How Vantage Point connected Claude to a mortgage team's Salesforce so every daily follow-up arrives ranked, routed to the right channel and drafted for approval.
Key takeaways
| What is it? | A Claude agent that triages a mortgage team's daily Salesforce tasks and prepares every reply for approval. |
| Key benefit | Email, text and call follow-ups worked from one chat, each on the channel the borrower last used. |
| Industry | Mortgage lending |
| Platform | Claude, Salesforce (Candid mortgage CRM), Front, Twilio |
| Best for | Loan officers and small lending teams buried in daily follow-up tasks across email, text and phone |
| Bottom line | A human approves every message, and every completed task books the next one automatically, so no borrower slips through. |
Quick answer
This case study shows how Vantage Point built a two-phase Claude agent for Jasmine Mortgage Team, an Atlanta residential mortgage team running on Candid, a mortgage CRM built on Salesforce. Each morning the agent ranks the day's follow-up tasks, picks the right channel from the borrower's real communication history, and drafts the email, text or call brief. Nothing is sent until a team member approves it. It is relevant to loan officers, mortgage operations leads and any regulated team that wants AI help with follow-up without giving up control.
Every mortgage team knows the feeling. The pipeline is healthy, but the task list is longer than the day. This is the story of how one team turned that list into a short, focused working session.
The challenge: too many follow-ups, too many places to look
Jasmine Mortgage Team is an Atlanta-based residential mortgage team. A small core team handles 100+ new leads a month on top of a multi-year database of past borrowers and referral partners. Their CRM is Candid, a mortgage-specific platform built on Salesforce, alongside Front for email, Twilio-powered texting inside Candid, and Aircall for phones.
Borrower follow-up was the team's most time-consuming Salesforce job. Every day brought 20 to 50 open tasks. Each one meant opening the record, reading its history, deciding how to reach the client and writing the message by hand.
Three things made it harder than it should have been. First, the conversation lived in three places: texts in one Candid object, emails in another, and logged calls in completed tasks. Second, the right channel was a guess, because nobody could see at a glance whether a borrower last replied by text or email, or had gone quiet on both. Third, the next follow-up date depended on memory. A new lead needs a touch within a business day; a borrower who did not qualify may not need one for over a year.
Texting was the sharpest pain point. The shared text inbox had drafts that disappeared and no clear ownership, so working texts in batches meant copying messages between systems. The team already used Claude to triage around 300 emails a day in Front. Their goal was to do the same for Salesforce: sit down for one focused session and clear the list.
The solution: a two-phase daily task agent in Claude
Vantage Point connected Claude to the team's Salesforce org through the Salesforce MCP connector and built a shared Claude project the whole team can use. Because Candid stores prospects in custom objects rather than standard Salesforce leads, the agent was mapped to read those objects directly. Access follows each person's own Salesforce permissions.
Phase 1: Triage
Each day the agent pulls every open Salesforce task that is due today or overdue and sorts it into three batches:
- Due today, ranked High to Low from the record and its comments
- Overdue, ranked the same way
- Red flags: duplicate tasks, several open tasks on one record on the same day, unclear channel or missing email data, test records, stale overdue tasks and unanswered texts
Every task shows the linked prospect or deal and its current status, the owner's name, and a recommended reply channel. To choose the channel, the agent merges the borrower's texts, emails and logged calls by date, then applies simple rules:
- Reply on the channel the client last used
- Recommend a phone call when several texts and emails have gone unanswered
- Flag a client's unanswered last message as urgent
- Default to email, with a note, when the client has never replied
Phase 2: Execution
The team member picks which tasks or batches to work. The agent prepares each reply on the channel from Phase 1, unless told otherwise:
- Email: a draft for review, sent through Front only after explicit approval
- Text: a draft for review; once approved, the agent writes it to Candid, which sends it through Twilio, and then re-checks the record to confirm delivery
- Phone: a short call brief of about 50 words, with history and talking points, to read before dialing
Every reply carries a prominent warning if the record is marked do-not-email, do-not-text or do-not-call. Nothing goes out without a human yes.
Closing the loop
When a task is complete (an email sent, a text sent, or a call the team member confirms), the agent creates the next follow-up task. The due date follows the record's loan status, from 1 business day for new leads and files with docs out, to 16 months for borrowers who did not qualify. Closed, clear-to-close and dead records get no new task. Each session ends with a tracker showing what was completed and what is still outstanding.
The results: one chat for email, text and phone
The agent went live in October 2026 with all three channels working from a single Claude chat. In the team's final review, emails sent through Front, texts were delivered through Candid and Twilio, and call briefs were ready before each dial, with tasks completing and rolling forward by loan status.
| Area | Before | After |
|---|---|---|
| Daily task review | Each of 20 to 50 tasks opened and read by hand | Batched, ranked High to Low, red flags surfaced automatically |
| Communication history | Texts, emails and calls in three separate places | Merged by date for every task |
| Channel choice | Best guess | Recommended from the borrower's last reply |
| Texting | Shared inbox with disappearing drafts | Drafted in Claude, sent through Candid, delivery re-checked |
| Call prep | None | A 50-word brief with history and talking points |
| Next follow-up | Set manually from memory | Created automatically, timed by loan status |
| Do-not-contact | Easy to miss | Prominent warning on every reply |
One moment in the live review stood out. Jeff Goddard, who leads business development for the team, had assumed sending email through Front from Claude was not possible:
"That's really amazing that it's sending it, because I thought it was a known limitation of the Front MCP."
Jeff Goddard, Business Development, Jasmine Mortgage Team
Thinking about AI for your follow-up? See how our Claude services work, or book a free 30-minute CRM assessment.
Key technologies and integrations
- Claude (shared team project)
- Salesforce MCP connector (read and write, permission-scoped)
- Candid mortgage CRM on Salesforce (custom prospect, deal and text message objects)
- Front (email)
- Twilio (SMS, through Candid)
- Salesforce Tasks and Content Notes
Why it matters: lessons for lending teams
The biggest win was context. Most follow-up time goes into working out what happened last and how to reply. Once texts, emails and calls sit on one timeline, the right next step is usually obvious, and an agent can propose it in seconds.
The second lesson is control. In a regulated business, an AI that sends on its own is a liability. This build keeps a person in charge of every message, respects each user's Salesforce permissions, and puts do-not-contact flags in front of the reader. That is what made the team comfortable using it every day.
The third is that a vendor-built CRM is not a blocker. Candid's custom objects meant off-the-shelf AI tools miscounted records and missed tasks. Mapping the agent to the real data model, rather than standard Salesforce objects, is what made the triage accurate.
If your team is weighing how Claude could work with your mortgage and lending CRM, Vantage Point can help you pick the first workflow and build it safely.
Frequently asked questions
Can Claude work with Salesforce data?
Yes. Through the Salesforce MCP connector, Claude can read and, where allowed, write Salesforce records using plain-English requests. In this project it reads tasks, prospects, deals, emails and texts, and creates follow-up tasks and outbound text records.
Does Claude send messages on its own?
No. Every email and text is shown as a draft first and only goes out after a team member approves it. Call briefs are for the person dialing; the agent does not place calls.
How does the agent choose between email, text and phone?
It merges the borrower's texts, emails and logged calls by date and replies on the channel the client last used. If several messages go unanswered, it recommends a call. If the client has never replied, it defaults to email with a note.
Will it work with a mortgage CRM built on Salesforce, like Candid?
Yes. Candid stores prospects and texts in custom objects, so the agent was mapped to those objects directly. Any Salesforce-based CRM can be connected the same way.
How are permissions and compliance handled?
Each user connects with their own Salesforce access, so the agent sees only what that person can see. Records marked do-not-email, do-not-text or do-not-call carry a prominent warning on every draft.
How is the next follow-up date set?
By the record's loan status. New leads and files with documents out get a task for the next business day; borrowers who did not qualify get one in 16 months. Closed and dead records get none.
Is your follow-up list longer than your day?
If your team spends more time finding context than talking to borrowers, start where Jasmine Mortgage Team did: one workflow, built on the CRM you already have.
- Claude services: connect Claude to Salesforce or HubSpot with the right guardrails
- Mortgage and lending: how we work with lenders
- Book a free 30-minute CRM assessment with a senior consultant
Published with the permission of Jasmine Mortgage Team.
