AI in Mortgage Servicing: Borrower Queries for Small Servicers

A small mortgage servicer can use AI to sort incoming borrower queries by type and urgency, draft replies for staff to review, summarise calls and explain escrow changes in plain English. It should not make decisions, send unreviewed answers on money or hardship, or handle complaints on its own. Every query that mentions a complaint, a dispute or a regulator goes to a person.

Servicing is a regulated activity with deadlines for acknowledging and answering certain borrower requests, so the point is faster, more consistent preparation, not fewer humans in the loop. The rules differ by market and loan type and I am not giving compliance advice, so confirm what applies to you. For a team of about ten servicing a few thousand loans, a fair illustrative aim is reviewed drafts for the routine 40 per cent or so of contacts, such as payoff and document requests, with every complaint, dispute and hardship message going straight to a person.

Follow me on Instagram@sagnikteaches

A 4,000-loan servicer and its monthly contacts

Picture a servicer with 4,000 loans and a team of ten: six on the phones and inbox, two in escrow and payments, one in default support and one compliance lead. In an average month it receives around 2,600 contacts: roughly 1,500 calls, 800 emails and 300 portal messages. A fair share are routine: payoff requests, payment date and amount, how to set up automatic payment, a duplicate statement, tax or insurance documents, a change of address.

Connect on LinkedInSagnik Bhattacharya

These numbers are illustrative. What matters is that the routine contacts take the same handful of steps each time, and a large part of each is reading the message, working out what it is, finding the record and typing the reply. That is where AI can help, as long as the triage is conservative and the reply is checked.

Subscribe on YouTube@sagnikteaches

Step one: triage by type, urgency and risk

The first job for an assistant is classification, not answering. Staff paste or forward a query (with the borrower's name and loan number removed, or masked by your system) and the assistant returns a category, an urgency rating and a flag for anything that must go to a person. The value is consistency: the same message is labelled the same way at nine in the morning and at five on a Friday.

Worked example: a morning inbox

Say the inbox holds 60 unread messages at 8:30. A staff member pastes them in batches of ten with identifiers stripped, using a fixed instruction:

Classify each borrower message into exactly one category: payoff request; payment query; escrow query; document request; change of details; hardship or difficulty paying; complaint or dispute; other.
Also give: urgency (same day / within 2 days / routine) and ESCALATE = yes if the message mentions dissatisfaction, a dispute about an amount, a lawyer, a regulator, bankruptcy, a death, a foreclosure notice or difficulty paying.
Return a table: message number, category, urgency, ESCALATE, and the exact words that triggered ESCALATE.
Do not answer the messages. If unsure, choose "other" and ESCALATE = yes.

An illustrative output for ten messages might show six routine items, two escrow queries, one hardship message flagged ESCALATE with the trigger phrase "I lost my job in January", and one message flagged ESCALATE for the phrase "this fee is wrong and I will complain". The batch takes two minutes to run and about five to check, compared with fifteen minutes of sorting by hand. Crucially, the "if unsure, escalate" rule means that errors fall on the safe side.

How the triage goes wrong

The common failure is a polite complaint. A borrower writes "just checking why my payment went up again, third time now". It reads calmly and may be classified as a payment query. Your triage prompt should include the cues for repeat contact and add them to your own list over time. Review a random 20 messages each week for the first month, compare the label to a person's judgement, and adjust the prompt. Track how many ESCALATE flags a person later downgraded and how many routine labels they upgraded; upgrades are the ones that matter.

Step two: draft the reply, in your words, from the record

For routine categories, the assistant can draft a reply from facts the staff member supplies: the payoff figure from the system, the date, the document requested. It should write only what it is given. A prompt template:

Draft a reply to a borrower from [servicer name] servicing team.
Facts you may use (nothing else): payoff quote $212,480.15 valid until 18 March; per-diem interest $31.42; to receive written payoff we need a signed request, which can be uploaded in the portal.
Tone: plain, polite, no jargon. Under 140 words. Do not give legal or tax advice. Do not promise anything beyond the facts. End with the team's phone line placeholder [phone].

The sample output states the payoff figure and expiry date, explains per-diem interest in one sentence, tells the borrower how to submit the signed request and gives the phone placeholder. A staff member checks every figure against the system, confirms the quote is still valid and sends it. If the figures are not in the prompt, the draft should say what is missing instead of guessing. Test this by deliberately leaving out a figure and checking that the assistant asks for it.

The related tutorial on keeping mortgage clients updated during an application covers the origination side of the same drafting technique, which is useful if you also run an advice or broking arm.

Step three: explain escrow changes without implying a decision

Escrow statements and payment changes generate more queries than almost anything else. The statement is technical, and borrowers see only that the payment has gone up. An assistant is good at turning the statement's lines into a plain explanation, as long as the numbers come from your system and the explanation does not make promises.

Worked example: a payment that rose by $96

Say a borrower's monthly payment rises from $1,642 to $1,738. The annual review shows property tax up $780, insurance up $372, and a prior shortage of $0, so the escrow portion rises by $1,152 a year, which is $96 a month. The staff member provides those figures and asks for a short explanation. The draft (illustrative) says the principal and interest portion is unchanged, the escrow portion rose because tax and insurance bills rose by the amounts shown, and the new payment starts with the stated month's payment. It points to the statement for the line items and invites the borrower to call with questions.

The check: recompute $780 plus $372 equals $1,152, and $1,152 divided by 12 equals $96, then match against the system's figures. Never let the assistant explain why a tax bill changed or advise the borrower to dispute it. If the borrower asks whether they should shop for cheaper insurance, that is a conversation for a person, and certainly not an automated answer.

Step four: call summaries that go into the record correctly

Servicing calls need notes: what the borrower asked, what was said, what was agreed, what happens next. Writing them takes three to five minutes per call, and the quality varies. A call-summary tool or a transcript pasted into an assistant can produce a structured note. The related tutorial on whether it is safe to let AI listen to mortgage advice calls sets out the consent and safety questions, which apply equally here.

A template for the note: reason for call, what the borrower said (their words for anything about hardship or dispute), what the agent told them, actions with owners and dates, and any follow-up promised. A sample summary (illustrative) might read: "Reason: borrower asked why the payment rose. Agent explained the escrow change using the annual statement. Borrower said the new amount is 'hard but manageable'. Action: agent to send the written breakdown by Friday. No complaint raised." The handler then listens for or reads the key moments, because a summary can flatten "hard but manageable" into "no issue". Language about difficulty paying is exactly what an agent must not miss, and where it appears the call must be handled under your hardship procedure.

Handling a hardship message: gathering facts without judging

The message that most needs care is the one from a borrower who cannot pay. Suppose a borrower writes that their hours were cut, they missed last month's payment and they do not know what to do. Triage flags it ESCALATE and a trained agent takes it. The assistant's role is limited to preparation. It can list the facts already on the account (payments made, the date of the missed payment, any previous contact), draft a short and kind acknowledgement for the agent to adapt, and produce a checklist of the documents your procedure asks for. It cannot assess eligibility, suggest an outcome or tell the borrower what will happen.

An illustrative acknowledgement draft: "Thank you for getting in touch, and I am sorry to hear things are difficult. A member of our team will call you on [date] to talk through the options we can offer. In the meantime, you do not need to send anything, but it may help to have your recent income details to hand." The agent checks that it matches your procedure's wording, because what you say about options and timing may be regulated, and removes any phrase that implies an outcome. Keep the borrower's own words in the file note exactly as they said them.

A practical safeguard: never let the assistant see the full account history for these cases unless your data terms and procedures allow it. A short, masked summary prepared by the agent is enough for a draft.

A library of approved answers for the questions you get every day

Once triage and drafting work, the best investment is a small library of approved answers that staff check against, instead of asking the assistant afresh each time. Take your ten most common queries and write, with your compliance lead, a plain-language answer for each, with the places where figures go marked clearly. For example: "How do I set up automatic payment", "Where do I find my annual escrow statement", "How do I request a payoff quote", "How do I change my mailing address", "When is my payment due and what happens if it is late".

The assistant then becomes a tool for personalising and shortening a vetted answer, not for inventing one. The instruction is: "Use only the approved answer below, fill in the facts I supply, and do not add anything." That keeps the wording consistent with your policies, makes review quicker, and lets you update one document when a rule or process changes rather than hunting for old phrasing in staff heads. Version the library with a date and the name of whoever approved it, and retire answers that are out of date.

Worked example: late-payment wording

The approved answer for "what happens if my payment is late" says what your loan documents say about the grace period, when a late charge applies and how to reach the team if the borrower is struggling. Your agent supplies the borrower's due date, the grace period end and the late charge from the system: say the due date is the 1st, the grace period ends on the 15th, and the charge is the amount in the note. The draft puts those three facts into the approved wording. The agent verifies each against the record, confirms the loan type matches the template, and sends it. Time saved: about four minutes per query, and the greater gain is that two agents give the same answer.

A realistic bad day, and what it teaches

Suppose a busy Friday, triage runs on 90 messages, and the reviewer, short of time, approves the drafts without opening the system. One draft quotes a payoff figure that was valid until the previous day, copied from an older note. The borrower relies on it and sends a payment that is short. Nothing in the assistant was wrong; the figure it was given was stale. The fix is a rule, not a better prompt: any figure in a draft must come from the system on the day it is sent, and the reviewer ticks a field in the log for each figure checked.

Another likely slip is tone. An assistant asked to be "firm" about a late payment produces wording that reads as a threat, which in a servicing context may breach your rules on communications. Keep tone instructions neutral and polite, ban words such as "final", "must" and "legal action" in your prompt, and let compliance review the library once before it is used.

Quarterly audit: sampling what the tool touched

Treat the AI steps as part of your quality assurance. Each quarter, pull a random sample of 40 queries that went through triage and drafting, and have someone who did not handle them score three things: was the category right, was every figure in the reply correct on the day, and was anything said that your procedures do not allow. Record the results, and fix prompts and the library where errors cluster. If the sample reveals a missed escalation, record it as an incident, find the phrase that triggered the miss, and add it to the escalation list. This is also useful evidence for an auditor or investor who asks how you control the use of AI tools in borrower contact.

Guardrails that sit around every step

• Human escalation. Complaints, disputes, hardship, bereavement, bankruptcy, legal threats and anything unusual go to a named person the same day. The triage prompt flags; the person decides what it is.
• No automated decisions. The assistant never decides on fees, waivers, payment plans, loss-mitigation options or escrow treatment. It can list the documents received and the documents still needed.
• Source of truth. Figures come from your servicing system and are pasted in. Anything the assistant did not receive it does not state.
• Data handling. Mask names, loan numbers and account details where the task does not need them. Use a business plan whose terms you have read, because personal free accounts may treat your data differently, and agree with your investors or clients what their servicing agreements say about third-party tools.
• Records. Keep the sent reply and the staff member who approved it. If your retention rules cover drafts, include them.

For the complaint-handling side, can AI handle customer complaints without making them worse explains why complaint replies need a person, and using AI in compliance investigations covers a safe workflow when a complaint becomes a formal matter.

Three numbers that show whether triage is safe

Pick three measures before you start, and take them for a month with and without the assistant on a sample of queries: minutes from receipt to reviewed reply, number of replies that needed a material correction, and number of escalations missed by triage. The third is the one that matters most, and its target is zero. If a single complaint or hardship message is mislabelled as routine and reaches a customer unflagged, stop using the triage step until the prompt is fixed.

A modest, realistic expectation for the servicer above: the routine contacts that are suited to drafting (payoff, document requests, escrow explainers, change of details) might make up 40 per cent of volume, and a draft might save three to five minutes of a ten-minute handling time. That is a saving of a few staff hours a day. It is useful but not dramatic, and it is worth having only if the quality of the reply holds. Evidence about wider results in servicing comes mostly from vendors, and I would not quote it without your own numbers.

An illustrative week-by-week start

1. Week 1. Write the escalation list with your compliance lead: every phrase and situation that must reach a person. Time a sample of 50 routine queries.
2. Week 2. Run triage in parallel with manual sorting on the morning inbox. Compare labels daily and tune the prompt.
3. Week 3. Add drafting for two categories only, document requests and payoff requests. Review every draft; log corrections.
4. Week 4. Add the escrow explainer and call summaries for volunteers, then review the three measures with the team and decide what stays.

If you would rather map this with someone, that is what my AI implementation consultation covers: choosing the steps worth automating in your own workflow and handing over something your team can run, with the escalation and decision points left where your policies put them. Before buying software, choosing a mortgage CRM with AI built in lists what to check.

Servicer questions about AI and borrower contact

Can AI answer borrowers directly?

For a small servicer the safer pattern is AI drafting and staff sending. Direct automated replies risk a wrong answer on a payment, escrow or hardship question, and some servicing rules limit what can be automated. If you ever allow direct replies, restrict them to fixed facts such as office hours and document requests.

What counts as a complaint that must be escalated?

Your own complaints policy and the rules that apply to you decide that, so ask your compliance adviser. In practice, treat any expression of dissatisfaction, any dispute about an amount or any mention of a regulator or lawyer as an escalation, and have the triage step flag it rather than judge it.

Is it acceptable to record and summarise borrower calls with AI?

Only with the notices and consents your rules require, and with a tool whose data terms you have checked. Keep the recording and summary under the same retention schedule as other servicing records, and have the call handler review the summary before it enters the file.

Can AI decide on payment plans or hardship requests?

No. Decisions about borrower relief, fees and payment arrangements stay with trained staff following your written procedures. AI may gather the facts, list the documents received and draft a request for what is missing, but the assessment and the outcome are human.

Further reads

• What Mortgage Brokers Can Automate With AI, and What Stays Advice — The broking side of the same advice boundary.
• Best AI Tools for Mortgage Brokers at Each Stage of a Case — Tools by stage of a mortgage case.
• AI Note Takers for Financial Advisers: Compliant Meeting Records — Compliant meeting records and note takers.
• Using AI to Spot Remortgage Opportunities in Your Client Bank — Finding remortgage opportunities in a client base.
• From Enquiry to Fact Find: An AI Workflow for Mortgage Brokers — From first enquiry to fact find.
• AI Tools for Mortgage Lenders: Processing, Disclosures, Audit Trails — A plain-English map of AI for small lenders: document readers, condition trackers and disclosure checkers, with the audit trail each one needs.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: general servicing practice and vendor data terms for business chat assistants. No regulation is quoted, so confirm the servicing and complaint rules that apply to you with a compliance adviser.