The letter arrived on a Thursday. A mortgage borrower named Anthony Reeves had submitted a Qualified Written Request (QWR, a written borrower inquiry to a mortgage servicer that triggers specific response obligations under the Real Estate Settlement Procedures Act, or RESPA) three weeks earlier, asking why his escrow account had been increased by $214 per month when the tax assessment on his property had not changed. The servicer's response, drafted by an AI tool and released by a servicing analyst who reviewed it for approximately 90 seconds before sending, explained that the increase was due to a "projected escrow shortage based on updated property tax data." The problem: the "updated property tax data" the AI cited did not exist. No tax assessment had been updated. The actual reason for the increase was a mid-year recalculation triggered by a change in hazard insurance premium, an explanation that was entirely in the servicing file but that the AI had not been provided with, and that the analyst had not checked. Anthony Reeves, who now believed his property taxes had been reassessed, filed a dispute with the county assessor's office, paid $300 for a certified copy of his property records, and then called his servicer angry. That single unverified AI-drafted RESPA response had consumed more staff time, borrower goodwill, and institutional risk than the correct response would have taken to produce.

What Servicing and Dispute Drafting Actually Require

Mortgage servicing is, in operational terms, a correspondence-intensive function. A servicer managing a large portfolio interacts with borrowers through a continuous stream of notices, acknowledgments, responses, and disclosures, each governed by specific regulatory requirements. Understanding the regulatory framework is a prerequisite for understanding where AI can help and where it creates risk if not properly governed.

RESPA and Regulation X. The Real Estate Settlement Procedures Act (RESPA, 12 U.S.C. 2601) and its implementing regulation, Regulation X (12 CFR Part 1024), govern mortgage servicer communications in several important areas. RESPA Section 6 requires servicers to acknowledge receipt of a QWR within five business days and to conduct a reasonable investigation and provide a response within 30 business days. The response must address the specific issue raised in the QWR; a generic or boilerplate response that does not address the borrower's actual question does not satisfy the regulatory requirement. RESPA Section 6 also governs servicing transfer notices, escrow account statements, and error resolution procedures. Each of these communication types has specific required elements and timing requirements.

ECOA and Reg B in servicing contexts. The Equal Credit Opportunity Act (ECOA, 15 U.S.C. 1691) and Regulation B (12 CFR Part 1002) continue to apply during the servicing phase of a loan. A servicer that makes decisions about forbearance, loss mitigation, or loan modification must apply consistent criteria and must not apply those criteria in a way that produces disparate outcomes for borrowers in protected classes. An AI tool that drafts loss mitigation decline letters using language that systematically differs in tone or substance across borrower groups creates ECOA risk in servicing, not just in origination.

CFPB mortgage servicing rules. The CFPB's loss mitigation rules, implemented through Regulation X, impose specific procedural requirements on servicers before a servicer may initiate foreclosure, including requirements for early intervention outreach, loss mitigation application acknowledgment, and written notification of loss mitigation decisions. These communications have specific content requirements that differ from general business correspondence standards. An AI tool that produces fluent servicer responses that do not contain the required CFPB-specified content elements is producing noncompliant output regardless of its professional quality.

Unfair, Deceptive, or Abusive Acts or Practices (UDAAP). The UDAAP standard, under the Dodd-Frank Act and enforced by the Consumer Financial Protection Bureau (CFPB), applies to all communications between a servicer and a borrower. A servicing response that provides inaccurate information, uses language that is likely to mislead the borrower about their rights or the servicer's obligations, or obscures the basis for an adverse servicing decision is a UDAAP risk independent of any specific regulatory provision.

Fair Debt Collection Practices Act (FDCPA). When a servicer that acquired a loan after default contacts a borrower about a delinquent account, the FDCPA (15 U.S.C. 1692) imposes requirements on the content, timing, and manner of debt collection communications. Servicers that acquired a loan before default and are collecting on their own loans are generally not covered by the FDCPA as debt collectors; servicers that acquired a loan after default are covered. All servicers, regardless of FDCPA status, remain subject to the CFPB's UDAAP prohibition in connection with servicing activities. An AI tool that drafts collection-adjacent servicing communications must produce output that complies with these requirements, and the institution must determine whether the FDCPA applies based on when and how the servicer acquired the loan.

The Dispute Drafting Workflow: From QWR to Compliant Response

Dispute response drafting is the most regulated and most consequential category of servicing correspondence. A QWR response that is late, incomplete, or factually inaccurate creates specific RESPA liability, and the penalties for RESPA violations include actual damages, statutory damages of up to $2,000 per individual violation (or up to $1,000,000 in a class action), and attorneys' fees. The stakes justify a careful, structured workflow.

The workflow has five stages, each with a defined human and AI role.

Stage 1: Classify and intake the inquiry. When a written inquiry arrives from a borrower, the first step is to determine whether it constitutes a QWR under RESPA Section 6 and to begin the five-business-day acknowledgment clock. AI tools can assist with classification: they can scan incoming correspondence to identify language that triggers RESPA's QWR definition (a written request that identifies the borrower's account and includes a statement of reasons the borrower believes an error has occurred or a request for account information), and they can flag items that require QWR treatment versus items that are general inquiries not subject to RESPA's specific requirements. Human review confirms the classification before the acknowledgment letter is sent, because misclassifying a QWR as a general inquiry and missing the five-business-day acknowledgment deadline is a RESPA violation.

Stage 2: Pull and review the servicing file. The analyst assigned to the QWR must pull the relevant account history and documentation before drafting begins. This is not a step AI can substitute for, because the accuracy of the response depends entirely on the accuracy of the servicing system data and the analyst's ability to understand what that data shows. In the opening story, the AI drafted a response about "updated property tax data" because it was not provided with the actual servicing file data. The analyst must retrieve the specific transactions, disbursements, correspondence, and account notes that are relevant to the borrower's inquiry and review them before drafting begins.

Stage 3: Draft with AI assistance, grounded in file data. With the relevant file data in hand, the analyst can use an AI drafting tool to produce the initial response. The key design principle here is that AI is drafting from the facts the analyst provides, not from the model's general knowledge of how servicing works. The prompt should include: the borrower's name and account number, the specific question or complaint the borrower raised in the QWR, the specific data from the servicing file that answers that question, the RESPA requirement for a complete and responsive answer, and any required elements that must be present in the response. The AI then assembles these elements into a professional, well-organized response letter. The quality of this step is entirely dependent on the quality of the information the analyst provides.

Stage 4: Verify the draft against the file and the regulation. The analyst reviews the AI draft against a verification checklist: Does the response directly address the borrower's specific question? Does every factual claim in the response match the servicing file data exactly? Does the response include the required RESPA elements (identification of the account, the servicer's contact information, the outcome of the investigation)? Does the response timeline comply with RESPA's 30-business-day requirement? Does the language satisfy plain-language standards that a borrower can understand? Are there any characterizations of the borrower's account or financial situation that the file does not support? This verification step, done carefully, takes 15 to 25 minutes for a typical QWR response.

Stage 5: Document, approve, and send. The analyst documents in the servicing system that the response was AI-assisted, that the factual verification was completed, that the response was reviewed for regulatory compliance, and that a named individual approved the response before it was sent. The response is then sent within the RESPA deadline.

Loss Mitigation Correspondence Without Creating UDAAP Risk

Loss mitigation correspondence presents the highest UDAAP risk of any category of servicing communication, for two reasons. First, the borrowers receiving this correspondence are typically in financial distress and the communications involve decisions of significant consequence: whether they will receive a loan modification, a forbearance, a repayment plan, or a foreclosure referral. Second, the CFPB's mortgage servicing rules are particularly specific about what these communications must contain and when they must be sent, leaving less room for generic or approximate responses.

AI-assisted loss mitigation drafting requires understanding three categories of communication and the specific requirements for each.

Early intervention outreach letters. CFPB regulations require servicers to make certain early intervention contacts with delinquent borrowers, beginning when a loan is 36 days past due. These communications must contain specific information about the servicer's loss mitigation options and must not make misleading representations about the borrower's options or the consequences of continued delinquency. An AI tool that drafts an early intervention letter using language that implies foreclosure is imminent when regulatory timelines have not yet run, or that lists loss mitigation options the borrower does not actually qualify for, creates a UDAAP violation in a communication specifically designed for vulnerable borrowers.

The AI-assisted workflow for early intervention letters starts with a prompt that specifies the exact delinquency status of the account, the specific loss mitigation programs the servicer has available and the borrower's preliminary eligibility for each, and the accurate timeline for regulatory and servicer action if the delinquency continues. The AI then drafts a letter that communicates this specific, accurate information. The verification step confirms that the delinquency status, the listed options, and the timeline all match the servicing system data exactly.

Loss mitigation decision notices. When a servicer makes a determination on a loss mitigation application, whether approving it for a specific option or declining it, the CFPB rules require specific written notification. An approval notice must identify the specific option approved and the terms. A denial notice must state the reason for denial and the borrower's right to appeal. An AI tool that drafts a denial notice with an inaccurate reason, an incomplete statement of appeal rights, or an incorrect description of the modification terms creates a CFPB mortgage servicing rule violation and a potential UDAAP problem. The verification requirement is the same as for QWR responses: every factual claim must be traced to the servicing file, and all required elements must be present.

Foreclosure referral communications. Before a servicer may initiate foreclosure under the CFPB's single-point-of-contact and loss mitigation rules, specific procedural steps must have occurred and specific notifications must have been sent. AI tools should not be used to draft foreclosure referral communications without careful legal review, because these communications have the most significant consequences for the borrower and the most detailed regulatory requirements. The human role in foreclosure referral communications is not verification of an AI draft; it is primary drafting with AI assistance on specific elements, and final approval by qualified legal or compliance staff.

Hands-On: Drafting a QWR Response Correctly

Let us return to the opening story and build the correct response to Anthony Reeves's QWR, using the AI-assisted workflow described above.

The actual facts from the servicing file: Anthony Reeves's account is a 30-year fixed mortgage at 6.25%. His escrow account covers property taxes and hazard insurance. The property tax assessment for his property has not changed this year. His hazard insurance premium increased from $1,140 annually to $1,716 annually at renewal on February 28, 2026, an increase of $576 per year or $48 per month. Under RESPA's escrow analysis rules, the servicer recalculated the escrow payment following the insurance premium increase, resulting in an additional $166 per month to cover the projected shortfall over the next 12 months. Total escrow increase: $214 per month ($48 for the premium increase plus $166 for the shortage catchup). This is entirely documented in the servicing system.

The correct prompt: "Draft a RESPA QWR response letter. Borrower: Anthony Reeves. Account: [number]. His QWR asked why his escrow payment increased by $214 per month. The reason is an increase in his hazard insurance premium from $1,140 annually to $1,716 annually at the February 28, 2026 renewal, plus a $166 per month escrow shortage catchup required under the resulting escrow analysis. His property tax assessment has not changed. The response must: directly answer his question, explain the insurance premium increase and the escrow shortage calculation, state that the property tax assessment has not changed, note that he can contact us to request an escrow analysis statement, include our contact information, and comply with RESPA Section 6. Use plain language. Do not add any other information."

The AI output (abbreviated for illustration): "Dear Mr. Reeves: Thank you for your inquiry dated [date] about the increase in your escrow payment. We have completed a review of your account and can provide the following explanation. Your escrow payment increased by $214 per month because of two factors related to your hazard insurance: your annual hazard insurance premium increased from $1,140 to $1,716 at your policy renewal on February 28, 2026, which increased your monthly escrow contribution by $48; and because the premium increase created a projected shortage in your escrow account for the current analysis period, your payment includes an additional $166 per month to bring the account to the required balance over the next 12 months. Your property tax assessment has not changed and is not a factor in this adjustment. You may request a detailed escrow analysis statement at any time by contacting us at [contact information]. If you believe there is an error in your account, you have the right to request further investigation by contacting us in writing."

Verification step: The analyst confirms: the reason stated (insurance premium increase) matches the servicing file. The premium amounts ($1,140 and $1,716) match the insurance disbursement records. The shortage amount ($166 per month) matches the escrow analysis output. The property tax statement is accurate. The contact information is correct. All required RESPA elements are present. The response directly addresses the borrower's specific question. Verification: complete. Time elapsed: approximately 18 minutes total, compared to an estimated 35 minutes for a fully manual RESPA response with the same level of care.

Governing AI in Servicing Correspondence at Scale

For a servicer managing a portfolio of any significant size, individual adherence to the workflow described in this lesson is necessary but not sufficient. The volume of regulated correspondence in a large servicing operation, potentially thousands of QWR responses, loss mitigation letters, early intervention contacts, and account maintenance notices per month, requires institutional governance infrastructure that ensures consistency across the team and maintains the audit trail that regulators will expect.

The governance infrastructure for servicing correspondence AI has three components that build on the individual workflow.

Communication templates and prompt library. The compliance and legal teams should develop and maintain a library of approved prompt templates for each recurring communication type: QWR responses, early intervention letters, loss mitigation denial notices, forbearance acknowledgment letters, and so forth. These templates incorporate the policy grounding, required elements guidance, and negative instructions appropriate for each type. Analysts use the approved templates rather than constructing prompts from scratch, which reduces the variance in AI output quality and ensures that required regulatory elements are systematically included in the prompt. The prompt library is a living document that is updated when regulations change or when error tracking identifies patterns in template-related failures.

Quality sampling and error tracking. A servicer using AI-assisted correspondence at scale must implement systematic quality review of outgoing communications. Not every communication can be reviewed by a second reviewer given volume constraints, but a random sample of every communication category, reviewed by a quality assurance team against the verification checklist, produces a continuous picture of where the AI-assisted workflow is performing well and where errors are appearing. Error tracking that identifies systematic failures (for example, a pattern of missing escrow analysis detail in QWR responses) enables prompt correction of the prompt template before the error pattern becomes a RESPA examination finding.

Complaint linkage and escalation. When borrower complaints mention the substance of a servicing communication, the complaint management system should route the complaint to the team responsible for QA of AI-assisted correspondence, not just to the general complaint resolution function. A complaint that says "you told me my taxes went up but they didn't" is a direct indicator of an AI output error and should trigger a review of the specific communication and the file, a correction to the borrower, and a review of whether the template or workflow that produced the error is generating similar errors in other files. OCC Bulletin 2026-13's model governance expectations include monitoring AI model performance through feedback loops including complaint data; for servicing correspondence AI, complaint linkage is the primary feedback mechanism.

The institution that builds these three components has a servicing correspondence AI program that is not only compliant today but is continuously self-improving and continuously generating the documentation that regulators will want to see. That is a defensible position under OCC Bulletin 2026-13, under CFPB supervision of mortgage servicers, and under any RESPA or UDAAP inquiry that might arise from a specific communication.

Key Takeaways

• Servicing and dispute correspondence is among the most heavily regulated communication a bank produces. RESPA, ECOA, the CFPB's mortgage servicing rules, UDAAP, and the FDCPA all apply to specific categories of servicing communication, each with distinct timing, content, and accuracy requirements that AI drafting must satisfy.
• The foundational principle of AI-assisted servicing correspondence is that AI drafts from facts the analyst provides, not from the model's general knowledge of how servicing works. An AI tool that is not given the specific file data will generate plausible-sounding but inaccurate responses, as the QWR example in this lesson demonstrates.
• The five-stage QWR drafting workflow (classify and intake, pull the file, draft with AI from provided facts, verify against the file and regulation, document and send) captures AI's speed benefit while maintaining the human accountability and regulatory compliance that RESPA requires. Each stage is essential and none can be safely omitted under volume pressure.
• Loss mitigation correspondence carries the highest UDAAP risk in servicing because it involves financially distressed borrowers and decision-of-consequence communications. AI-assisted drafting in this category requires the most rigorous prompt grounding and verification, and foreclosure referral communications should involve qualified legal or compliance review rather than routine AI-assisted drafting.
• The common failure modes in AI-assisted servicing drafting are fabricated reasons (the model fills in plausible reasons the file does not support), missing required elements (the model produces fluent responses that omit regulatory required content), and inappropriate characterizations (the model adds helpful-sounding context about the borrower's situation that the file does not warrant). A structured verification checklist catches all three.
• Governing AI correspondence at scale requires three institutional components beyond individual workflow adherence: a prompt template library maintained by compliance and legal, systematic quality sampling and error tracking, and complaint linkage that routes borrower feedback about communication errors to the AI quality assurance process. These components convert individual best practice into sustainable institutional compliance.
• OCC Bulletin 2026-13 requires documented governance controls on AI tools in all consumer-facing applications. A servicer using AI-assisted correspondence without documented verification workflows, prompt templates reviewed by compliance, and a quality monitoring program is not positioned to demonstrate the governance that the 2026 interagency model-risk guidance expects.
• Accountability for every AI-assisted servicing communication stays with the analyst who verified and approved it. RESPA liability and UDAAP enforcement run against the servicer, not the AI tool. The analyst who approves a QWR response owns its accuracy, its completeness, and its regulatory compliance.