Every mortgage operation runs on one rule, and you will not find it in any policy manual: push every task down to the cheapest person who can do it. Processors do what underwriters shouldn't have to. Junior underwriters do what senior underwriters shouldn't have to. Loan officer assistants do whatever the loan officer has decided is beneath them this week.
Cost per loan is a staircase, and the staircase has exactly one design principle: the step below you is cheaper than you are.
The rule has survived every technology wave this industry has bought. The LOS, the POS and the digital mortgage, whatever that turned out to mean, each just changed who stood on which step.
AI broke it. The cheapest person who can do the work is now a model, and it does not go home at five.
The numbers come from the company that has run this experiment further than anyone. In October, Anthropic published a study of its own workforce. By August, the AI agents its employees directed were logging six hours of work for every human hour. Code commits per engineer had quintupled in six months. And then the finding that should keep every ops leader awake: even non-technical staff would need half their work week just to read what their agents produced. The agents themselves sat idle 80% of the time, waiting for a human to look.
Read that again with a loan file in mind. The work is no longer the bottleneck. The review is.
The other half of the story is the one almost nobody is pricing in. Over the same stretch, junior hires fell from 27 percent of Anthropic’s technical hiring to 9 percent. Anthropic is careful to call that a hiring choice rather than layoffs, and a correlation rather than proof. That is the "cheapest-person rule", applied to AI, deleting the bottom of the staircase. Which would be fine, except the bottom of the staircase was never just cheap labor. It was the place where people learned what a bad file looks like before anyone trusted them to say so.
The question that built your existing org chart was "who is the cheapest person who can do this work."
The question that will build the next one is: "who is the cheapest person who can tell when the machine got it wrong, and how do we make more of them."
Different questions produce different organizations. This piece is about the second one.
The mortgage industry has done what it always does when something new shows up: it wrote a framework. Three of them, actually, within about ten months. Freddie Mac's Guide Section 1302.8 went live in March. MISMO released FRAME in June. Fannie Mae issued a lender letter in April, then superseded it in October with a new Section 4 of its Information Security and Business Resiliency Supplement, effective January 18, 2027, and renamed the whole topic "Super Intelligence" while assuring everyone the rename changes nothing. Only in mortgage does a governance document get a marketing refresh.
All three do the same sensible thing. They tier every AI use case by risk, and the tier sets how much testing, validation and monitoring you owe. FRAME's tier table is the clearest. A system is Critical when it "directly produces a binding credit decision." It drops to High when there is "human review present but output carries material weight." It drops again to Medium when it "influences decisions with meaningful human oversight."
What moves the tier is the human. Your income-calculation AI is High rather than Critical because somebody looks at its output. Your document classifier is Medium rather than High because somebody oversees it meaningfully. The entire risk structure is load-bearing on a person.
FRAME describes that person in exactly one place, a field in its inventory workbook called “Reliance on Output and Human Oversight.” It is a dropdown with four values: Required, Optional, None, Unknown. The implementation guide adds an optional scale (Informational, Decision Support, Primary Driver, Automated Decision) and asks for "an honest assessment of actual usage, not stated policy." Then it offers its example of good practice: "Underwriters typically follow the score unless exceptions are triggered."
That is the whole reviewer. A dropdown, and a sentence describing someone who agrees with the machine.
The Governance Policy template lists nine roles with named responsibilities, from the Board down to the technology team. Not one of them is the person who looks at AI output on a file and decides whether it is right. Accountability is "individual and named," which sounds rigorous until you notice it is named at the system level, in a Business Owner, and never at the decision level. The training section covers employees "with AI governance responsibilities," meaning training on the policy itself. Nothing in any of the three documents says what a reviewer has to know, how long they get per item, or what happens to the tier when they stop looking.
To be fair to MISMO, this is version one and it is better than what most lenders had, which was nothing and a prayer. The tiering is sound. The gap is not a drafting error. It is the industry's shared assumption that human review is a thing that just exists, like gravity, and does not need to be engineered.
It needs to be engineered, because the default human is persuadable. In a separate Anthropic study published October 27, researchers gave professional analysts an AI agent that had been secretly instructed to favor certain analytical choices. Participants followed the agent's preference 86% of the time. When they had already stated their own view and the agent disagreed, they abandoned their view 60% of the time. Most of them later explained their decision using the agent's reasoning, and only 15% said the reasoning came from the AI.
Those were trained analysts with time to think. Your underwriter has forty files in the queue and a turn-time dashboard glowing red. Most lenders’ policies put their AI at Decision Support, meaning the underwriter decides and the AI advises. The honest answer FRAME asks for is usually Primary Driver: the AI decides and the underwriter ratifies. Write that down and the tier table turns on you.
The human review that kept your income AI at High risk was the only thing keeping it out of Critical, and you have just admitted the human is a formality. It was a Critical-tier system all along.
Picture the underwriter again. The AI has already calculated income, flagged two conditions, and recommended an approval. Her job is to confirm it. She has the file open for ninety seconds, the queue has forty-one more behind it, and the turn-time dashboard the COO installed last year is tracking her by the hour. She approves. She approved the last six the same way.
She is responding rationally to the incentives everyone handed her. Management measures the team against what the AI can produce, so a human who slows the line looks like a defect. The reviewer sees a queue that never shrinks. The technology team sees an expensive agent spending most of its day waiting for a click. Everyone's complaint is legitimate, and together they produce one outcome:
Approve more, inspect less, keep it moving.
On the AI Riff podcast, one of the hosts put it plainly: "I guess it's right. I don't have time to check it." That sentence is the operating procedure at more lenders than will admit it.
Call it approval theater. The approval step still exists. There is a person, a timestamp, a documented procedure, and a control that passes every audit. What is missing is the judgment. The reviewer is accepting a recommendation because the volume makes serious scrutiny impossible, and the control's value has quietly gone to zero while its paperwork stays perfect.
Your reporting cannot see it. Your governance inventory shows that a reviewer exists. Your post-close QC shows a defect rate. Nothing in between connects the two. You cannot tell whether your income AI's 2% defect rate was caught by the reviewer and fixed, or sailed through because the reviewer approved everything. In the first case the control works. In the second, the review your tiering depends on is not happening.
One number exposes it and you're probably not computing it:
Override rate per reviewer, measured against the known defect rate of the task they are reviewing.
A reviewer who overrides the AI on 3% of items in a category with a 3% defect rate is working. A reviewer at zero in the same category is a stamp. Zero is the worst possible number, and it is the number most lenders would find if they looked.
The industry has a long tradition of not looking. We did not audit every loan in 2021 either; we sampled, because there was no physical way to do otherwise, and the industry survived. The difference now is that nobody has decided what the sample is. The AI's review queue defaults to everything, the human cannot do everything, and so the sample becomes whatever gets a real look before the reviewer's attention runs out. That is not a control. It is a lottery with an audit trail.
Ask a senior underwriter what she knows that a new hire doesn’t. The answer won’t be the Selling Guide, since the new hire can read that. It will be something like: I know what a self-employed borrower's income looks like when the business is quietly dying, even when the Schedule C passes the math. That knowledge did not come from a course. It came from a few thousand files, a few hundred corrections from someone senior, and a handful of loans that went bad with her name on them.
That is production expertise. Build the answer from the documents and the rules. It took a decade to grow and the industry has been running short of it for years. As one of the AI Riff hosts pointed out, how many underwriters can actually do a clean manual underwrite anymore? Not many. The ones who can are not just competent. They are trusted, and there are fewer of them every year.
The frameworks don't need production expertise from most of your reviewers. They need something else. FRAME's System Risk Assessment asks for "Observed Model Limitations," defined as the "time periods, loan types or other pertinent data where the model does not perform at a desired level." Nobody can fill in that field from guideline knowledge. You fill it in by having watched the model fail, a few hundred times, with the evidence in front of you.
That is detection expertise: knowing what wrong looks like, where this particular model is weak, when the file holds something the model could not see, and when to stop the line. It is closer to the skill of a QC auditor or a model validator than to the skill of an underwriter, and it is the skill the current org structure quietly assumes you have on staff.
You do not have it on staff, because nobody has ever had a reason to grow it. And you cannot hire it.
Anthropic's hiring managers now screen for judgment over domain skill; one of them said he wants people who are "fungible across domains" because he expects models to stay jagged. But Anthropic is importing judgment from a labor market that still produces it the old way. A mortgage lender cannot import two hundred people who know where its income model breaks on a declining-trend Schedule C. There is no such labor market. The lender has to grow them, and it has to grow them from a starting position of fewer entry-level jobs, not more.
The AI Riff podcast crew named the consequence before anyone had a framework for it: "When you're looking to fill the positions and you're looking for five years experience, nobody's got five years experience because they can never get started." The AI Riff podcast call this judgment debt.
Every time you automate work without preserving a way for people to learn to evaluate the automation, you borrow against your future capacity to say no.
The savings show up this quarter. The bill arrives when the last trusted underwriter retires and the person replacing her has only ever clicked approve.
None of this means recreating the old apprenticeship. Nobody is going to pay a junior to calculate income by hand for three years so they can eventually check a machine that already did it. What’s needed is a different apprenticeship, one whose curriculum is the model's error distribution instead of the guidelines, and which produces detectors faster than the old one produced underwriters. A small bench of true production experts still has to exist at the top, for the manual underwrite and the exception nobody has seen before. Everyone else needs to be good at detection.
The old staircase sorted work by how cheap the person was. The new structure sorts work by how likely the AI is to be wrong and how much it costs when it is. There are three tiers, and each one is a job.
Tier 1 is high-accuracy, reversible work: document classification, W-2 income, VOE reconciliation, condition prediction on a clean file. Nobody reviews all of it. Juniors sample it, the way QC samples loans, at a rate set by the defect rate your own post-close QC traces to the task, not by a vendor’s confidence score.
Tier 2 is anything that reaches a borrower or feeds a decision: self-employed income, asset sourcing, the condition list that goes out, the adverse action reason codes. Every item gets a human, and the human is a junior or a mid-level reviewer working inside a surface designed for the job. The design of that surface decides whether any of this works.
Tier 3 is the exceptions and the governance act. A small senior bench owns it. They review the escalations nobody below them could resolve, and they do the one thing that actually compounds: deciding, on QC evidence, which task categories have earned a lower tier.
The reason the junior rung disappeared is that reviewing AI output felt like clicking approve, and nobody learns anything clicking approve. Making juniors redo the work would fix that at double the cost. The cheaper fix is to change what the screen asks of them, so that the fastest path through the queue is also the path that builds detection.
From the junior's seat, it looks like this. The pay stub and the tax return come up first, with the fields the AI used already highlighted. The AI's number comes up second, after she has seen the evidence, because screen order is the cheapest defense against the persuasion problem. To approve, she clicks the highlighted fields, not the conclusion; she is asserting the inputs, and the file records that a named person asserted them. Overriding the AI is one click and a reason code, cheaper than agreeing. Roughly one item in fifty is a plant, a known-wrong AI output seeded by the system, and her catch rate on the plants is her performance review. Airport security has used a version of this for years, projecting fictional threat images into screeners’ X-ray feeds, and it works only because screeners can’t tell the plant from the real thing. When post-close QC finds a defect on a field she asserted, it lands back in her queue within days, with the correction, from a loan she remembers.
None of that costs the CFO a second pass. It is seconds per item on work that used to take a processor minutes, and the lender gets a catch rate per reviewer per task category for free. That number is the evidence you hand Fannie when they ask about safeguards, the promotion gate between tiers, and the first thing that would have told you approval theater had set in.
The most expensive person in the building stops reviewing files. Reviewing files is a volume job and it does not scale, which is the whole lesson of the six-to-one number. Instead the senior bench reviews the machine's error distribution. Monthly, they look at the seeded-defect catch rates, the override patterns, the QC defects by task, and they vote: document classification has run four quarters at a defect rate below threshold with the reviewers catching what little there was, so it moves from Tier 2 to Tier 1. Self-employed income stays where it is. Something new drifted and goes up a tier until someone understands why.
Tier migration is the ROI line. Every task that moves down a tier is cost per loan that stays down through the next cycle, and the lender can show its regulator exactly why it was safe to move it.
This is also the only version of "senior underwriter" that a 30-year-old with a strong catch rate can grow into without a decade of manual underwrites behind her, which matters, because the industry is not going to get that decade back.
Every lender knows the cycle. Volume halves, the senior underwriters go, volume returns, the lender rehires them at a premium from whoever poached them. The pyramid changes which layer flexes. Tier 1 and Tier 2 scale with volume, and they are staffed by the people who are cheapest to add and, frankly, cheapest to lose. The Tier 3 bench is small, and it is now the governance control the GSEs require you to maintain whether or not you are funding loans this quarter. You stop firing the last person who can say no, because her review is what makes your Tier 1 safe to run.
You do not need a transformation program. You need five things, and two of them cost nothing but honesty.
What changes by channel is less than you would think. A credit union with a compliance person and a bank with a compliance division need the same five things at different scale. Wholesale has one extra problem: the broker is a reviewer you do not employ, and the tier math has to assume the broker's package arrives at Tier 2 at best. Correspondent buyers already run a version of this and call it due diligence.
The industry's reflex when a framework lands is to comply with it and move on. That reflex will produce lenders with immaculate AI inventories, named Business Owners, annual policy reviews, and underwriters who approve everything because the queue made them. They will be fully compliant and fully exposed.
The lenders that come out of this ahead will be the ones that noticed the frameworks left a person-shaped hole and decided to fill it on purpose. They will have reviewers whose catch rates are known, tiers that move on evidence, and a senior bench that spends its time making the machine safer to trust rather than re-checking its arithmetic. They will also, almost as a side effect, have an entry-level job again, and in five years they will be the only ones with anyone to promote.
The article this grew out of ended with the observation that the organization with stronger governance may ultimately move faster, not slower. That is true but incomplete, because governance is paper.
What moves faster is a lender that knows, per task and per person, who can tell when the AI is wrong. The cheapest-person rule never asked that question. The right person was always the one who could say no. You used to get those people for free, as a byproduct of the staircase. Now you have to build them.
Sources: Anthropic, How employment is changing at Anthropic, October 16, 2026. MISMO FRAME v1.0 Governance Policy, System Risk Assessment, AI System Inventory and Implementation Guide, June 2026 (MISMO member materials). Fannie Mae Supplement Bulletin 26-01 and Section 4, effective January 18, 2027, via BrokerLumen's summary. Freddie Mac Guide Section 1302.8, effective March 3, 2026. AI Riff podcast, episodes on the Anthropic employment study, and "The Last Person Who Can Say No."