The Infinity Loop — Value Model

• $17.1MNet value · 36 months
• Month 5Payback
• $456Value per funded loan

Most mortgage technology is bought on a promise of efficiency and evaluated on a single number: how much faster, how much cheaper, measured once. That framing misses what actually determines whether an automation investment pays. The question is not how much better the operation gets. It is whether it keeps getting better — and who owns the improvement when it does.

This paper describes a system built to compound. Processing confirms which exceptions are real. Underwriting receives files that arrive pre-conditioned rather than assembled. Quality control converts its findings into rules enforced upstream, where they prevent the defect instead of documenting it. Each pass around that loop retires a question the organization has already answered, permanently, for everyone. There is no end state — which is precisely why the economics look nothing like a conventional software return.

To test what that is worth, we built a value model around a representative lender: 15,000 funded units today, growing to 30,000 over thirty-six months, with 175 processors and a $333,000 average loan. The model derives capacity per person from the lender’s own volume and staffing rather than importing an industry benchmark, treats loop maturity as a saturating curve rather than a step change, and excludes anything that would flatter the result. On those terms the loop returns approximately $22.2 million of value over thirty-six months against $5.1 million of platform cost — a net position of $17.1 million, turning positive in month five.

The number that matters most, though, is not the total. It is the shape. Year one nets $1.4 million; year two, $5.9 million; year three, $9.8 million. The return accelerates because the asset appreciates — and the asset is the lender’s own accumulated judgment, encoded.

01  The loop

A loan file passes through three distinct kinds of expertise on its way to funding, and in most lending organizations each of the three throws its knowledge away as it goes.

A processor determines that a flagged name mismatch is an LLC naming convention rather than a discrepancy. The loan moves; the determination evaporates. An underwriter works out how a particular income structure should be conditioned. The file closes; the reasoning stays in that underwriter’s head. Quality control finds a defect pattern across forty files and writes it up. The report circulates, someone schedules training, and six months later the pattern is back — because nothing structural changed.

None of this is a failure of the people involved. It is a failure of retention. The expertise is generated correctly and then discarded, over and over, at industrial scale.

Orchestration closes the circuit. When a processor confirms or dismisses an exception and documents why, that judgment becomes a rule. When an underwriter’s condition logic is encoded rather than remembered, the same fact pattern produces the same condition on every file regardless of who holds it. When quality control identifies a defect, the finding is enforced at processing — before the file ever reaches underwriting again.

The result is a system with no natural stopping point. Each pass makes the next loan marginally cheaper to manufacture and marginally harder to get wrong, and those margins accumulate against a base that is itself growing.

02  One curve, not three scenarios

Capacity models in this industry are usually presented as scenarios: conservative, moderate, aggressive — three futures, pick one, hope for the middle. That framing is a category error when the underlying mechanism is cumulative learning. Conservative and aggressive are not different futures. They are the same future at different times.

The model treats loop maturity as a saturating curve — m(t) = 1 − e^−t/τ — calibrated so that the observed improvement milestones fall naturally along it. Approximately +10 percent capacity appears around month six. Roughly +28 percent arrives near month twenty-four. The +45 percent figure is not a best case; it is the asymptote the curve approaches as the rule set stops encountering patterns it has not seen before.

This reframing changes the buying conversation in a specific way. A lender evaluating three scenarios is being asked to bet on which one comes true. A lender looking at one curve is being asked a different question: how fast do we want to move along it? That is a question about their own discipline — comment quality, rule-refinement cadence, how quickly monitoring extends past the closing document set — rather than a question about whether the technology works.

It also explains why the model’s most interesting periods are not the early ones. The curve is steepest in the middle. An organization that measures results at month six and stops has sampled the flattest, least representative part of the return.

03  A number the model derived, and the floor confirmed

One design decision in the model deserves specific attention, because it produced the only genuine validation available so far.

The model refuses to assert an underwriting capacity gain. Asserting one would be the easiest thing in the world and the least credible. Instead it derives the figure from two operational inputs: the share of an underwriter’s file time consumed by clerical work rather than judgment, and the proportion of that clerical work orchestration removes. At 40 percent clerical share and 80 percent of it removed, 32 percent of the underwriter’s time is returned — which converts to a throughput gain of approximately 47 percent at maturity.

• +47%Derived by the model
• +50%Observed in production
• 40%Clerical share assumed

That derivation was performed without reference to observed underwriting data. Separately, a lender running the platform in production reported underwriter throughput moving from approximately 50 loans per month to 75 — a 50 percent gain.

A model that derives 47 percent from first principles and then encounters 50 percent in the field is not proof of anything on its own. But it is the kind of corroboration that should raise confidence in the surrounding structure, and it suggests the clerical-share assumption — the one input most likely to be challenged — is conservative rather than optimistic.

04  Where the value comes from

Capacity is the largest component of the return but not the only one, and the smaller components are worth understanding because they are where most efficiency models quietly inflate themselves.

That last item is the tell. At current rates the lender is in positive carry: a loan on the warehouse line earns more than it costs to hold, so faster delivery forfeits interest income. The model subtracts that forfeited carry, leaving only the genuine benefit — released line capacity and freed equity. A model built to sell would have shown faster delivery as pure savings and booked several hundred thousand dollars that does not exist under today’s rate structure. Netting the carry costs the result money and makes it defensible.

05  What it adds up to

Platform cost over the same horizon totals $5.11M. Value per unit nearly triples between year one and year three. That is the signature of a compounding asset and the single most important thing to understand about evaluating it: a pilot measured over two quarters will systematically understate the return, because it samples only the flattest part of the curve.

The loop returns approximately $456 per funded loan at maturity. Against the $727 of pre-tax net production profit per loan that independent mortgage banks earned in the first quarter of 2026 — on roughly $11,898 of production cost, a margin of about 16 basis points — that is a 63 percent lift on the bottom line of a loan. At margins that thin, per-loan improvements are heavily levered.

06  What this model refuses to claim

Value models in this industry have a credibility problem, and it is earned. The most useful thing this one does is decline several opportunities to overstate itself. Those declines are listed here so they can be checked.

• It does not reduce anyone’s team. Every capacity benefit is expressed as hiring avoided against future demand, never as positions eliminated. An operation that is not growing gets a materially smaller number from this model, and that is correct.
• It does not import an industry throughput benchmark. Capacity per person is calculated from the lender’s own trailing volume divided by their own staffing.
• It does not assert the underwriting gain. The figure is derived from clerical time share and clerical work removed, so the assumption is visible and arguable rather than buried.
• It excludes per-loan incentive pools. Incentives scale with volume rather than headcount, so they are identical in every scenario and cannot influence the decision.
• It subtracts positive carry. Faster delivery currently costs interest income, and the model books that cost against itself.
• It marks its speculative inputs. Underwriting headcount and salary, clerical share, defect and repurchase rates, suspense behavior, cost of equity and several secondary-market assumptions are estimates from public benchmarks, not lender-verified figures.
• It does not display a price. The model reports the value ceiling per unit — what the loop returns — rather than what it costs. Price is a conversation, not a screenshot.

07  The asset you own

Strip away the projections and the argument reduces to a question about ownership. Most lenders today rent their automation intelligence: the vendor’s rules, the vendor’s thresholds, the vendor’s definition of what counts as an exception. The arrangement works until the contract ends, at which point years of accumulated operational learning walk out the door with the software.

The loop inverts that. The rules are built from the lender’s own confirmed exceptions. The workflows are designed by the lender’s own operations leaders, in a visual designer based on Business Process Model and Notation, so the people who understand the process shape the logic rather than describing it to someone who will approximate it. The condition patterns come from the lender’s own underwriters. What accumulates is not the vendor’s product knowledge. It is the lender’s institutional memory, in executable form.

That is what makes the return curve bend upward instead of flattening. Conventional software delivers its full value on day one and depreciates from there. An asset built out of an organization’s own accumulating judgment does the opposite — which is why the third year in this model is worth seven times the first.

The compounding is the product. Everything else is delivery mechanism.

Methodology & assumptions

Modeled operation: 15,000 funded units trailing twelve months growing linearly to 30,000 at month 36; $333,333 average loan amount; 70% application-to-funding pull-through; 175 processing FTE at $52,000 base; 36 underwriting FTE at $110,000 base; 27% employer burden. Loop maturity follows m(t) = 1 − e^−t/τ with τ = 24 months, adjusted for volume. Capacity per FTE is derived from the operation’s own trailing volume and staffing rather than an industry benchmark. Processing capacity gain at maturity is set to 45%; underwriting gain is derived from a 40% clerical share of file time with 80% of that work removed, yielding approximately 47%; defect reduction at maturity is set to 35%. Warehouse value is computed net of carry, which is positive at the modeled 6.66% note rate against a 5.90% warehouse cost and therefore reduces the benefit of faster delivery. Suspense days are netted from warehouse dwell to prevent double counting. Per-loan incentive pools are excluded as volume-driven and decision-neutral. Benchmark net production profit of $727 per loan is drawn from Mortgage Bankers Association reporting for Q1 2026.

Speculative inputs. Underwriting headcount and salary, clerical share of underwriter file time, defect and repurchase rates, cure cost and loss severity, suspense incidence and duration, lock extension behavior, execution improvement captured, cost of equity, and facility fee are estimates drawn from public benchmarks and market research rather than lender-verified data. They exist so the model runs end to end and must be replaced with actuals before any decision rests on them. This document is an analytical tool and a starting point for discussion — not a proposal, quote, or guarantee of results.