Revenue scores you can audit, backtest, and argue with.
Every score is readable arithmetic with declared weights and published caveats. Every prediction is written down when it is made, graded once the outcome settles, and frozen. And when there is not enough evidence yet, VERA says so instead of showing you a green light.
For the people who have to defend the number.
Published weights · Graded predictions · Honest when the evidence is thin.
Northwind Labs
Health score, and how it got there
- Product health× 0.503.1
- Engagement× 0.305.0
- Relationship× 0.207.5
Adoption is not mapped for this tenant. Scored neutral, never counted against them.
Called it. Flagged at risk 62 days before the renewal. Outcome: churned.
Reads what you already have
Numbers you can argue with
Every score here is arithmetic you can read, with the evidence under it and the honest limits attached. When the evidence is thin, it says so.
Scores you can read
Health and churn risk are an explicit weighted formula, not a model nobody can inspect. Declared weights, published thresholds, and every number expandable to the evidence under it. If a score cannot be explained in a paragraph, it does not ship.
An AI that gets a vote, never a veto
Sentiment enters the health score as a modifier capped at one point in either direction. It can nudge a number. It cannot overrule the formula, and confirmed churn caps the score no matter how good the arithmetic looks.
Predictions that get graded
Every risk call is written down the day it is made, graded once the outcome settles, and frozen. Correct a close date in your CRM afterwards and the historical grade does not move. Most tools will never tell you whether last quarter was any good. This one keeps the receipts.
Honest about what it does not know
"Not enough evidence yet", "we could not read this", and "we looked and it is fine" are three different answers, and VERA keeps them three different answers. A thin sample says so instead of showing you a confident green.
Signals that earn the interrupt
A signal has to justify the notification before it gets one. Deduped per person, capped per window, opt-out honoured in one place, and a new signal defaults to a surface you already visit rather than a message you did not ask for.
Reads your stack, does not own it
A method reads logical fields. You map them to whatever your CRM calls them. No method hardcodes a field name, so the same arithmetic runs on Salesforce or HubSpot and stays comparable across both.
If any of this sounds familiar
These are the problems VERA was built around. Not job titles, arguments you have already had about a number nobody can defend.
You inherited a health score nobody trusts
Everything is green and accounts still churn, so the team stopped looking at it. Start from a score you can read, and a record of when it was right.
Your forecast is a feeling
Rep confidence goes into a spreadsheet and nobody grades it afterwards. Keep every call, score it at fixed horizons, and find out whether you run optimistic.
You are asked to defend a number
A board deck is a bad place to discover you cannot explain where a figure came from. Every number here expands to its inputs and its caveats.
You own the scoring nobody agrees on
Scoring debates never end because nothing is testable. Put the method in version control, backtest a change before shipping it, and let the argument be about evidence.
Compute once, write it down, grade it later
This is the life of a single score. It is calculated in one place, recorded the day it is made, graded against what actually happened, and only then put into words. The last step never gets to change the first.
One place, deterministic
Every score is a pure function with declared inputs and unit tests. It runs in exactly one place, because two implementations of the same rule will always drift apart eventually.
Append-only
Scores and predictions are written down when they are made, not re-derived on read. That is the difference between a track record and a number that quietly changes when someone tidies the CRM.
After the fact
When an outcome lands and settles, the prediction is graded against it and frozen. Coverage travels with every rate, so you never read an accuracy number without seeing how many cases it rests on.
Narrate, never decide
AI puts a computed result into words. It never produces the number. Anything a model says can be traced back to arithmetic you can check yourself.
Why you can trust the number
Two questions decide whether a score is worth anything. Can you audit how it was reached, and can you trust where its inputs came from? Both answers are structural here, not policy.
The method is readable
Weights, thresholds and caps are published, not buried in a rules engine. You can disagree with a number on the merits, which is the only kind of disagreement worth having about a score.
The track record is frozen
Graded predictions are append-only and immutable. Nothing recomputes them on read, so a correction upstream can never quietly improve how the system looks in hindsight.
Every rate ships its denominator
Accuracy figures travel with the number of cases behind them, and a sample too thin to support a claim reports insufficient evidence rather than a percentage.
Questions, answered
You check. Every risk call is stored the day it is made and graded once the outcome settles, so the system publishes its own hit rate, how much warning it gave, and how many cases that rests on. If the sample is too thin to support a claim, it says so instead of quoting a percentage. That is a question almost no scoring tool will answer about itself, and it is the first one you should ask any of them.
Could you defend your health score in a board meeting?
Bring us the score you already have and the calls it got wrong. We will tell you what we would measure differently and how we would grade it. If that turns into software, good. If it turns into a better spreadsheet, also good.