Seasonality
Seasonality is expected, so a landscaping supplier going quiet in January is not a warning.
For the skeptic
Skeptical is the right starting point.
A tool that reads your books and your inbox and warns you about your own customers should have to explain itself. This page covers exactly what ChurnRisk reads, how a warning is produced, where AI is allowed and where it is banned, and what happens when we cannot see enough.
Scope
What we read, and what we never touch
From QuickBooks Online, read-only: customers, invoices, credit memos, payments, and estimates.
We never write to your books, never change a record, and never create one.
From email, only the mailboxes you explicitly connect, minus any labels you exclude: message timing, reply patterns, and the content needed to detect complaints, competitor mentions, and contact changes. We store structured signals and short quotes, and link back to the thread rather than copying your inbox.
We do not read calendars we are not given. We do not touch personal mailboxes. We do not train any model on your data.
The connector pages list this scope permission by permission: QuickBooks and Gmail and Outlook.
Baselines
Every customer compared to themselves
There is no universal definition of a healthy customer. A customer who orders every three weeks is late after five. A customer who orders twice a year is not.
ChurnRisk learns each customer's own rhythm from up to two years of history: order cadence per product family, typical value, quote conversion, payment timing, and reply patterns.
Warnings fire when a customer breaks from their own normal. Never from an industry average.
Exceptions
Why the obvious false alarms do not fire
Four things move the numbers without meaning the relationship is in trouble, and each one is handled as an exception rather than a warning.
A large one-time project
A large one-time project inflates a baseline, so we treat it as an exception rather than the new normal.
Raw material prices
A raw-material price change moves invoice values without meaning anything about the relationship.
A cause on your side
A customer whose silence has an operational cause on your side, like a stockout, is your problem to fix, not theirs to be nagged about.
Exception handling is configured during onboarding and refined as you correct us. Every suppressed warning is still visible in an exceptions view, so "the system decided not to tell you" is never silent.
Four readings
Four dimensions, never one score
They are stored separately, shown separately, and never merged.
Risk
The probability something is wrong.
Revenue at stake
What it costs if it is.
Confidence
How much of the picture we can see.
What to do
Whether acting now would plausibly help, and what the action is.
The right response to high risk at low confidence is to check coverage. The right response to high risk at high confidence with $180,000 at stake is to pick up the phone. One blended number cannot tell you which situation you are in.
Where AI sits
Where AI is allowed, and where it is banned
Language models do four jobs here: extract structured events from text, summarize evidence, explain a risk in plain language, and draft messages for your approval.
They never assign a risk score, and they never decide whether something is a risk.
Scoring is deterministic, versioned, and reproducible: the same inputs always produce the same result, and we can show the calculation.
Every high-severity extracted signal must carry a verbatim quote that matches its source text. If the quote does not match, the signal is automatically downgraded. This is a mechanical guard, not a policy.
Coverage
When we cannot see enough
Customers who order by phone, EDI, or a portal are partly invisible to us. Mailboxes that are not connected leave gaps. A stale connector means stale data.
In every one of those cases, confidence drops and the interface says so. A customer we cannot see is marked unknown, never healthy.
This is enforced in the data model itself and tested on every release, because treating silence as good news is how tools in this category lie.
Limits
The honest limits
What ChurnRisk cannot do: see orders that never touch QuickBooks, read conversations that never touch a connected mailbox, know about a competitor's handshake at a trade show, or guarantee that acting saves the account.
What it can do: make sure that everything your systems already know shows up in one place, ranked by money, early enough to matter, with the proof attached.
Use the working product now.
Create and verify an account, complete onboarding, and reach every screen. QuickBooks can connect sandbox companies only until Intuit approves production credentials.
Real-company QuickBooks access is pending Intuit review.
