Predict customer churn with your AI assistant

Train a churn model on your own customer data, scored on rows it never saw.

Export your customers to CSV, ask your assistant, and get a model that names what actually predicts churn — with an honest score and a refusal when the data is too thin to support one.

Run this with your AI assistant

meelu-analytics-mcp is a free, open source MCP server that runs on your machine. One command installs it and configures your assistant.

meelu-analytics-mcp View on GitHub All open source

The problem

Why this is hard to answer today

You know your churn rate. You do not know which accounts make it up, or which of the things you believe about churn are actually true.

The monthly number tells you how many left, never which ones are about to.
The reasons everyone repeats — low usage, no onboarding call, small plan — are folklore until something measures them against your data.
A model built in a hurry is usually graded on the same rows it learned from, which makes a useless model look excellent.

The solution

What the engine does instead

Ask in plain English. The engine picks the method from the shape of your data, trains on part of it and scores on the part it never saw, then tells you what the model is actually paying attention to.

A model trained on your customers, not a generic one shipped in a box.
Scored on held-out rows, so the accuracy figure means something.
Ranked drivers via SHAP — what moves the prediction, not just what correlates.
A trust rating on every result, and a plain refusal when there are too few rows to answer.

Built on Supervised machine learning, one of nine tool categories in the server.

How it works

Step by step

Step 1

Export your customers to CSV

One row per customer. Anything you already record is fair game — the engine works out which columns are usable and which are ID codes wearing a number's clothing.

customers.csv

customer_id,signup_date,last_active,monthly_sessions,support_tickets,plan,seats,churned
CUST_001,2024-01-15,2025-08-20,15,2,pro,4,1
CUST_002,2024-03-10,2026-09-01,45,0,team,12,0
CUST_003,2024-06-01,2025-08-15,8,5,starter,1,1

Step 2
Install the MCP server
Install meelu-analytics-mcp once, then restart your assistant. The install page has the one command and the README covers the rest.

Step 3

Ask the question

Say “using meelu” the first time so your assistant reaches for the engine instead of improvising its own analysis.

In your assistant

Using meelu, load ~/Downloads/customers.csv and tell me
what predicts whether a customer churns. Which model did it
use, how was it scored, and how much should I trust it?

Step 4
Push back on the answer
The follow-ups are where the engine earns its keep, because each one has a recorded answer rather than an improvised one.
Follow-ups
```
Which features matter most, and in which direction?
How much should I trust this?
Is that actually causing churn, or just related to it?
```

Example output

What comes back

The exact wording is your assistant's, but the substance underneath is fixed: the method it chose, the score on held-out data, the ranked drivers, and a trust block.

Model: gradient-boosted trees (chosen for mixed
       numeric and categorical columns)
Scored on: 1,204 held-out rows the model never saw
Trust:  high — 4,812 rows, no class imbalance flagged

What drives the prediction, most to least:
  1. days since last_active
  2. monthly_sessions trend over final 60 days
  3. support_tickets in final 30 days
  4. seats  (fewer seats, higher risk)
  plan tier ranked below all of the above

Caveat: "churned" was derived from last_active, so it
partly encodes its own predictor. Treat driver 1 with
care.

The caveat in that last block is the point. An engine that flags leakage in your target column is doing the job a hurried script does not.
Ask for per-customer scores and you get them, as a column saved back alongside your data so the next question can build on it.

In practice

Who asks this

Seed-stage SaaS founder

Five hundred customers, a churn rate you can quote and no idea which accounts are behind it.

Rank every account by predicted risk, then spend the week on the top twenty rather than emailing all five hundred.

Customer success lead

Leadership believes onboarding calls prevent churn. Nobody has checked.

Put the belief to a real test. Either it survives contact with the data and you fund it, or it does not and you stop spending the hours.

Analyst without a data team

You can write the SQL but not defend a model in a review.

The method and its justification are recorded in the result, so the review is about the decision rather than about your homework.

FAQ

Common questions

What is churn prediction?

Churn prediction is estimating which customers are likely to stop paying or stop using a product, before they do it. It works by learning the patterns that preceded past cancellations — falling usage, longer gaps between logins, failed payments, support complaints — and scoring current customers against them. The output is a ranked list, not a verdict, and it is useful because it tells you where to spend limited retention effort.

How do you calculate churn rate?

Divide the number of customers lost in a period by the number you had at the start of that period. Lose 12 of 400 customers in a month and monthly churn is 3%. Revenue churn uses the same formula with money instead of headcount, and often tells a different story, because losing one large account matters more than losing five small ones.

What is a good churn rate?

It depends on who you sell to. Business software sold to large companies typically runs at 0.5% to 1% of customers a month; small-business and self-serve products commonly sit between 3% and 7%. Consumer subscriptions are higher again. The useful comparison is against your own trend rather than an industry figure, because the number moves with contract length, price and payment method more than with quality.

What causes customer churn?

Most churn comes down to four things: the customer never reached the point where the product was useful, the value faded once their need changed, something went wrong and was handled badly, or the card simply failed. That last one, involuntary churn, is often a tenth of all cancellations and is the cheapest to fix. Price is blamed far more often than it is the actual cause.

How accurate is churn prediction?

Accuracy depends entirely on your data, and anyone quoting a single figure across all businesses is selling something. What matters is the score measured on rows the model has never seen, and whether that beats a simple rule such as "flag anyone inactive for 30 days". When customers are few or the churn signal is weak, the honest result is a refusal rather than a confident number.

Keep going

Related use cases

Customer LTV

Find out what a customer is worth, and which traits predict it

Cohort retention

See which signup months held up, and find the month things changed

Customer segmentation

RFM segments and natural groupings found in your data

Try churn prediction on your own data

Install meelu-analytics-mcp once and every use case on this site is available to your assistant — it is one server with forty-five tools, not six separate products.