NRR Fell 22 Points. New Business Was Fine.
For SaaS founders, heads of growth, and anyone who has spent the night before a board meeting re-checking how retention was calculated.
Net retention has an unpleasant property: by the time it visibly drops, you are three quarters late.
It has a second unpleasant property. You bring the drop to the leadership meeting, and the next hour goes to arguing — not about what to do, but about how the number was computed. Someone wants annual contracts only. Someone says upsell shouldn't count as expansion. Someone finds last year's definition and discovers it isn't this year's.
We published a conversation you can open right now. A fictional SaaS company, 219 customers, 24 months of subscription detail — real generated data, not screenshots.
No signup, no login. Every number below was pulled live by the agent on that page.
First question: how far did it fall
The prompt was: what was net revenue retention over the last 12 months, and how does it compare to a year ago?
The agent didn't lead with a number. It first stated which definition it was using — Bessemer's 2023 formula, (starting ARR + expansion − contraction − churn) ÷ starting ARR — and then laid out the intermediate figures for both windows:
| Line | Current window (2025-08 → 2026-07) |
|---|---|
| Starting MRR | ¥1,211,773 |
| Expansion MRR | ¥429,737 |
| Contraction MRR | ¥158,440 |
| Churned MRR | ¥175,838 |
| NRR | ≈ 107.9% |
The comparison window came out at 130.7%. A fall of 22.8 points in one year.
Worth pausing here. 107.9% is not, on its own, an ugly number. It is above 100%; existing customers are still net-growing. A report that shows only the current figure turns this page. What's alarming is the delta — and a delta only means anything when both windows were computed under the same definition.
The agent flagged the caveat itself: under the BVP formula a single outsized expansion can carry the whole number, and if you cap each customer at 100% you get GRR, which does not belong on the same chart. That wasn't a template we wrote into a prompt. It came attached to the definition the agent used — a definition forked from Bessemer's public standard, living in the workspace as a record you can trace and audit.
Second question: whose fault is it
So is new business weak, or are existing customers shrinking? Break it out by segment.
This is every CEO's next sentence, and it is the one most BI dashboards can't answer — because it doesn't need another chart, it needs a recalculation.
Three findings:
New business is fine. Enterprise new MRR ¥320k, mid-market ¥210k, SMB steady. Nothing collapsed at the top of the funnel.
Enterprise is the engine. Expansion MRR of ¥230k, far above contraction plus churn. Net retention on existing enterprise accounts ≈ 117%.
Mid-market is leaking. Contraction of ¥44,717 — the highest of the three segments — against expansion of only ¥75,381, almost exactly consumed by contraction and churn. Net retention on existing mid-market accounts ≈ 83%.
Now go back to the two ratios from the first question and the story closes. In the comparison window, contraction plus churn was 9.5% of starting MRR. In the current window it is 27.6%. Outflow nearly tripled. The 22.8 points did not come from weak acquisition. They came from mid-market customers walking down.
At that point the required action has changed. It is no longer "increase marketing spend." It is "list the mid-market accounts with the largest contraction." Those two cost an order of magnitude apart.
Third turn: make it something you can send
The last message in the conversation was: turn these two findings into a data video for the leadership team.
The agent produced a 3 minute 23 second two-host video, 16:9, ready to play in a meeting room. It did not invent figures. Every number in the script has to trace back to a query this conversation already ran; anything unsupported gets refused. You can play it on the page.
We'll be blunt about why this matters. There is no shortage of tools that will have a language model write you a video script. Their problem isn't that the output looks bad — it's that the numbers have no origin. For material going to a board, looking good is optional; being traceable is not.
What you're actually looking at
That page looks like a chat log. It is four things stacked on top of each other:
- A dataset — raw subscription detail for 219 customers, re-queried on every answer, not a cached conclusion.
- A definition — NRR as a standalone record forked from a public standard, with an audit trail. Anyone can ask "how was this calculated," and the answer is the same one.
- A line of reasoning — two turns, where the second builds on the result of the first.
- A deliverable — a video whose script is forced to line up with the numbers already pulled.
Remove any one of them and the rest stops working. Data without definitions puts you back in that hour-long meeting. Definitions without data is a slide deck. Both, but only if you personally run them each time, means you are the bottleneck.
Try it with your own numbers
At the bottom of that page there's Remix with my data. Download the template, fill in your own subscription detail, upload it. The questions and the definitions cross over; the data does not. You get an agent in your own workspace, pointed at your own table.
The person who shared the page never sees a single row of yours.
Luxiva turns a spreadsheet and a definition you can defend into an agent your whole team can question.