Do More by Doing Less
Growth should create leverage, not more administrative work.
If revenue doubles and your finance team is just reviewing twice as many receipts, that’s not scale. That’s linear workload growth.
That’s why I use Brex. Brex built an Intelligent Finance platform with AI agents that automate repetitive work — receipts, categorization, policy enforcement, reconciliation — so that finance output improves as the business grows, and I can focus on scaling the MM empire.
If you care about operating leverage, your finance stack should reflect that.
See why 35,000 companies like Anthropic, DoorDash (and me!) use Brex to spend smarter and move faster.
When you beat plan on recurring EBITDA adjustments.
Yo, it's CJ! I’m hosting a happy hour in NYC. Come hang with fellow Mostly Metrics readers on Wednesday April 8th.
I’ll be telling my favorite LTV to CAC riddles.
RSVP below. It’s at an undisclosed location guarded by my ferocious Berne doodle Walter.
Really pumped to meet you IRL
Reading the Winds
Bruno Annicq is a multi-time CFO who's forecasted across wildly different business models: from AOL's legacy ad-based business to Wellhub's B2B2C fitness marketplace.
But when he explained his forecasting philosophy to me, he didn't start with spreadsheets or ERPs. He started with water sports.
"I do a lot of water sports. Like kite surfing. Wind is really important. And if you look at how the best wind apps show you what the wind is going to be, they don't show you one line. They show you multiple. They're showing you the actual output of different models."
Here's the deal with weather forecasting: there's a big North American model, a different European model, and a bunch of regional ones that are strong locally but weaker at the macro level. Some nail short-term predictions, while others are better at longer-range.
So the best wind apps don't pick a winner; they show you all the lines at once so you can deduce a line of best fit.
"When I look at that and I see tomorrow at 5:00 PM, the wind is going to be 15 knots from the east — and all the forecasts are close together — I can have a lot of confidence. But if some forecasts think it's going to be 20 knots and others say 5 knots from the west? I attach a lot less certainty to that. And that's a very valuable input."
When the lines converge, you go out. When they diverge, you stay home. Also, I’ve been kite surfing exactly 4 times and almost drowned.
Most CFOs wouldn't take forecasting advice from a wind app. Bruno rebuilt his entire approach around it, and cut forecast error by 80%.
Why Wellhub Breaks Traditional Forecasting
To understand why Bruno needed a different approach, you have to understand Wellhub (formerly Gympass’s) business model. It's a bit of a brain twister.
"Our business is a B2B business. We sell a platform to clients — companies — but on top of that, we have a B2B2C marketplace. Once we sell into a company, their employees can sign up for one of our plans and use our partner network. We pay partners based on that usage."
TL;DR:
Companies pay Wellhub.
Employees subscribe.
Employees visit gyms.
Wellhub pays the gyms based on those visits.
Lots of matching. Lots of moving pieces. Lots of variability. 5 dimensional chess type stuff.
The tricky part is utilization — how often employees actually show up and use the network. That's not something you can pull from a CRM. It's subject to forces that most finance teams never think about.
Like (wait for it) weather.
"In some countries, when the weather is bad, nobody goes to the gym. In other countries, bad weather means everybody wants to go because it's indoors."
Gotta love that. Your forecast accuracy depends on whether it's raining in São Paulo.
At Wellhub's scale, small misses get expensive fast. Being off by 1% means tens of millions of dollars. When Bruno joined, they were running about 9-10% MAPE — Mean Absolute Percentage Error — in either direction.
The only sane response? Get conservative. Plan for worst-case utilization. At least you won't get caught with your pants down.
But here's the thing about conservative forecasting: it has a cost.
You don't hire the engineers you could have.
You don’t expand into new geos and hire more sales reps.
You leave cash on the balance sheet because you couldn't see clearly enough to put it to work.
Bruno needed a better way.
The Messy Middle
Before we get into what Bruno built, let's talk about why normal forecasting wouldn't work here.
Most finance teams run on deterministic forecasting. Price × Quantity. You count your inputs — reps, pipeline, deal sizes, sales cycles — and you math your way to the output. It can get complicated, sure. A lot of lookups off an assumptions tab at the front of the model. But at the end of the day, you're going to get the same answer every time. It's auditable. Everyone can follow along.
This works great for traditional seat-based SaaS. Contracts recur. Licenses have known costs (marginally zero). You've got decent visibility into your base and can make educated guesses on renewals and expansion. It's multiplication. Finance 101 stuff.
But deterministic breaks down when your business gets weird.
High variability in inputs? Problem. External factors outside your control? Problem. Usage-based dynamics that don't follow neat little patterns? Big problem.
Wellhub has all three. In spades.
I think of this as the messy middle — that zone where your business is too complex for rules-based forecasting but too structured to just wing it. Marketplace dynamics. Weather (yes, weather). Employee behavior that shifts by country and season. None of that fits in a spreadsheet formula. But it's not random either. There are patterns. You just need a different way to see them.
That's where probabilistic forecasting comes in. Instead of spitting out one number, you're modeling a range of outcomes based on what the data is telling you. It's not "here's what will happen." It's "here's what's likely, and here's how confident we should be."
Bruno looked at Wellhub and realized: we're living in the messy middle. We need to think in probabilities and ranges.
The Ensemble
So Bruno did what any reasonable CFO would do: he called up multiple vendors and said, "Build me competing models."
Today Wellhub runs four models, independently developed by two different vendors, all predicting where the business is going to land. They put them together — just like the wind app — and look at where the lines agree and where they don't.
Then they add their own finance judgment on top.
Bruno calls this an "ensemble." (He credits ChatGPT for the term. Asked it how to describe the approach and it said, "Well, it's a multimodal forecast, so it's an ensemble forecast." Nice.)
The magic is in the convergence.
"It's so fascinating to see when they agree and when they don't. Especially when there's something more extreme — like people are predicting a lot of wind — you can see more variability because it's outside of the normal. They don't always agree on where it's going to come from and how it's going to hit."
Think of it like a hurricane's cone of uncertainty. Except Bruno doesn't want just the cone — he wants to see the actual lines of each model.
When models converge → high confidence, lean in.
When models diverge → low confidence, dig deeper.
The results are astounding. What used to be a 10% margin of error is now around 2%. Sometimes even less.
The Human + The Machine
The tech gets a lot of credit, but it’s not completely "let the robots do it."
The ensemble approach only works because Bruno built a team that can actually engage with the models — not just receive their outputs. They have hard skills.
"I hire people that can talk in the native language of the people that build these models. Mostly they use Python. When we hire a firm to build a model, we insist: 'No, I want you to not just give me the output and something that's a black box. I want you to feed us the actual model. I want you to teach my team so they can keep it up to date, add dimensions, retrain it over time.'"
Every day, every week, new data comes in. The team reruns the models. The models get smarter. It builds on itself. It's a flywheel.
And it's important to note where traditional tools still fit. They still use Anaplan — but it's downstream in the pipeline, rolling everything up across revenue and costs. The actual prediction work happens at the most granular level, in Python, with the team's hands on the wheel.
The Explainability Payoff
Now here's the elephant in the room with any AI-driven forecasting: explainability.
The knock on probabilistic models is that they're black boxes. They might be more accurate, but good luck explaining to your board why the number is what it is. And in finance, "trust me, the model said so" doesn't fly.
Bruno gets this.
"Especially in a marketplace where you have some control over what happens, you want to understand what's driving the behavior so you can potentially change it if the outcomes aren't what you want."
Here's where the ensemble approach gets clever. Running multiple models doesn't just improve accuracy — it actually improves explainability.
Because some of the models are more interpretable than others.
You can use the "simpler" models to explain what the "complex" models are seeing.
You can change single variables and watch how the lines move.
You can isolate which drivers matter most.
"By having multiple models, some of which are more explainable, you can have that model say it's probably because of these drivers. It helps with the explainability."
Real example: Wellhub has a fraud team. They don't need to understand every nuance of the complicated models. They just need to know where to look. The ensemble tells them, "The dimension to investigate is X." And they go.
Most CFOs think AI means trading explainability for accuracy. Bruno's showing you can have both — if you're willing to build the infrastructure and hire the people willing to work within it.
When You Need This (And When You Don't)
Look, not every business needs to run four models in parallel.
If you're selling $50K/year SaaS seats with 12-month contracts, deterministic forecasting will get you there. Price × Quantity. Count your reps. Model your pipeline. Sleep easy, my friend.
But if your business lives in the messy middle — usage-based revenue, marketplace dynamics, external dependencies like weather or macro or human behavior — then you're fighting with one hand behind your back.
The ensemble approach gives you something rare: accuracy and explainability. You’re dealing with confidence bands with traceable factors.
"I love seeing the actual lines of the models when a team presents their forecast."
Not one line. All the lines.
The wind don’t care about your spreadsheet. But you can still learn to read it.
Quote I’ve Been Pondering
I been sellin' dreams to sleepers,
Tell the truth, that's the perfect business
'Cause in the drought, I was payin' double
For some work that wasn't even worth the ticket
Hoping you can pronounce ‘probabilistic’, bc I can’t
CJ
