When the Loan Is FHA or VA, the Waterfall Is the Investor's, Not the CFPB's: Government Default Servicing With AI After the 2025-2026 Overhaul

One Waterfall Is Not Enough

The most common design mistake I see in default-servicing AI is a single loss-mitigation agent built to Regulation X §1024.41 and pointed at every delinquent loan in the book. The CFPB rules are real and they bind, so the instinct to build to them is right. The problem is that they are the floor, not the whole structure. When the loan is insured by FHA or guaranteed by VA, the option the borrower actually gets is set by the investor's waterfall, and that waterfall sits on top of the CFPB timeline rather than replacing it. An agent that knows Reg X cold and does not know whose loan it is looking at will confidently offer a borrower something the investor retired months ago.

That failure went from theoretical to urgent in the last twelve months, because both government waterfalls were rebuilt. A servicer running an agent that was accurate in early 2025 is running an agent that is wrong now, and the borrower on the other end of a default conversation is the one who pays for the lag.

What FHA Actually Changed, and the Date That Matters

FHA issued Mortgagee Letter 2025-12 on April 15, 2025, with a mandatory implementation date of October 1, 2025. It did not tune the existing framework. It ended it. The COVID-19 Recovery Options, the COVID-19 Advance Loan Modification, and FHA-HAMP all sunset on September 30, 2025, and after that date a servicer may not send final loss-mitigation documents on any of them. In their place is a new permanent framework with a hard 24-month cumulative cap on permanent loss-mitigation eligibility, and the amendments land in HUD Handbook 4000.1, the servicing rulebook every FHA servicer is examined against. FHA has also said publicly it is still evaluating whether the Payment Supplement stays in the program at all, which means the framework is not only new, it is still moving.

Here is the operational trap. An agent whose knowledge base was loaded before October is not merely out of date, it is dangerous, because the option it recommends most confidently, the one it has the most training examples for, is exactly the one that no longer exists. We hit a version of this in shadow mode on an FHA book: the agent kept surfacing FHA-HAMP as the recommended path on files reviewed after the cutoff, because the underlying guidance in its retrieval store had no effective-date on it and the model had no reason to prefer the newer rule. Nobody's loan was harmed, because it was shadow mode and a human made every call, which is the entire reason we run shadow mode before an agent decides anything. The fix was not a better prompt. It was structural, and I will come back to it.

What VA Changed, on a Different Clock

VA moved on its own timeline, and it moved the opposite direction from where it had been. The Veterans Affairs Servicing Purchase program, VASP, wound down, and the VA Home Loan Program Reform Act signed on July 30, 2025 authorized a replacement. The VA Partial Claim program opened for submissions on June 15, 2026, and servicers have until November 28, 2026 to implement it in their systems. Mechanically it is a different animal than VASP: instead of VA buying the defaulted loan, the servicer advances the missed payments after a three-month trial period and VA reimburses it, and the arrearage becomes a separate subordinate lien the borrower repays when the home is sold, refinanced, or paid off. It sits inside a new standardized loss-mitigation waterfall for VA loans.

So as of this writing a servicer with a mixed book is running three different rule sets at once: the CFPB timeline that applies to all of them, an FHA waterfall that reset in October 2025, and a VA waterfall with an implementation deadline still ahead of it in November 2026. The dates do not line up, the options are not the same, and the consequence of getting it wrong on any of the three is a borrower steered toward foreclosure on a path they should never have been on.

The Design Decision: Version the Waterfall, Gate on Investor and Note

The lesson from that shadow-mode failure is that a loss-mitigation agent cannot treat guidance as a flat pile of documents. It has to treat each waterfall as a versioned artifact with an effective-date range and an investor tag, and it has to resolve which version applies before it evaluates anything. The first thing our agent establishes on a delinquent file is not the borrower's hardship. It is two facts: who owns or insures the loan, and what date governs the evaluation. Only then does it load the waterfall that was in force for that investor on that date and walk it in order.

That single decision, gating on investor and effective date before evaluation, is what turns a retired option from a recommendation the borrower sees into a version that never loads. It also produces the record a servicing examiner wants, because the file now shows which waterfall version was applied and why, dated, rather than a determination that has to be reverse-engineered after a complaint. We version the CFPB timeline the same way, so the Reg X clocks, the acknowledgment window, and the anti-evasion rules against dual tracking under §1024.41 run alongside the investor waterfall instead of being bolted on.

What the Agent Computes, and What It Never Decides

On a government-loan default file the agent does the work that is arithmetic and tracking, which is most of the work and the slowest part when a human does it. It reads the loan and the hardship documentation, confirms the investor and the note, computes eligibility against the correct waterfall step by step, tracks every CFPB and investor timeline running on the file, and drafts the evaluation notice. An illustrative figure we hold internally is agreement with an independent re-review in the high nineties on the eligibility computation before that task goes live, set per investor because an FHA eligibility check and a VA eligibility check fail in different ways and do not deserve the same bar.

What the agent does not do is make the determination that carries a foreclosure consequence. A denial of loss mitigation, or an approval that starts a trial payment plan, is a decision a human owns, because the cost of the agent being wrong is not a rework ticket, it is a home. The agent hands the human a completed evaluation with the waterfall version cited and the math shown, and the human decides. That line is the same one we draw across mortgage loss mitigation under Reg X, and it is not a limitation we apologize for. It is the point.

There is a governance layer under this too. An eligibility model whose output feeds a servicing decision is a model, so it sits inside the model risk discipline SR 11-7 describes: documented intended use, validation before it is relied on, and monitoring tied to the outcomes that matter, which here means the rate at which humans reverse what the agent proposed and the defects a later review finds on files the agent touched.

Why This Is the Servicer's Problem Now

The reason to care about this precisely now is that the window where an agent can be quietly wrong is open. Both waterfalls changed inside a single servicing year, one deadline has passed and one has not, and the borrowers in default on government loans are among the least able to absorb a servicer's error. A default agent that was validated in March 2025 and has not been re-versioned is recommending at least one retired FHA option today and will be unprepared for the VA program when its November deadline arrives.

At Sei we treat each investor waterfall as versioned, effective-dated content that the agent resolves before it evaluates a file, and we re-validate against the current guidance whenever an investor moves, because the alternative is an agent that is accurate on the average and wrong on exactly the loans where being wrong costs the most. The agent makes default servicing faster. The versioning is what keeps faster from meaning wrong.