed, but require competent human oversight and verification (e.g., research, first drafts and issue analysis). Green tasks are low importance and low risk and may, and sometimes, should, be delegated to AI, with minimal required human oversight (e.g., document reformatting, routine correspondence preparation and generation of ideas). If you supervise a team, you should be thinking about how you triage and how you want your team to triage matters. A breakdown in expectations can produce a “garbage-in, garbage-out” cycle.

Evaluating AI outputs critically

Our key takeaway is that AI’s greatest value lies in refining professional judgment, not replacing it. Generative AI is probabilistic, not deterministic. This means that the same prompt can produce different outputs in the same tool across different sessions. Models predict the next likely word in a sequence; they do not understand your question or verify their own answers.

Our recommendation to be effective with this technology: tell the tool what you need and be dynamic in your approach to prompting and task execution. We teach a simple prompting framework that is easy to recall and apply: RCTF—role, context, task, format. R: assign the AI a role. C: provide relevant context. T: define the task precisely. F: specify the output format. We think of this framework in the same way as ordering at a drive-thru. You would not pull up, say “food,” and expect to get what you want. You need to say what you are ordering, how you want it and where to hand it to you.

Other effective strategies we recommend professionals are:

    • Chunking. Breaking tasks into smaller pieces to keep tools on task.
    • Few-shot prompting. Provide examples of good work products to the tool before commencing a task.
    • Iterative refining. Adopting a “the first answer is a first draft” mindset.
    • Flipping interactions. Ask the tool to guide you on how to use it for a particular task.
    • Perspective switching. Assign the tool competing perspectives to pressure-test your work.

Managing hallucinations & overreliance

AI tools are known to generate plausible-sounding outputs that contain errors and invented citations. They also misread sources and silently drop items from long documents. These “hallucinations” are not bugs that will be patched away; they are inherent to how large language models work.

A deeper risk for inexperienced users of AI tools is what Wharton researchers Steven D. Shaw and Gideon Nave call “cognitive surrender.” In their 2026 study spanning three experiments and more than 1,300 participants, they found that participants were highly susceptible to following incorrect advice from AI tools. Access to an AI chatbot during the experiments appeared to inflate participants’ confidence in their answers, even when the answers were wrong. Observations like these point to a broad human tendency towards cognitive surrender: When a fluent, confident-sounding tool delivers a coherent answer, the pull to accept it is powerful.

We want to draw a sharp distinction between cognitive surrender — letting AI do your deliberate thinking and accepting its output uncritically — and “cognitive offloading” — handing defined steps to AI while retaining control of the overall analysis. The first is a professional hazard. The second is a legitimate productivity strategy. After every substantive AI-assisted task, ask yourself: Have I thought this through as fully as I would have without the tool? If not, dig back in.

Building reusable templates & checklists

One of the highest-value applications of generative AI is converting complex source documents into workflows a team can reuse, such as checklists, trackers and comparison matrices. In our workshop, we demonstrate how to take a dense document and instruct AI tools to produce a structured checklist to capture desired variables, like task status, assigned parties, deadlines, source references and risk flags.

We also teach benchmarking: uploading a set of similar documents and directing the AI tool to create a comparison matrix of key terms among the documents. AI tools offer value in their continually improving (but imperfect) ability to accurately extract and categorize information from new documents based on historical templates. For professionals with high accuracy needs, this skill can offer considerable leverage by accelerating the manual steps in these types of workstreams (initial review, identification and extraction or summarization of terms).

The bottom line

Whether you work in law, healthcare, finance or any field built on complex documents and careful analysis, the starting point is the same: Develop your own judgment first, verify before you rely and triage every task before you hand it off.

Storm-Induced Delinquencies

Posted on by

The Urban Institute’s Housing Finance Policy Center has released its November 2017 Housing Finance at a Glance Chartbook. The Introduction looks out how this summer’s big storms have pushed up delinquency rates:

The Mortgage Bankers Association recently released the results of its National Delinquency Survey (NDS) for Q3 2017. The non-seasonally adjusted NDS data for Q3 2017 showed a significant increase in delinquency rates across all past due categories (30-59 days, 60-89 days and 90 days and over). The increase was largest–and most noteworthy–for the 30-59 day category, spiking by 57 basis points from 2.27 percent in Q2 2017 to 2.84 percent in Q3. The D60 rate increased by a much smaller 12 basis points, from 0.74 to 0.86 percent, while the D90 rate increased the least, by 9 basis points, from 1.20 to 1.29 percent. The rise in delinquencies was broad based, affecting FHA, VA and Conventional channels with FHA D30 seeing the largest increase (4.57 to 5.92 percent).
While early payment delinquency rates were expected to increase in the wake of the storms Harvey, Irma and Maria for the affected states, the magnitude of increase in the D30 rate is quite remarkable. The reported Q3 2017 D30 rate is the highest in nearly four years. The 57 basis points increase in a single quarter was also the largest in recent history. The last time D30 rate increased by more than 50 bps in one quarter was in Q4 2000, when it rose by 61 bps. In comparison, both D60 and D90 rates, while slightly higher in Q3, are well within their recent range.
MBA’s state level NDS data confirms that storms were a major driver behind the increase. For Florida, the non-seasonally adjusted D30 rate more than doubled from 2.12 to 4.64 percent, the highest ever D30 rate recorded. The D30 rate for Puerto Rico also nearly doubled from 4.98 to 9.12 percent, while Texas D30 rate increased from 5.05 to 7.38 percent. The increase in FL and PR was larger than in TX because of the statewide impact of hurricanes Irma and Maria. In contrast Harvey’s impact was limited to Houston and surrounding areas. The increase in the D90 rate is not storm-related as not enough time has elapsed since the storms made landfall (Harvey made landfall in Houston on August 25, Irma made landfall in Florida on September 9, and Maria made landfall in Puerto Rico on September 20).
Besides storms, there are other factors that are driving the D30 rate higher. As the figure shows, there is a very strong seasonal pattern associated with 30 day delinquencies. The D30 rate typically witnesses an uptick in the second half of each calendar year after declining in the first half because of tax refunds. Another reason for the Q3 increase is that the last day of September was a Saturday, which means that payments received on this day were not processed until Monday Oct 2nd and were identified as past due (mortgage payments are due on the 1st of the month; D30 rate is based on mortgages unpaid as of 30th of the month).
There is one more thing worth pointing out. Many borrowers affected by recent storms have received forbearance plans that allow them to defer mortgage payments for a few months. Under the NDS methodology, these borrowers are considered delinquent. Many will likely resume making monthly payments once they regain their financial footing or after forbearance ends. Others unable to afford payments could get a loan modification. Therefore, although it will take several quarters before the eventual impact of storms on delinquency rates becomes clear, many borrowers who are currently 30-days delinquent might not enter D60 or D90 status.
While the Chartbook does not look at the longer term impact of climate change on mortgage markets, it is clear that policy makers need to account for it in terms of mortgage servicing, flood insurance, land use and building code regulation.

FHA Annual Check-up

Posted on by

The Department of Housing and Urban Development released its Annual Report to Congress Regarding the Financial Status of the FHA Mutual Mortgage Insurance Fund. The MMIF fund is the FHA’s main vehicle for insuring mortgages. As we saw last week, FHA reverse mortgage (formally known as “Home Equity Conversion Mortgage” or “HECM”) portfolio is not doing so well. FHA standard (sometimes referred to as “forward”) mortgages are doing better, although their performance is also slipping.

The MMIF declined from its 2.35 percent FY 2016 Capital Ratio to 2.09 percent. This still exceeds its statutorily-required level of 2.00 percent.  The Economic Net Worth of the MMIF was $25.6 billion while the MMIF Insurance-in-Force was approximately $1.23 trillion at the end of FY 2017. The decline was driven by the negative Economic Net Worth of the reverse mortgage portfolio, as the capital ratio for the forward mortgage portfolio on its own was 3.33%.

The report contains a multitude of useful tables and charts about the FHA’s mortgage portfolio. The FHA has an 18 percent share of the mortgage market, which is pretty high. (Table A-2) Indeed, it is in the same range of its market share during the financial crisis years (2008-2010). The FHA remains a strong force in the first-time homebuyer market, with an 82.2 percent share. (Table B-2)

The FHA’s objectives for FY 2018 are worth reviewing:

Play a Significant Role in Disaster Recovery. In the wake of Hurricanes Irma, Harvey, and Maria, and wildfires in California, in FY 2017 and the first part of FY 2018, FHA has played a significant role in relief and recovery efforts in affected areas, while taking immediate actions to protect its Single Family assets and financial exposure. (78)

Make Necessary Changes to the Home Equity Conversion Program (HECM). During FY 2017, FHA revised the HECM initial and annual Mortgage Insurance Premiums (MIPs), and Principal Limit Factors (PLFs). These revisions were necessary to enable FHA to continue to endorse HECM loans in FY 2018, protect the program for seniors, and balance serving FHA’s mission with taxpayer protection. (79)

No less important than these objectives is the FHA’s second-to-last one, Technology Modernization:

FHA is working to update its systems over the coming years to allow the Agency to work more effectively with lenders participating in the program, while operating FHA with greater efficiency and control. The technology systems that support FHA’s Single Family business have an average age of more than 18 years, with the Computerized Homes Underwriting Management System (CHUMS) exceeding 40 years. Similarly, the systems supporting the servicing, default, claims and REO areas have an average age of 14 years. FHA’s systems have been maintained, modified and enhanced over the years, but it has become fundamentally difficult and exceedingly expensive to maintain systems beyond their usable life. FHA’s outdated systems make it more difficult to work with lenders and to collect and manage important data. FHA remains a largely paper-processing entity while the rest of the industry has increasingly migrated to digital processes. FHA needs systems that can capture and effectively process the extensive volumes of data now in use, with enhanced storage and processing capabilities to handle the migration from paper forms to digital ones. Additionally, FHA requires the ability to analyze and manage insured loans comprehensively over the many phases of the mortgage life cycle. (80)

When you stop and think about how bad the state of the FHA’s technology is, you think that maybe this should be their top priority.

Fannie, Freddie and Climate Change

Posted on by

NOAA / National Climatic Data Center

The Housing Finance Policy Center at the Urban Institute issued its September 2017 Housing Finance At A Glance Chartbook. The introduction asks what the recent hurricanes tell us about GSE credit risk transfer. But it also has broader implications regarding the impact of climate-change related natural disasters on the mortgage market:

The GSEs’ capital markets risk transfer programs that began in 2013 have proven to be very successful in bringing in private capital, reducing the government’s role in the mortgage market and reducing taxpayer risk. Investor demand for Fannie Mae’s CAS and Freddie Mac’s STACR securities overall has been robust, in large part because of an improving economy and extremely low delinquency rates for loans underlying these securities.

Enter hurricanes Harvey, Irma and Maria. These three storms have inflicted substantial damage to homes in the affected areas. Many of these homes have mortgages backed by Fannie Mae and Freddie Mac, and many of these mortgages in turn are in the reference pools of mortgages underlying CAS and STACR securities. It is too early to know what the eventual losses might look like – that will depend on the extent of the damage, insurance coverage (including flood insurance), and the degree to which loss mitigation will succeed in minimizing borrower defaults and foreclosures.

Depending on how all of these factors eventually play out, investors in the riskiest tranches of CAS and STACR securities could witness marginally higher than expected losses. Up until Harvey, CRT markets had not experienced a real shock that threatened to affect the credit performance of underlying mortgages (except after Brexit, whose impact on the US mortgage market proved to be minimal). The arrival of these storms therefore in some ways is the first real test of the resiliency of credit risk transfer market.

It is also the first test for the GSEs in balancing the needs of borrowers with those of CRT investors. In some of the earlier fixed severity deals, investor losses were triggered when a loan went 180 days delinquent (i.e. experienced a credit event). Hence, forbearance of more than six months could trigger a credit event. Fannie Mae put out a press release that it would wait 20 months from the point at which disaster relief was granted before evaluating whether a loan in a CAS deal experienced a credit event. While most of Freddie’s STACR deals had language that dealt with this issue, a few of the very early deals did not; no changes were made to these deals. Both Freddie Mac and Fannie Mae have provided investors with an exposure assessment of the volume of affected loans in order to allow them to better estimate their risk exposure.

So how has the market responded so far? In the immediate aftermath of the first storm, spreads on CRT bonds generally widened by about 40 basis points, meaning investors demanded a higher rate of return. But thereafter, spreads have tightened by about 20 basis points, suggesting that many investors saw this as a good buying opportunity. This is precisely how capital markets are intended to work. If spreads had continued to widen substantially, that would have signaled a breakdown in investor confidence in future performance of these securities. The fact that that did not happen is an encouraging sign for the continued evolution of the credit risk transfer market.

To be clear, it is still very early to reasonably estimate what eventual investor losses will look like. As the process of damage assessment continues and more robust loss estimates come in, one can expect CAS/STACR pricing to fluctuate. But early pricing strongly indicates that investors’ underlying belief in these securities is largely intact. This matters because it tells the GSEs that the CRT market is resilient enough to withstand shocks and gives them confidence to further expand these offerings.