Learn more →
All Work
Agentic Workflow DesignMortgage Servicing

An Agentic Workflow for Escrow Waiver Decisioning

Designed the end-to-end operating process, data model, and workflow configuration for escrow waiver requests at a top-10 US mortgage servicer — a decision governed by two federal investor rulebooks that disagree with each other, one client contract, and federal law. The interesting part was not the automation. It was proving exactly where automation had to stop.

0Rule Sets Codified
0Judgment Calls Isolated
0Messages Configured
0Data Entities Modeled
The Challenge

A decision where getting it wrong is a legal problem, not a conversion problem.

Most homeowners with a mortgage do not pay their property taxes and insurance themselves. They send extra money with each monthly payment, the servicer holds it in an account — the escrow account — and the servicer pays those bills when they come due. An escrow waiver is the borrower asking to stop doing that and pay the bills themselves.

The servicer cannot simply say yes. Property tax liens jump ahead of the mortgage in almost every state, so an unpaid tax bill can force a sale that wipes out the investor's position entirely — a $340,000 loan extinguished over a $6,000 bill. That is the reason escrow exists, and the reason the rules around releasing it are strict.

Three things made this hard to automate. The two investors who own these loans, Fannie Mae and Freddie Mac, publish rules that genuinely contradict each other — the same borrower qualifies under one and is denied under the other. Some bills can never be waived at any borrower quality, because federal law says so. And the workflow tool available to build in could send messages but never receive replies, and could not write back to the system that actually holds the loan.

Scope

A design engagement, not a production deployment. Delivered as a 32-page operating specification, a data model, the evaluation logic in pseudo-code, and a working simulator. Process, rules, owners, and message copy are our proposals; the two investor guides are quoted verbatim.

The Approach

Start from the source text, not from the workflow diagram

The temptation in a regulated process is to draw the happy path first and handle the exceptions later. We went the other way: read both rulebooks verbatim, find every place they disagree, then let the disagreements dictate the shape of the process.

Phase 1
Read the Rulebooks

Worked the two governing guides line by line — Fannie Mae's Servicing Guide B-1-01 and Freddie Mac's § 8201.1 — and wrote each rule down verbatim before designing anything. That produced a divergence table: the exact clauses where the two investors disagree, which is where a single blended process would quietly break the law.

Phase 2
Model the Data

Every rule was traced back to the specific field it reads. Ten entities, grouped into eight dictionary sections, each section citing which rules depend on it — plus a deduplication audit and an explicit list of the fields the servicer does not hold today, with a fallback for each.

Phase 3
Separate Rules from Judgment

Sorted every step into one of two kinds: the data alone decides, or a person must. Converting judgment into arithmetic wherever it was defensible left exactly seven cases that a person still has to own — each with a named owner, a deadline, and its complete set of options.

Phase 4
Configure and Prove It

Expressed the whole process inside the servicer's own workflow tool — datasets, triggers, and messages — then built a working simulator that runs a synthetic borrower through every gate in order and pauses wherever a human decides, so the design could be tested rather than asserted.

The Gate

Six questions, ranked by how much say each one has

Five questions decide whether a waiver is granted. Keeping it comes down to one more. The order matters for a practical reason: the cheapest and most authoritative checks run first, so the one step that costs money — pulling a fresh credit score — runs last, and only on requests that have already passed everything else.

5
The requestCan we act on this at all?Every borrower has agreed and no legal block stands.
1A / 1B
The investorWho owns the loan?Fannie Mae lists disqualifiers — one failed test denies. Freddie Mac grants a permission, with an appeal.
3
The clientDoes the client's stricter policy allow it?A client may tighten the rules, never loosen them.
4
Our own riskCan we carry the risk?Caps on money the servicer would have to front. Free checks first, the paid one last.
2
The billsWhich escrow items can come off?Each escrowed bill is cleared or kept, one by one.
6
The signatureWhose yes is it, and did they sign?The authority holder signs; the borrower signs within 30 days.
Yes · full waiver
Yes · partial
No · denied, with the date
↻ After yes — re-checked every night, until payoffBills paid? Still allowed off? Still visible to us? Any no and we pay the bill, escrow goes back on, and the borrower gets a letter explaining which of those three failed.
The Arguments

Five design decisions the whole process rests on

Each one is a case where the obvious build produces a wrong — sometimes unlawful — answer, and the correct build is only slightly more work.

01

The rulebook forks, and it has to fork in the design

Fannie Mae's guide is a list of disqualifiers — fail any one test and the request is denied, with no appeal. Freddie Mac's is a permission with an alternative: be current, not excluded, and satisfy either an equity test or a payment-history test. A borrower at 85% of original value with six clean months passes under Freddie and fails under Fannie. One blended rulebook would silently deny eligible Freddie borrowers. Forking makes each branch independently auditable and lets each one update when its own guide changes.

02

Every gate returns three answers, not two

Pass, fail, and unknown. Missing data is the servicer's problem, not the borrower's — so unknown routes to research and gets its own letter, never a denial. Collapsing unknown into no is the most common way a servicer denies someone who was actually eligible, for a data problem that was never theirs.

03

Escrow state is per bill, not per loan

Flood insurance in a flood zone can never be waived — federal law, regardless of anything else. Neither can monthly-paid mortgage insurance. So a borrower who qualifies on every test still receives a partial waiver: some bills come off, the protected ones stay. A single on/off flag at the loan level cannot represent that, and would produce an unlawful outcome.

04

A subjective clause became four objective tests

Freddie Mac requires weighing whether a borrower can absorb large one-off charges, and forbids deciding on loan-to-value alone. Rather than an analyst's opinion, the design uses the borrower's own record: 24 consecutive on-time payments including escrow, zero returned payments, shortages repaid in one sum, zero advances made on their behalf. The servicer has already watched them accumulate this exact amount, monthly, for two years. Deterministic, auditable, and it requires no income documents.

05

Judgment moves to onboarding, not decision time

Client waiver policies are validated as typed rows when the client is signed. A policy clause that cannot be expressed as a row fails onboarding and never reaches an analyst. That converts what would be tens of thousands of case-by-case reads a year by rotating staff into a handful of policy decisions a year by senior staff.

Where Automation Stops

Seven judgment calls — the complete list

Every other step decides by itself. These are the cases the rules deliberately do not decide. Each arrives as a message to one named person with the numbers already attached, the full set of options listed, and a deadline that reminds on day two and escalates to their manager on day three. Everything else was either converted into a number or resolved at client onboarding.

Lump-sum review

Passes everything except the capacity test — can this borrower absorb large one-off bills?

Options

Approve · deny · request income documents

Owner

Escrow underwriting

Exception filing

Failed one Freddie Mac rule narrowly, passed every one of the servicer's own.

Options

File the exception under § 8201.1(d) · let the denial stand

Owner

Freddie Mac relationship owner

Late payment review

A late payment sits in the lookback window with no reason recorded.

Options

Classify as servicer error · approved forbearance · declared disaster · borrower fault

Owner

The loan's servicing owner

Client carve-out

The request hits a clause the client kept in prose rather than in numbers.

Options

Apply the carve-out as written · decline it

Owner

The carve-out's named owner

Client approval

Passes everything, and the contract puts the pen in the client's hand.

Options

Sign the approval · refuse it, with the reason recorded

Owner

The client

Unresolvable research

A fact the rules need cannot be established — history shorter than the lookback and held by a prior servicer who does not answer.

Options

Deny on the missing fact · grant and carry the risk · keep digging past the deadline

Owner

Escrow operations lead

Lost visibility

The watcher on a waived bill lapsed while the waiver was live. A waiver the servicer cannot monitor is exposure it cannot cap.

Options

Restore detection through another source · resume escrow for that bill

Owner

Escrow operations lead

The Configuration

Two questions sort every message the process will ever send

The tool does exactly one thing: run a query and send a message. So we listed all seventeen messages the process sends and asked two questions of each — who gets it, and does something have to come back? Two questions, two answers each, four boxes, and every message lands in exactly one. The box it lands in determines which extra fields the configuration needs: a deadline, a chase, a fallback.

This is also the discipline that keeps the system maintainable: a workflow tool fills fields, it never composes prose from data. Composed sentences break and multiply templates. One template covers all seven judgment calls — only the referring rule, its inputs, and the option set change, and all three of those are data.

The heads-up
Our teamNothing comes back

Something happened that a team needs to see. Nobody has to reply. Every automated approval and denial writes one line to the log channel — that channel is the paper trail.

The to-do
Our teamSomething must come back

A person has to act, so the message is their work queue. It chases by itself: a reminder on day two, an escalation to their manager at the deadline.

The notice
OutsideNothing comes back

A final decision, on a compliance-approved template. The denial letter is a notice, and it carries the exact future date on which the answer would flip to yes.

The ask
OutsideSomething must come back

Something is needed from outside by a deadline, and the consequence of the deadline passing is written down in advance. Proof of payment in 30 days, or the servicer pays and escrow resumes.

Capabilities Demonstrated

Designing agents for processes that cannot be wrong

The transferable part of this work is the method, not the mortgage domain. Any heavily regulated process — claims, underwriting, licensing, clinical intake — has the same structure.

Regulated Process Design

Translated two federal investor guides and one client contract into a single ordered decision gate, with each rule tracing to the clause that authorizes it and the data field that answers it.

Agentic Workflow Architecture

Designed the process as a nightly evaluation over data already held, writing one eligibility row per loan. The workflow tool computes nothing — it queries that row and fires a message. Every decision is reproducible from the record.

Data Modeling

Ten entities and eighteen grouped inputs, each group citing the rules that read it, with a deduplication audit and a named fallback for every field the servicer does not hold today.

Human-in-the-Loop Boundaries

Isolated the exact seven points where the rules deliberately stop deciding. Each arrives as a message to one named person, with the numbers attached, the full option set listed, and a deadline that chases itself.

Auditability & Controls

Every stage writes a record: which rule answered, the exact data it read, who decided, what they chose, and a required reason note. Each of the servicer's own rules also answers three questions on file — the harm it prevents, the authority for it, and the number that would prove it wrong.

Build-to-Constraint Engineering

The tool sends but never receives, and cannot write to the system of record. Rather than designing around that, the process was shaped by it — every automated step is a message, and the four steps that must stay manual are named explicitly.

Deliverables

What was designed and handed over

The Process
  • Six-phase decision gate, fully ordered
  • Forked Fannie Mae and Freddie Mac rule sets
  • Seven named judgment calls with owners and deadlines
  • Post-grant monitoring and revocation rules
The Data & Logic
  • Ten-entity data model with rule-to-field mapping
  • Nightly evaluation logic in pseudo-code
  • Three-valued gates — pass, fail, unknown
  • Deduplication audit and data-gap fallbacks
The Configuration
  • Seventeen messages sorted into four shapes
  • Datasets, triggers, and merge fields per message
  • Twelve-week, seven-phase rollout plan
  • Interactive simulator for testing any borrower
Outcome

A process a regulator could read, and an agent could run

The finished design runs the entire decision on data the servicer already holds, re-evaluated nightly, with every rule tracing to the clause that authorizes it and the field that answers it. Approvals and denials go out with no human touch — passing every rule is the yes. Seven cases route to a named person with a deadline. Every stage writes a record of which rule answered and the exact data it read.

The rollout was deliberately conservative: ship with humans deciding, then promote gates one at a time, each running in shadow first — the query decides, a person decides, compare the two — and promote only after a full quarter of agreement. Twelve weeks to live, in seven sequential phases, starting with stakeholder alignment rather than with the build.

The lesson we carry into every regulated build: the value is not in automating the judgment. It is in shrinking judgment to the smallest honest set, naming who owns each remaining piece, and making sure that missing data never gets mistaken for a no.

Have a regulated process that needs an agent?