Learn more →
Back to WorkCase Study · Credit Due Diligence

Franchise Durability Diligence
on a U.S. Mortgage Lender

A three-week, interview-led diligence for a credit investor weighing a structured investment in a large U.S. mortgage lender. The qualitative question: will the franchise hold? We ran 25 expert interviews on one shared scoring rubric, supported by an AI-native operating system that turned every call into comparable evidence.

ClientConfidential
Expert Interviews25
Workstreams4
Client Check-insWeekly
Timeline3 Weeks
The Challenge

Is the franchise's strength structural, or is it bought?

The lender runs a broker-only model and has led its channel for years. It had also paid out aggressively to shareholders, taken on more leverage, and drawn rating-agency scrutiny. A credit investor weighing a structured investment needed to know whether the lender's position rests on real advantages (scale, efficiency, genuine broker loyalty) or on pricing and incentives that a stretched balance sheet may not keep funding.

The mandate was strictly qualitative, with no forecasts and no projections. Every judgment had to be an up, down, neutral, or watch read, backed by evidence the investment committee could interrogate. Because the target is public, the work was also conducted under strict information barriers.

The work split into four workstreams. I owned two of them, technology & AI and regulatory risk, covering the desk research, interview content, synthesis, and slides.

01

Competitive Landscape & Broker Relationships

Why do brokers place loans here, and what would make them leave?

02

Technology Platform & AI

Led

Is the technology edge durable, or is AI about to commoditize it?

03

Servicing & Refinance Recapture

Does owning the servicing book help the franchise, or erode broker trust?

04

Regulatory, Litigation & Industry Risk

Led

Could a regulator, a court, or a headline change broker behavior?

How the Engagement Ran

Hypothesis-first, on a tight operating rhythm

We wrote straw-man hypotheses on day one, one per diligence question, and every interview after that either strengthened or weakened a specific read. The rhythm kept the hypotheses honest.

Daily · AM

Standup

A prioritized task list for the day, assembled before the call from the calendar, inbox, team chat, and yesterday's carryover.

Daily · PM

Check-out

What closed today, what carries to tomorrow, and which interview answers changed a hypothesis.

Twice weekly

Partner problem-solving

Thirty minutes with the partners to pressure-test the storyline, cut pages, and decide what the evidence can and cannot yet support.

Weekly

Client check-in

Answer-first readouts with an up, down, neutral, or watch read on each diligence question. The first one fed the client's investment committee.

The Interview Process

From a blinded screener to a single source of truth

Primary research was the backbone of the diligence. Over eight business days we ran 25 interviews, and each one passed through the same six-step pipeline, so a working broker's answer and a competitor executive's answer ended up in the same workbook, on the same scales.

01

Front-load the desk research

Before booking a single expert, we mapped the origination workflow end to end, inventoried every competitor's technology stack, and built a litigation and regulatory dossier. That let interviews start at the second-order questions instead of spending twenty minutes on how the industry works.

02

Source through a blinded screener

A screener went to two expert networks, plus direct outreach to working brokers. It described a market study and never named the target. It set quotas across five segments and asked specifically for people who had left the target within the last year.

03

Write guides for the segment, then the person

Each segment got a general guide, and each high-value expert got a tailored overlay on top of it. Guides came in two versions: a clean, shareable Word document for the interviewer, and a richer internal prep with hypotheses and follow-ups.

04

Capture on one rubric, live

Every call ended in the same structured answers: a 100-point allocation across the criteria that drive placement, a 1–5 score per lender, and a Yes / Partial / No chain on AI. Everything was captured during the call or right after it, in the same order and on the same scales.

05

Clean, attribute, and verify

Recordings became house-format transcripts with a speaker roster, topics, and timestamps. Every score that reached a slide was checked line by line against the raw transcript. Scores an interviewer had prompted were cut from the numbers.

06

Roll up into a single workbook

One master workbook held a tab per expert and an aggregate layer. Heatmaps, distributions, and quote banks were all generated from it, so every chart traces back to a named call.

Who We Spoke To

Five segments, each one tagged for bias

Many brokers are contractually exclusive to one lender, so they can't rate its rivals. We deliberately prioritized non-exclusive brokers and filled the gap with competitor executives who see the whole field.

SegmentCallsWhat They Told UsRater Tag
Working mortgage brokers~14The customer. Placement drivers, lender ratings, switching behavior, and how their own AI use is changing the job.User · Competitor's user
Competitor lender executives~6Cross-lender benchmarks that exclusive brokers couldn't give, plus market-share and pricing dynamics.Neutral
Technology & servicing operators~3What is buildable, what can be bought, and how far AI has really moved each step of the loan journey.Neutral
Former target employees~2How the platform, the broker incentive programs, and the sales motion work from the inside.Insider · discounted
Regulatory specialists1+Federal versus state enforcement, private litigation, and fair-lending exposure from AI underwriting.Neutral
Interview Design

Six rules that made opinions comparable

Expert calls are easy to run and hard to aggregate. Most of what made this program work was decided before the first call and enforced after the last one.

01

Every comparison ends in a number

“Their tech is better” is not evidence. Every comparative answer had to finish with a score, a reason for that score, and a hard anchor such as days from submission to clear-to-close, clicks to submit a file, or resubmissions per loan. If two answers can't be compared at least semi-quantitatively, they don't go on a page.

02

Don't lead the witness

The first pass of guides told experts the thesis. The rewrite opened every block with an open question and saved the hypothesis for the follow-up. The same test applied after the call: one expert's composite score had been built by the interviewer rather than given by the expert, so it was excluded from the numbers and used only for color.

03

Tag who is talking

A broker who sends the lender half their volume rates it differently from one who left it. Every rater was tagged as a user, a competitor's user, or neutral. A finding counted only if it held among neutral raters, and a former insider's praise was discounted heavily.

04

Ask about outcomes, not features

Asking brokers to score a dozen technology capabilities produced noise. We rolled the capabilities up into three jobs (win the deal, execute it, run the workflow) and asked about each one's outcome in the broker's own pipeline, plus one question on switching.

05

Keep pricing out of the tech score

Preferential rates earned through gamified rankings are pricing. Marketing collateral is table stakes. Balance-sheet-funded incentives are capital. Pulling all three out of the technology score was the only way to answer whether the software itself was the moat.

06

Benchmark against what anyone can buy

Besides named competitors, experts also scored the best third-party stack a lender could license today. The target's score minus that buyable-stack score gives a numeric estimate of how much of the edge can't be replicated.

Measuring the Moat

Target minus buyable stack = the part that can't be copied

Each expert scored every lender from 1 to 5 on each capability, with a reason and a hard anchor. The last column is the key one: the best stack any lender could license off the shelf today.

When the target's lead over a buyable stack is one point or less, the advantage can be rented. The rows where the gap held up pointed us away from software and toward distribution and capital.

Illustrative scores shown. Actual figures are confidential.

TargetCompetitor ACompetitor BBuyable stack
Speed to clear-to-close5433
Underwriting automation5434
Portal & workflow UX5433
Pricing & product tools4434
Broker-facing AI4523
Borrower-facing AI3322
The Operating System

An AI-native workspace behind a traditional diligence

I ran my workstreams out of a Claude Code workspace built for the engagement. A single router file described the project, the team, the folder map, and the house conventions. A set of slash-command routines handled the mechanical work every day, so the hours went to thinking. With 25 calls in eight days, memory stopped being the bottleneck.

/check-inEvery morning

Reads the calendar, inbox, team chat, and task log, and returns a ranked list for the day before the standup.

/check-outEvery evening

Records what was finished and the small amount carrying over, and updates the living task log.

/clean-transcribeAfter every call

Turns a raw recording into the house transcript format, with speakers attributed from a roster. It only cleans up; it never interprets.

/catch-upOn re-entry

Scans the newest working documents and reconciles them into persistent project memory, so no session starts from stale context.

/researchOn demand

A structured desk-research routine for the technology and regulatory questions, with every claim sourced and time-sensitive figures flagged for re-checking.

/build-deckOn demand

Generates on-brand slides from the firm's template: action titles, source footers, and read chips.

Evidence stays traceable

Raw transcript → cleaned transcript → workbook tab → heatmap → slide. Any number on a page can be walked back to the line in the call where it was said.

Memory doesn't go stale

A distilled knowledge base carries a synced-through date. Reconciling new documents into it is a routine, not an afterthought, so each session picks up exactly where the last one ended.

Judgment stays human

The transcript routine is forbidden from interpreting. Synthesis, reads, and storyline calls happened in the standups and problem-solving sessions, with AI doing the assembly work around them.

Key Findings

What the evidence said about technology and regulation

Generalized, industry-level reads from the two workstreams I led. Target-specific findings remain confidential.

01

Technology was a lead, not a moat

Brokers rated the target's platform best in class, and neutral raters agreed. But when they allocated 100 points across what actually drives where a loan goes, technology ranked well behind speed, price, and service. It reduces friction; it doesn't win the loan.

02

Switching cost holds the share, not the software

The stack is increasingly buildable or buyable: a licensed loan-origination system, AI underwriting, and a pricing engine now replicate most of it. What keeps brokers in place is habit and workflow lock-in. Brokers also tend to move new business before old, so share erodes gradually rather than all at once.

03

AI's productivity lift is broker-led and front-of-funnel

The real gains came from brokers using general-purpose AI for marketing, lead generation, and product comparison, and some had built their own agents. Adoption of the lender-built AI tools was low. Underwriting, appraisal, closing, and retention were all “not yet.” We found no evidence that AI had grown loan volume.

04

Regulatory risk is migrating, not melting

Federal enforcement has receded, but the risk is moving to state attorneys general, the private plaintiffs' bar, and fair-lending scrutiny of AI underwriting. Brokers' placement behavior hadn't moved on litigation headlines, so the risk is to relationships and reputation more than a legal one.

05

Durability routes back to the balance sheet

If the software is replicable, the edge comes from distribution and the capital that funds broker incentives. That turned the diligence's central question, whether the franchise's strength is structural or bought, into a question about financial capacity, which is exactly what a credit investor prices.

Outcome

Answer-first reads, in time for the investment committee

The first weekly check-in went to the client's investment committee as an answer-first executive summary. It took each diligence question in turn, led with the read, and named the path to validating it. Later check-ins added scored heatmaps, rating distributions split by rater type, and quote-backed pages on AI and regulation.

Behind the pages sat the full evidence base: every cleaned transcript, a master scoring workbook, a quote bank organized by theme, shareable interview guides, and regulatory and technology dossiers. The client could audit any read back to its source.

The method carries over to other deals. Interviews become comparable when each one ends in a number from a tagged rater, verified against the transcript, and benchmarked against what the market can already buy.