Full Transcript: How to Use AI to Spot Churn Before the Non-Renewal Notice

Why Churn Signals Go Unnoticed

AI can rank your most at-risk service agreements by combining visit history, billing patterns, service notes, and external signals into a single churn risk score. The output identifies which agreements are in trouble, the specific evidence behind each, and which ones are still saveable, giving leaders time to act before the non-renewal notice arrives.

If your leadership or PE sponsor asks you tomorrow which service agreements are about to not renew and what you're going to do about it, could you answer with confidence? Could you back it with data? Many operators can't, because the signals are spread across four or five systems and it takes a lot of resources to pull them together quickly.

By the time a facility manager sends a renewal notice, the save call is really purely theater. The relationship was over months ago. But here's the thing. Many of those customers were telegraphing it for months. Missed preventive maintenance visits, disputed invoices, a callback that dragged on for three visits, a competitive bid your team never saw because the note went into the service platform without a flag. The signals were there. They just weren't assembled.

In this episode, I'm going to show you how AI can rank your most at-risk service agreements, tell you exactly why each one is in trouble, and give you a recommendation for saving each account. All with just four files you already have: your agreement registry, visit history, accounts receivable and billing, and service notes. Twenty-five accounts is the recommended run size. I'm going to use the twenty-five highest annual contract value accounts renewing in the next 90 to 180 days.

Setting Up and Preparing Your Data

The data preparation for churn risk analysis requires four files from systems the business already runs: a service agreement registry, service visit history, accounts receivable and billing data, and service notes. You will also run a separate external research step to check for ownership changes, RFPs, and other outside signals at each account.

As always, we're going to start with the data prep checklist, which has a note about which AI model to use. The simple way to explain it: use the most recent and the most advanced one. This is going to be the thinking modes or the pro versions.

We have two chat windows open. One is going to be used for external research and the other for the main analysis and the follow-up analysis. The first thing we need to do is pick your twenty-five accounts, because AI won't do well with more than that. It's just too much data and too many opportunities for it to get confused. Then go ahead and get your files: your service agreement registry, service visit history, accounts receivable and billing history, and service notes. Just like every time we do this, anytime we use a spreadsheet, we want it in CSV format.

Running External Research

The external research step checks for recent ownership changes, facility management turnover, posted RFPs, expansion or closure signals, and competitive bids across all 25 accounts. This gives AI outside context that internal data alone cannot provide, and "no signal found" is a perfectly acceptable result.

We're going to do some external research to get the latest news or information that might exist for our 25 accounts. As always, you could read the entire prompt if you want, but basically it's telling AI to go out, find research, make sure it's credible, and provide the actual URL so we know it's not hallucinating.

We copy the prompt, paste it into ChatGPT, and at the bottom it says "here are the accounts to research." We paste our list of accounts. Even though my data is synthetic, the customers are actually real companies. Then we send the prompt and ChatGPT goes out onto the web researching news and updates about these 25 customers.

The results come back as a table with the signal type, details with a URL source, and the signal date. Some accounts have no news related to the business, and it says "no signal found." That's actually a good thing. We don't want AI to make things up. We'd rather it say no signal found than fabricate something.

Pre-Analysis: Computing Health Metrics

Before we do the main analysis, we need AI to compute per-agreement health metrics: PM compliance percentage, response time trends, callback rates, AR aging, and disputed invoices. We paste the pre-analysis prompt, upload all four files plus the external research table, and let AI do the math.

This is kind of a micro analysis, a pre-calculation prompt. Basically what it's going to do is compute the per-agreement health metrics. This is just some pre-homework for the AI to do before we have it do the full analysis.

We paste the prompt, upload all four files, then copy and paste the external research output from the other tab. AI does the health calculation and puts it out into a table. We don't actually need to review this table in detail. It was really just doing math to set up the analysis. It does have some data quality notes, but in testing it always comes out really strong.

The Master Analysis: Ranking Churn Risk

The master analysis prompt uses the health metrics, source files, and external signals already in context to produce a ranked churn risk table. Each agreement gets a risk score, a primary risk pattern drawn from a fixed taxonomy, supporting evidence citing specific work orders and invoices, and a recommended save action.

Now we paste the master analysis prompt into the same chat. The AI already has the four files, external signal table, and health signal table in context, so no re-uploading. It confirms it has everything and starts its analysis.

The top-ranked account is UC Health. If you manage this account, this should not come as a surprise. It's a high-value contract and the account notes say they're getting quotes from other contractors, so that's already a known risk. But what makes this interesting is that AI didn't just rank it because of the competitor shopping risk and the high value. It saw other signals pointing to urgency.

In addition to the competitor shopping, there's a new facility manager in role for six months who wants to review the full scope before renewal. There's a warranty callback on a temperature issue that returned after a prior repair. And there's a disputed billing invoice. All those things stacking on top of each other is why it got a score of 88 and a high risk tier, even with a long time to renewal. The risk has been compounded.

AI feels it could be savable. The account manager should lead a branch-manager-supported scope pricing call within two weeks and address the vendor review quotes before the renewal discussion.

Weber State is ranked number three. It's a $95,000 account, less than half the value of Common Spirit Health, but it's labeled high risk with a score of 76 versus Common Spirit's 70. Why? There's a warranty callback. Response time degraded from 21 hours to 36 hours in the last 90 days. The university is running on an old controls platform that's getting flaky. And the account is asking for extended payment terms, which is never a good sign. It's a lot of signals stacking on top of each other.

One more interesting case: the Boyer Company at number six. The contract value is actually quite low, so no one is really losing sleep over this one. But the risk is really high and it's savable. Even though it wouldn't come up as a priority, it's still a customer you want to save. And it might actually be a canary in the coal mine for other things.

Building a Save Plan

We know UC Health is the highest risk and a big account. We want to put a lot of effort into saving it, and AI probably has some ideas. So we ask: "For UC Health, create an action plan to save the account."

It gives us a summary of everything we've learned, explains the evidence, and then lays out a thirty-day action plan. Within 48 hours, the account owner assembles an account brief and goes into the conversation with facts, not being defensive. Within five days, they take specific actions: acknowledge the vendor review process and position it as a scope review, not a sales pitch. It takes us all the way through the next thirty days.

It even creates a recommended customer message, which I would take as a draft and have the team member review and edit so it's in their voice.

The Pattern View

The pattern view prompt groups results by risk pattern, branch, account manager, facility type, and renewal window. It surfaces which risks are systemic versus isolated, and identifies the top five easiest saves where one targeted action could flip the trajectory.

We want to find out what the patterns are. Maybe there's something going on that AI can identify and that we can then take action on from an organization level, not just the account level.

By risk pattern, competitor shopping risk accounts for the most value at risk. But service quality risk is nearly equal in ACV exposure and appears across more agreements, four versus three. Then there's quiet disengagement with three agreements, and others.

By branch, we can see the dominant risk pattern per location. For Salt Lake City, it's service quality risk. For Colorado Springs, it's a combination of competitor shopping risk and quiet disengagement. Maybe there are actions the teams in these branches can take.

By account manager, Chris Lundgren has the biggest concentration, but that's not surprising given his book size. The interesting insight is that service quality risk keeps appearing, and we can dig into why. Sarah has a relatively high amount at risk with quiet disengagement popping up.

By facility type, healthcare has the biggest exposure, with the most accounts at risk and the highest dollar amount. Maybe that's something we can proactively address for other accounts in that space.

The good news from the renewal window view: the biggest amount at risk is between 91 and 180 days out, so there's still time to act.

Going Deeper: Coaching Insights

We saw that a pattern coming up with Sarah is quiet disengagement. So we asked: "Can you tell me more about Sarah's quiet disengagement flag and how we can address it?"

It tells us her book is mostly Colorado Springs. It's not about missing data. It's about customers creating friction around scheduled service access, postponing PMs, or showing signs that they may be less engaged before renewal. It gets into specifics of where the flags are showing up with evidence. If we want to have a conversation with her, we can point to specifics to help her figure out where she can improve.

It's less about customers saying they're leaving and more about customers not prioritizing the relationship. Then it gives actions on how to address it and recommended coaching for her.

There are a lot of ways you can go deeper. We even have a section in the starter kit about going deeper, with questions you can copy and paste or just to get you thinking about what you can ask.

I like to end the videos with examples of going deeper because I think that's really what becoming an AI-first worker and an AI-first organization is about. It's not stopping with the first answer that AI gives you. It's about going deeper, following up, and using your curiosity as an edge to enhance your own skills, experience, and perspective. Use AI to augment yourself and your team, doing the things that AI can't do on its own.