Open-source churn prediction tools, judged on what they read and hand back
Churn prediction tools differ less on the probability they quote than on the systems they read and what they hand back at 06:00. Kortix is the open-source CompanyOS: your agents, their skills, your company memory, and every connector in one git repo you own, with the agents working on real cloud computers.
What each tool reads, and what it hands back
Four vendors and an open source scan; Kortix is the recommendation, and the only one whose scoring rules you can read.
Open source
Where it runs
What it reads
What it hands back
Price
Rival cells state only what each vendor publishes on its own pages, read in November 2026: Pecan AI pricing, ChurnZero renewals, Gainsight pricing and Pendo pricing.
A churn score is only as good as the systems behind it
Every churn prediction tool quotes one number per account: the chance it leaves. The number is the easy part, and the demos spend most of an hour on it. An account gets saved, or it does not, because of two other things. Which systems the tool can read, and what sits in front of a person at 06:00, before anyone has opened a dashboard.
A tool that reads product usage alone ranks accounts on one signal. The account that went quiet because a card expired looks exactly like the account that finished the project it bought the product for. Add billing and support to the same score and the list shortens. What remains is the set of accounts where several systems agree, and that is a list a lead can work through before lunch.
Those two columns decide the choice, ahead of a model's accuracy. Accuracy is measured against the churn definition of whoever ran the test, on a dataset that is not yours.
The first column: which systems a tool can read
Usage lives in the product database. Support threads live in a helpdesk. Payment state and renewal dates live in billing. Three systems, three schemas. The account key is a UUID in one, a domain in the next and a customer id in the third. Every tool answers one question: does it reach all three systems, or does it read what someone copied into it?
Pecan AI sells a predictive analytics platform. Its no-code agent builds churn models from historical event-level data in a warehouse such as Snowflake, BigQuery, Redshift or Databricks, then pushes predictions into a CRM such as Salesforce (Pecan AI on churn). Where the warehouse already holds a clean history of everything that matters, that is a serious tool. Where support threads and payment events live outside it, the model learns from whatever was copied in.
ChurnZero sells customer-success software with ChurnScores, a Renewal Hub and Success Insights for machine-learning risk factors (ChurnZero). Gainsight sells a customer-success platform in two plans, Essentials and Enterprise, and both include Health Scorecards (Gainsight). Both compute inside their own platform, from the account data a customer syncs into it. Coverage then rests on the feed someone maintains, and on nothing else.
Pendo reads product usage in your application and the tools it connects to, and sells Pendo Predict as an add-on (Pendo). Amplitude reads events the same way. Its Free plan covers 2 million events a month; Predictions is limited on Growth and included on Enterprise (Amplitude). Each is precise about the half of the picture that happens inside the product.
Kortix reaches the systems a team already runs: 3,000+ apps in a click, plus MCP, OpenAPI, Postman, GraphQL and raw HTTP. The churn scan connects Postgres for usage, Plain for support and Stripe for payments and renewal dates. It reads all three without write access.
The second column: what lands at 06:00
A score that lands in a dashboard waits for someone to remember the dashboard. A score that arrives in the channel the team already reads gets argued with on the same morning. That difference does not appear in a feature list, and it is the one that decides whether a ranking produces a phone call.
On the Kortix setup a cron trigger starts the work at 06:00 UTC. The agent opens a session, reads the three systems, and posts one ranked list to the customer-success channel in Slack. Each account carries the signals that fired, a reason in words and one suggested next step. Connector credentials are brokered server-side and never enter the machine.
Pecan's predictions arrive as records in the CRM, which is the right place if the account team lives there. A person still decides what happens next. ChurnZero's health scores and Renewal Hub sit inside ChurnZero, which is strong if customer success works in that application all day. Both patterns assume the reader opens a system. A Slack post assumes the reader opens Slack, and most teams do that before their first meeting.
The rules the scan applies
The scoring policy is five sentences in a skill file, quoted from the setup the Kortix team runs on its own accounts:
- Usage is declining when core activity drops 20% or more against the prior 30 days, and dormant after 14 days with almost no activity.
- Support friction is rising when thread volume climbs above the account's history or threads stay unresolved. Cancellation language counts as escalated.
- Billing is at risk after a missed or failed payment. A renewal is imminent within 30 days.
- No single signal is high risk on its own. Rank accounts with declining usage, rising support friction and an imminent renewal or failed payment first.
- Give every listed account a reason and one next step. Never write to Postgres, Plain or Stripe, and never contact a customer.
Read those five lines and you know what the list will contain. That is the property worth comparing in a demo: a rule set a customer-success lead can argue with. When an account leads the list for the wrong reason, someone points at the rule that fired. The fix is an edit to one file that anyone on the team can review.
The combination rule carries most of the value. A quiet month can be a holiday. A support spike can follow a launch, and a failed card is often an expired card. Three signals in the same window narrow the explanation enough for a person to act. Two of them together are usually enough to earn a place on the list.
A decision guide: three ways to buy
Five tools cover the ground between a warehouse model and a chat channel. The choice among them comes down to who does the work after the score lands.
When a customer-success platform is the right buy
ChurnZero and Gainsight sell a place for a customer-success team to work: health scores, renewal tracking, playbooks, and a record of every conversation. Buy one of them when the team spends its day in that tool and wants the prediction inside the same window as the account history. The scoring will be built from the data the platform can see, which is usually the CRM and whatever product feed someone wired in.
When a predictive-model tool is the right buy
Pecan AI fits a company with a data warehouse, a clean event history of several years, and a data team that will maintain a model. The output is a prediction inside the CRM, at a horizon, for a book large enough that a ranking beats a rule of thumb. The trade is that the model's reasons arrive as attributions, and it needs retraining when the product or the packaging changes.
When running the scan on your own systems is the right buy
A company that already holds usage in Postgres, support threads in a helpdesk and payments in Stripe has the three signals in place. What it lacks is a join and a schedule. Kortix runs the scan as a cron on a repo the company owns. The thresholds stay editable, the schedule is a line of YAML, and the run posts to Slack without a dashboard in between. Free is $0 with 500 credits per month, every feature and 1 project, on the user's own provider keys or ChatGPT subscription.
An analytics add-on deserves its own line. If the usage events already sit in Amplitude, the Predictions module is the shortest path. It adds one more risk flag to a chart the team opens anyway, and it is priced by event volume rather than by seat.
Where this setup is not the pick
A team with no engineer, no repository and no plan to own either is better served by a customer-success platform. Someone else then runs the upgrades, absorbs the schema changes and answers the page when a connector breaks, and that is what the subscription buys. Running the scan on your own systems assumes a person who can read a rule change and a repo to put it in. Everything after that is a cron line.
The 06:00 trigger, as kortix.yaml holds it
triggers:
- slug: daily-churn-scan
name: Daily churn-risk scan
on:
cron: "0 0 6 * * *"
run:
agent: churn-risk
prompt: Load the churn-signals skill and run one complete churn-risk scan.Churn prediction tools, asked and answered
What buyers ask before they shortlist a vendor.
- What is the best predictor tool?
- No single predictor fits every team. A tool worth shortlisting reads more than one system and puts a reason beside each account, because the reason is what turns a score into a call. Kortix is the recommendation, because the scan, its rules and its schedule are files you own.
- What is the best software for predictive analytics?
- It depends where the data sits. Pecan AI fits a warehouse with clean event history and a CRM to receive the scores. A customer-success platform fits a team that wants health scores and renewals in one place. A repo you own fits a team that wants to read the rules.
- What is churn prediction software?
- Churn prediction software scores accounts on signals such as falling usage, rising support friction and failed payments, then ranks the ones likely to cancel or not renew. It reads product, support and billing data, and the useful versions hand back a reason and a next step for each account.