I publish a free library of mortgage regulations: FHA Handbook 4000.1, the Fannie and Freddie guides, the VA handbooks, USDA, the CFPB regulations, Ginnie Mae. Verbatim, section by section, 3,127 of them. It exists so an AI assistant has accurate text to stand on instead of recalling a rule from memory and getting it subtly wrong.
Assistants use it constantly, which gives me a narrow window nobody else has: I can watch what they come for.
Not who asked. Not what anyone typed into an assistant. Not what file they were working on. Just which pages get pulled, how many at a time, and whether they ever come back. That turns out to be enough to see a pattern.
What the assistants actually come for
Over two days, question-time AI agents made 1,637 fetches against this library. Not crawlers building an index, which I exclude: these are the live retrievals that happen when somebody asks an assistant a mortgage question.
- 81% were rule text. Straight to a regulation page.
- The remainder was almost entirely
robots.txt, the homepage, and data files: section-scoped spreadsheets, and in a couple of cases the entire library as a CSV archive. - Roughly 1.1 fetches per page. They read a section once and move on. Almost nothing gets read twice.
That last number is the one worth sitting with. An assistant working through a problem — testing a conclusion, going back to check a definition, reconciling two sections that disagree — would revisit pages. These do not revisit. They arrive, take the text, and leave.
And here is what they never fetch: the pages on this site that explain how to use any of it. Not the methodology, not the guidance on running this against your own documents, not the regulatory update kits. Assistants come for the rule and only the rule.
The eleven questions I can see, and why I will not overclaim them
Microsoft publishes the queries Copilot used to ground an answer. I have exactly eleven:
38 u.s.c. Sec. 3737 · chapter 18 of gnma guide · fannie mae community second guidelines · fannie mae credit report requirements · fannie mae offer letter · fannie mae paystub requirements · fnma allowable foreclosure attorney fees · fnma gift funds · fnma leasehold requirements · gross up social security fannie mae · loan administrators interest rate resets rollovers rate change notices
Every one is a lookup. Find me the rule. Not one contains if, does my, qualify, or eligible.
It is tempting to present that as proof of how people talk to assistants. It is not, and the reason matters. A grounding query is not what the user typed; it is the query the assistant generated to go find sources. Someone could ask a genuinely rich question — does my borrower on maternity leave qualify if her father wires the down payment — and the assistant would still emit fnma gift funds to go fetch.
So eleven machine-rewritten queries cannot tell you how people ask. What they do tell you is what the assistant decided it needed, and in all eleven cases it decided it needed one rule.
The retrieval pattern says the same thing far more loudly, and that is where I would rest the argument.
That is a real gain, and people are right to take it
None of this is a criticism, and I want to be clear about that before I say anything else.
Agency guidance is genuinely hard to navigate. The Selling Guide is enormous, 4000.1 is enormous, and the search box on a government website is worse than useless for finding the paragraph you actually need. An assistant that puts the right section in front of you in eight seconds beats twenty minutes of hunting through a PDF, every time.
It also says something good about the people asking. Reaching for a named rule means you already know the regulatory landscape well enough to know where the answer lives. And going to a verbatim source at all is exactly the instinct you want in someone doing this work, especially now that assistants will confidently produce regulatory text that does not exist. These are not naive questions. They are experienced people using a new tool for the thing it obviously does.
They adopted what helps, and they adopted it fast.
What is being left in the box
Here is what strikes me as the miss.
Finding the rule is a transactional use. One question, one rule, one loan, one moment. It helps the person in front of the file, and then it is over.
But compliance work at the level that actually matters is not transactional, and the people who do it know that better than anyone. It is program-level. Is our policy current with the rule as it stands today. Which of our procedures does this Mortgagee Letter touch. Where does our written policy say something the regulation no longer says. Which vendors sit inside a requirement we are accountable for. What evidence would we hand an examiner who asked how we know any of this.
That work is repetitive, high-volume, cross-referencing, and unglamorous. It is holding forty requirements against thirty documents without dropping one. It is precisely the kind of clerical labour these tools are good at, and precisely the kind of judgment work they are not good at, which is why the compliance officer stays in the chair and the tool does the fetching and the first pass.
And nobody is asking for it. Not once in 1,637 retrievals.
The assistants are already doing the harder work, unasked
Here is the part that changed how I read all of this.
Those two facts sit together in a way worth unpacking: each page is read once, but a great many pages get read. The assistant does not fetch one section and stop. It fetches the neighbours, once each, and assembles. In one day of retrievals against this library, the Fannie underwriting sections came down in this order:
B3-1-01 general underwriting → B3-3-2-02 employment income → B3-3-3-01 and B3-3-3-07 self-employment → B3-3-4-06 other income → B3-3-8-01 income documentation → B3-4-3-04 gift funds → B3-4-3-06 assets → B3-6-02 debt-to-income → B3-6-05 monthly debt obligations
That is the entire qualifying chain. Income, then how income is documented, then what else counts, then what is in the bank, then where it came from, then what is owed, then the ratio. You cannot answer a DTI question honestly without walking it, and the assistant walks it.
The correlation is already happening. It is just not being pointed at anybody’s documents.
That is not a technology gap or a competence gap. It is a framing gap: the question being typed is smaller than the work the tool is already performing.
What the correlated version sounds like
Take the eleven grounding queries above and follow each to the question underneath it, the one somebody running a program would actually need answered:
Seen: fnma gift funds Correlated: Our gift funds procedure was written in 2023. Walk it against B3-4.3-04 as it reads today, plus the donor and documentation requirements it depends on, and show me every requirement in the regulation our procedure does not address, with the sentence from each side.
Seen: fannie mae credit report requirements Correlated: Here is our credit policy and a sample of what our credit vendor actually returns. Which of Fannie’s credit report requirements does our process fail to evidence, and which of those is the vendor’s responsibility rather than ours?
Seen: fnma allowable foreclosure attorney fees Correlated: Here is our attorney fee schedule. Which line items exceed the allowable, quote the limit for each, and tell me which of our procedures would need to change to bring us inside it.
Seen: chapter 18 of gnma guide Correlated: Chapter 18 governs delinquency and default. Which of our default servicing procedures implement it, where do our timelines differ from the Guide’s, and which difference creates the most exposure?
Seen: 38 u.s.c. § 3737 Correlated: The VA partial claim is new and VASP is winding down. Which of our loss mitigation procedures still reference the old options, and what specifically changes in each?
Seen: gross up social security fannie mae Correlated: Our income policy permits grossing up non-taxable income at a fixed percentage. Compare that against current agency treatment across Fannie, Freddie, FHA and VA, and flag every place our single rule is wrong for one of them.
Notice what the second version of each has in common. It names your documents, not just the rule. It asks for a comparison, not a retrieval. And it produces a work list rather than a passage.
That last one is worth sitting with, because a single internal policy applied across four agencies with four different treatments is exactly the kind of quiet, long-running exposure that never shows up in a lookup and always shows up in an exam.
Four real files, built from the sections you cite most
The pages AI assistants cite most from this library are not the easy ones. They are the awkward ones: leasehold estates, notes receivable income, temporary leave, gifts, borrowed funds, Community Seconds, offer letters, non-citizen eligibility.
Those sections do not appear one at a time in real life. They collide in the same file, which is precisely why they are hard. Here are four deals that do exactly that, with the question worth asking about each.
1 · She goes back to work in six weeks, and her father is wiring the down payment. Borrower is on paid maternity leave at 60% of base pay, returns in six weeks, and $14,000 of the down payment comes from her father directly to title. No gift letter yet. Sections in play: B3-3.3-09 temporary leave income, B3-4.3-04 personal gifts, B3-3.2-01 employment and income documentation, B3-6-05 monthly debt obligations.
Evaluate this borrower against each of those four sections in turn. For temporary leave, tell me which income figure qualifies and what documentation of return the section requires. For the gift, tell me what has to be evidenced given the funds went straight to title and no letter exists yet. Quote the sentence behind every conclusion and list what is still missing.
2 · The down payment is a Community Second and the reserves are a 401(k) loan. First-time buyer using a municipal down-payment assistance second, with $9,000 borrowed against her own 401(k) to cover reserves. Sections in play: B5-5.1-01 Community Seconds, B3-4.3-15 borrowed funds secured by an asset, B3-6-05 monthly debt obligations, B5-6-01 HomeReady eligibility.
Does this second meet the Community Seconds requirements as written. Do the 401(k) loan proceeds count as acceptable funds, and does its repayment have to be counted in the ratio. Show me where each answer comes from, and whether the two interact.
3 · Starts the new job in three weeks, and he is not a U.S. citizen. Borrower has an offer letter with a start date after closing, holds an H-1B, and has a thin domestic credit file. Sections in play: B3-3.3-03 employment offers or contracts, B2-2-02 non-U.S. citizen borrower eligibility, B3-5.2-01 requirements for credit reports.
Walk each of those three sections against these facts. What does the guide require before closing on an offer letter, what does it require to document his status, and what does it accept when the credit file is thin. Tell me which requirement is most likely to kill this file and what would cure it.
4 · Leasehold land, builder buydown, and a principal curtailment at closing. Condo on leased land, builder offering a 2-1 buydown plus a closing credit the borrower wants applied to principal. Sections in play: B2-3-03 leasehold estates, B2-1.4-04 temporary interest rate buydowns, B2-1.5-05 principal curtailments.
Confirm the leasehold requirements this project must satisfy, whether the buydown structure is permitted as described, and whether the credit can be applied as a curtailment. Quote the limit for each and flag anything the three requirements together make impossible.
Every one of those is a lookup question a person could ask four or five times in a row and then assemble themselves. The point is that they no longer have to. And note what the fourth one asks for at the end, because that is the part a sequence of lookups can never give you: whether the requirements, taken together, rule the deal out.
Nothing in two days of retrievals looks remotely like any of these.
The capability is already there
The part that makes this a genuine miss rather than a wish is that nothing needs to be built or bought.
A Project is a persistent workspace you load once with your own context: your policies, your procedures, your vendor list, the regulations that govern you. Every conversation inside it starts already knowing your program. You stop re-explaining who you are on every question, which is most of what makes one-off chat feel shallow.
A Skill is a repeatable procedure you write down once so it runs the same way every time, by anyone, with the same steps and the same output shape. Consistency is the point. A gap analysis that is done differently each quarter is not a gap analysis, it is an anecdote.
Put the two together and the shape changes completely. Instead of what does Fannie say about gift funds, the question becomes here is our gift funds procedure, here is the current section, walk them against each other and show me every place our document says something the regulation does not, with the sentence from each.
Same assistant. Same regulations. Different unit of work: a program instead of a loan.
It is the same shift the data shows people have not made yet. Today the tool is being asked to locate text. Nothing stops it being asked to hold two documents against each other and report where they disagree, which is the actual job.
What that looks like in practice
Three that are worth the setup, in order of how quickly they pay for themselves:
Currency check. Take one policy. Take the governing section as it reads today. Ask for every place the policy states a requirement the regulation no longer states, and every place the regulation states one the policy is silent on, with the sentence from each side. Do it for the policy most likely to be examined first, not the easiest one.
Change-to-document mapping. A Mortgagee Letter drops. Instead of reading it and hoping you remember which documents it touches, ask which of your procedures contain the affected requirement and what specifically would change in each. The answer is a work list, and it takes minutes rather than an afternoon.
Evidence assembly. For a requirement you are accountable for, ask what documentation would demonstrate compliance to someone who does not trust you, and what you would be unable to produce today. The gaps are the useful output.
Each one ends with a person deciding. The tool produces a draft with the source sentence attached to every claim, so it can be checked in seconds rather than taken on faith. That is the arrangement I would insist on if it were my licence.
And if none of this sat well with you
There is an obvious discomfort in this piece and I would rather name it than let it sit there.
I opened by telling you I can see how the industry asks. That should give a compliance officer pause, because your research pattern is itself information. A servicer whose people spend a fortnight pulling dual-tracking, loss mitigation timelines and appeal windows is telling anyone watching something about where they think their exposure is. Nobody wants their reading list to be a leading indicator.
So here is exactly what I can and cannot see, and then how to switch it off.
What I can see: which pages an assistant fetched, and the questions Microsoft publishes as having earned a citation. That is it. What I cannot see: who you are, what you typed, what file you were working on, or what the assistant told you. Every number in this article came from aggregate page and query counts, which is why it is about the shape of demand and never about a person.
How to switch it off entirely: download the corpus from Claude for Compliance and run it locally.
3,127 regulations, 7,685 obligations, free, no signup, no licence to accept. Download it, load it into a Project alongside your own policies, and every question you ask runs against a copy that sits inside your environment.
I never see the fetch, because there is no fetch. No page view, no query, no pattern for me or anyone else to read. Your research stays yours.
And it is the better setup for the governance work anyway, for reasons that have nothing to do with privacy:
- The whole rulebook is in context at once. An assistant working from a local copy can hold every relevant section against your documents simultaneously, instead of deciding which single page to go fetch.
- Nothing depends on retrieval luck. No wondering whether it found the current section, an archived copy, or a lender’s blog post about the section.
- It is reproducible. The same question against the same downloaded file gives you the same grounding next quarter, which is what makes a gap analysis a control rather than an anecdote.
I would rather tell you this than have you work it out later and wonder what else I had not mentioned. The library is more useful to me being used than being watched.
The honest caveats
Use your own judgment about what goes into the tool. Anything you paste into a third-party assistant leaves your control. Your firm has a position on that, or needs one. Test the method on hypotheticals before it touches anything real.
The output is a draft for review, not a decision, and never a filing. Regulations are prose, not checklists. An assistant inferring a requirement list from a paragraph will sometimes skip one. That is why every conclusion should arrive with the sentence it came from.
Tell it to flag what it cannot source. Instructed to say when it cannot find a basis, an assistant will say so. Not instructed, it fills the gap with something that reads well. That single sentence in a prompt is most of the difference between a draft you can check and one you cannot.
Currency is still on you. This library mirrors published text and stamps every page with the date it was captured. It does not yet tell you whether a Mortgagee Letter has amended a section since. Confirm at the source before relying on it. That is the honest state of it, and it is what I am building next.
What I will be watching
Right now, nothing reaching this corpus looks like program-level work. Every retrieval is a rule being fetched. I will report that number again when I have another month of it. If it moves, I will say what moved.
If you try any of the three above on something real, I would like to hear where it fell short. The failures are more useful to me than the successes.
Download the corpus from Claude for Compliance and run it locally. 3,127 regulations, 7,685 obligations, free and ungated, nothing to install: Excel · CSV · source files · or browse it online first.
Everything here is a working draft for your own compliance and legal review, not legal advice.