Why a Loan Officer's AI Call List Kept Repeating, and the One Setting That Fixed It

A loan officer I work with built his own AI call list before we ever met. It ran every morning at 8, pulled realtors out of his CRM, and told him who to call and why. Smart setup. It had one problem: it kept handing him the same names.

This is the story of why that happened, what we did about it in two working sessions, and what it says about where AI actually breaks down for a business owner. Short version: the AI was fine. It just had no memory, because nobody had given it one.

What was he trying to do?

He's a mortgage loan officer in the Triangle. A big share of his business comes from realtor referrals, so staying in front of realtors matters more than almost anything else on his calendar. His plan was simple: call 20 or so realtors every Monday, split between the ones who just sent him a deal (say thank you) and the ones who have gone quiet (re-engage).

He set up two scheduled tasks in Claude himself, a few weeks before our first session. Each morning, one task pulled a list of past clients to call and the other pulled realtors, with suggested talking points for each. He was ahead of most owners I meet. He had already figured out what AI should do for him and built a first version on his own.

Then he opened it on day three and saw the same three realtors from day one. And day two.

In his words: "It's giving me the same realtors to call almost every day."

Why did the list keep repeating?

Here's the part most people miss. The AI wasn't broken. It was doing exactly what it was asked: look at the CRM, find realtors who haven't been contacted in a while, list them.

The problem was that nothing told the CRM he had made the call. His connection to the CRM was read-only, so Claude could look but not write. After he called a realtor, the "last contact" date never changed. The next morning, that realtor still looked untouched, so back on the list they went.

He also admitted what most of us would admit about logging calls: "I don't do a really good job at that." His team's convention was to add a line to each contact's notes after every touch, with initials in front. Good practice. Easy to skip when you have 20 calls to make and a closing at 2.

So the real bottleneck wasn't the call list. It was the logging. The list was just where the missing logs showed up.

What did we build in session one?

Two things, in this order.

First, we gave the AI permission to write to his CRM. His CRM is built on Salesforce, which has a setting for exactly this. We turned on create and update, and left delete off. There is no scenario where an AI should be deleting records in your CRM, so we didn't give it the option.

This took a couple of tries. The first connection we set up wouldn't accept the write permission. The second one needed a credential we had to go find in a different settings page. That's normal. Connecting AI to real business software is mostly a series of small permission puzzles, and the reason I sit on the screen with owners instead of sending instructions.

Second, we turned his morning prompt into a skill. A skill is just a written set of instructions the AI follows the same way every time. His has two jobs:

• Build the list. Pick 10 realtors (he can change the number), sorted into thank-you calls and re-engage calls, with talking points.
• Log the call. When he says "I talked to so-and-so, here's what we discussed," the AI writes the note into the contact's record with his initials in front, the way his team already does it, and updates the last-contact date. That realtor then drops off the list for 21 days.

We tested it on one real contact. The note showed up in the CRM, initials and all. He laughed at the initials. Then the morning task was swapped to a one-line prompt that just runs the skill.

What went wrong between sessions?

Two weeks later, the list was still repeating.

Not because the skill failed. Because logging through it hadn't become a habit yet, and a couple of edits to the skill in the meantime had pointed the notes at the wrong field. When we logged three calls together, the AI said "notes added," and the notes weren't there.

We found it in about 10 minutes. Logged one contact at a time, confirmed the right field, watched for the approval prompt the AI shows before it writes anything. The notes landed. "This is exactly what we wanted to do. Look at that."

I'm including this part on purpose. The first build is rarely the last build. The second session is where it starts to stick, because now you're fixing something you've actually used.

What happened next?

With the call list working, we had time left in the session, so I showed him how I prep for my own meetings. I have a skill that pulls the calendar invite, the email thread, and anything I know about the person into a short brief. He watched it run and asked, "What did you just do to get all of that that fast?"

So we built him one. His company runs on Microsoft, so we built it in Copilot instead of Claude, using the tools he already has. It runs at 5 a.m., reads his calendar and email, and writes a brief for each meeting that day. We tested it on a real coffee meeting he had that week. It caught a detail from the email thread that he had forgotten. "That's pretty wild, isn't it?"

That brief now goes out to two teammates too. It still needs him to click send, which is a company setting, and that's the next thing on the list.

What's the lesson for your business?

If you've tried AI and it felt dumb, check whether it can remember anything.

Most AI tools, out of the box, start from zero every time. They can read your data and give you a smart answer, but unless you connect them to something that keeps state, like your CRM, they have no idea what happened yesterday. That's not a flaw in the AI. It's a setup step that almost nobody does.

The loan officer's first version was 80% right. The missing 20% was permission to write back. Once the AI could close the loop, the whole thing worked. And the fix wasn't a new tool. It was a setting in the tool he already paid for.

Questions owners ask

Can AI work with the CRM I already have?

Usually, yes. Most modern CRMs, including anything built on Salesforce, HubSpot, or GoHighLevel, can connect to AI tools for reading and writing. The write part is the one people skip, and it's the one that makes the AI useful day after day.

Why does my AI keep giving me the same answer every day?

Because it has no record of what you did with yesterday's answer. Connect it to the system where that record should live, and give it permission to update it.

Is it safe to let AI write to my CRM?

Yes, with limits. Turn on create and update, keep delete off, and use a tool that asks for approval before it writes. You should see every change before it happens.

Do I need to know how to code to set this up?

No. The loan officer in this story isn't a programmer. The setup is permissions and plain-language instructions, and we do it together on his screen.

Want to see where this fits in your week?

I'm Justin Farmer, and I run Ignitara out of Raleigh, NC. I help owners put AI to work inside the tools they already use, and we build it together so they own it.

If you have a task that keeps repeating, or an AI setup that never quite stuck, book a free 15-minute AI assessment at ignitara.com. We'll find the one thing worth fixing first.