The answer is 14. Now what?
AI will build you 87 flawless charts before your coffee is cold. The old data hands knew the only question that mattered came after the number, and nobody is asking it anymore.

Dashboards are having a moment.
Open any AI group chat. Somebody handed a model eleven years of Strava running data and got back 84 interactive charts. Somebody else wired up the smart fridge and now knows, in real time, the household’s oat-milk-to-cow-milk ratio. There is a guy with a live heatmap of his dog’s naps. Thanks, AI.
To be fair, it’s impressive. The models got very good, very recently, at turning a pile of data into something that looks like a design team made it. The charts are right. The colors are tasteful.
Nobody in the chat asks the follow-up. What did you do differently on Tuesday because of chart 61?
That’s a hobby, and a hobby doesn’t need a business case. The trouble is that the same reflex just showed up at work with a warehouse connection.
The answer is 14
Anyone who did analytics in the pre-Frontier AI era knows this Slack message by heart:
“Can you put together a dashboard that shows the revenue impact when the segment is X, Y and Z?”
The junior analyst said “sure” and vanished for two weeks. The dashboard shipped. It got opened four times, twice by the analyst.
The old hands had a different move.
“Funny, I ran that last week. That segment runs about 14% above average. What are you going to do with that?”
Silence, usually. Sometimes for days.
The gambit works because the request was never for a dashboard. It was a question, and under the question there was supposed to be a decision. Hand over the number in one line and the decision has nowhere to hide. Either there is one – move the budget, change the price, kill the segment – or there never was, and you just saved two weeks.
That analyst was friction, and the friction was the service. Now the same Slack message goes to a model, and the model does not ask what you plan to do. It builds the dashboard. Twelve tiles in four minutes: MRR, churn, CAC, LTV, a funnel, a cohort heatmap. Every number reconciles. It looks exactly like the one your competitor generated on Tuesday.
We watched the same thing happen in sales, where engineering used to say no for you. This is the analytics version. The person who said “now what?” got automated out of the loop.
The numbers that survive “now what?” have made-up names
Early Slack had a signup number, and it failed the test completely. More than 90% of the people who filled out the form never invited anyone or started using the product. Signups are up. Now what? Nothing.
So the team went looking for what separated the teams that stayed from the ones that left. The answer was 2,000 messages: “regardless of any other factor, after 2,000 messages, 93% of those customers are still using Slack today.” Stewart Butterfield’s conclusion: “Only you can really determine your company’s magic numbers” (‘From 0 to $1B’, First Round Review, February 2015).
Now what? Get every new team to 2,000 messages. That’s a number with a job.
Webflow’s version was a site with 10 views. Not an account. Not a published page. A site somebody else had looked at ten times, which Melissa Tan, who ran growth there, called “a clear success metric that meant you are more likely to retain” (‘How to design a high-impact growth org for a PLG startup’, First Round Review, June 2023).
Neither of those is in a KPI glossary. They aren’t definitions. They’re findings: somebody did the work, and the work said this condition, in this business, is the one that predicts whether the customer is still here next year.
That is what a number that runs a company looks like. A term somebody invented, sitting on top of a calculation nobody outside the building would think to run.
It can’t generate what was never written down
Ask a model for a SaaS dashboard and you get the SaaS dashboard, because “SaaS dashboard” has appeared next to “MRR, CAC, LTV, churn” ten thousand times in what it read. Nobody chose those tiles. They’re just the ones that show up.
Google researchers describe the mechanism: models train on text with a steep power-law distribution, scale has improved the average case, and “persistent failures on low-frequency, domain-specific, cultural, and temporal knowledge remain poorly characterized” (‘Long-Tail Knowledge in Large Language Models’, Google Research, 2026).
Your magic number isn’t even in the long tail. It was never in the training data at all. It’s in your warehouse, undiscovered, or in the head of the one analyst who found it.
Don’t expect a different model to reach for something different. Across 22 LLMs on standard creativity tasks, “LLM responses mirror other LLM responses far more than humans do other humans” (‘Large language models are homogeneously creative’, PNAS Nexus, March 2026). That wasn’t a dashboard experiment. It is the same pull toward the middle we covered in AI makes you average.
Same industry, different number
Two SaaS companies land in the same row of the same benchmark report. One raised $80 million and will lose money on every new logo for two years, on purpose. The other is bootstrapped and exists to send its owner a distribution every quarter.
Whatever predicts winning for the first is close to irrelevant for the second. Industry is a classification. Strategy is the thing the number has to answer to.
The generated dashboard only knows the classification. You asked for yours and got somebody else’s.
Green tiles, wrong direction
The familiar numbers can all improve while the business gets worse.
Microsoft’s experimentation team documented two. Raise prices and short-term revenue likely goes up while lifetime value goes down as users leave. Make search results worse and people search more, so query share rises while abandonment climbs (‘Pitfalls of Long-Term Online Controlled Experiments’, Microsoft Research, December 2016).
Revenue up. Engagement up. Both tiles green. Both pointing the wrong way.
The hunt is the part AI is good at
Here’s the part that should annoy you. Finding the magic number used to be expensive: months of an analyst’s time, six data sources, a standing argument about definitions. That is exactly the work that just got cheap. And we’re spending it on heatmaps.
Airbnb’s data scientists had the standard quality tile, the star rating, and knew it was close to useless: reviews skew so heavily to five stars that, in their words, “we cannot purely rely on the reviews stars.” So they went hunting. They ran listing descriptions, review text and photos through embedding models, clustered the results, and kept the clusters where guests rebooked within a year at unusually high rates. Then they turned what set those clusters apart into one number per listing, “crafting a novel metric” they named the listing quality score.
Tested on 500,000 listings and 800,000 stays it had never seen, the score tracked one-year return rate more tightly than star ratings or review sentiment: a 0.903 correlation across score buckets. And it came with its own “now what”: the paper proposes using it to decide which listings get featured on marketing pages (‘Understanding Product Quality with Unstructured Data’, Airbnb working draft, KDD workshop, 2024).
That is the job. Not “chart my revenue.” Tell the model what winning looks like, hand it everything, and make it come back with: this is the thing driving your north star, and here is the evidence.
Then make it prove it
A number that predicts the outcome is a candidate. It is not yet a lever.
Slack saw that teams past 2,000 messages stayed. That is not the same as knowing that pushing a team past 2,000 messages makes it stay. Announce the target, and somebody ships a bot that says good morning in every channel. The number moves. The thing it used to mean is gone.
Google’s recommender team shows what closing that gap looks like. Researchers tested a broad set of behavior patterns and found the subset that predicted how often users would visit five months later. Then they made the system chase those behaviors and checked, in multiple live experiments, that the long-term outcome actually improved (‘Surrogate for Long-Term User Experience in Recommender Systems’, KDD 2022).
Skip that step and you have built a new way to fool yourself.
The hunt gives you a suspect. The experiment gets you a conviction.
Start from the objective
The order of operations is backwards. Forward from available data gets you fifty tiles in four minutes. Backward from the objective gets you five, and some of them need a name you’ll have to invent.
Then give the number a brake. Microsoft’s experimentation group notes that the big, slow measures everyone already tracks “are poor indicators of success, but they are good guardrails” (‘STEDII Properties of a Good Metric’, Microsoft Research, April 2022). That is where the standard tiles belong. Not at the center of the screen. On the edge of it, watching for the day your clever number starts climbing while customers quietly leave.
AI can do the heavy lifting on all of it now. It will not volunteer. Ask for a dashboard and you get a dashboard, and nobody on the other end is going to play the old hand for you.
So play it yourself. Before the next one gets built, answer the question in a sentence and see what’s left.
The answer is 14. Now what?
Zwischen

