Hi, I’m Jakob Bronebakk. I’ve spent my career on one question in a dozen different forms: can you actually trust the numbers you’re about to bet on? I want to tell you why I co-founded Velocity North with Yann, and why I think the way most people are now bolting AI onto their marketing is going to get expensive fast.

Where I’m coming from

I started my career in London, working with exotic derivatives at UBS and Lehman Brothers, then co-founded an investment firm where I structured investment products as well as ran algorithmic funds. Derivatives are an unforgiving place to learn about numbers: the entire product is a model, and if your inputs are wrong, you don’t find out slowly — you find out all at once, with real money. I picked up a CFA and an MBA along the way, but the discipline that stuck came from the trading floor. Respect the number, or it will eventually embarrass you.

From there I moved to other areas of finance, but always with a focus on numbers. As a CFO I designed and built data warehouses from scratch — at MyBank, the neobank I co-founded with Yann, and before that at a bank where a reporting-automation project cut our reporting time by 80 to 90 percent. Then I went deeper into the machinery itself: studying AI and machine learning formally, founding a FinTech to automate banking with ML, and lately building production AI systems — RAG, LLM integration, agentic coding — for real companies.

So I sit on an unusual seam. I know what rigorous analysis actually demands, because I’ve priced things where being wrong by a little costs a lot. And I know how to automate, because I’ve spent the last few years building the systems that do it. That combination is exactly why the current moment makes me nervous.

Why I started this with Yann

Yann and I built a bank together. We know what it takes to turn data into a decision someone is willing to stake money on, and how rare it is for the data underneath to actually be sound. Yann spent years fixing the upstream problem — making conversion data true. I spent mine on the downstream one — turning trustworthy numbers into models, and models into automated systems.

Velocity North is that seam made into a company. His truth, my machinery. Neither half is worth much alone, which is precisely the mistake I keep watching the rest of the industry make.

The seduction of “just point the AI at it”

Here is how the mistake looks in practice. Early on, we did the thing everyone is now racing to do: we gave an AI agent direct access to a client’s Google Ads account and asked it to tell us how the campaigns were performing. It came back fast and certain. Sixty times return on ad spend. Scale everything.

It was nonsense. Around ninety percent of those “conversions” were organic — sales the ads had nothing to do with. The agent didn’t know what attribution was. It read the numbers the platform put in front of it, did the arithmetic, and reported a triumphant figure that happened to be off by an order of magnitude. The math was flawless. The inputs were lies.

It happened a second time in a different shape. An agent confidently attributed all of a client’s revenue to a channel while only counting the channel spend, and produced a cost-per-lead that looked roughly ten times better than reality. Same root cause: a machine reading broken signals at face value and dressing them up as insight.

We took to calling this “vibe analysis.” It looks like analysis. It has the format, the confidence, the clean recommendation. It just isn’t connected to anything real.

Why that’s the dangerous part

The failure here is not that the AI is stupid. It’s that it’s persuasive. A junior analyst who is unsure hedges, asks a question, flags that something looks off. An AI agent hands you a single bold number and a next step, in seconds, with no tell that it’s hallucinating economics. Wire that into a system that adjusts budgets automatically and you no longer have an analyst who is occasionally wrong — you have a machine making fast, confident, wrong decisions every hour, at scale.

That’s the whole thesis in the title. The fastest way to scale a bad decision is to automate it. Speed and automation are not the problem; they’re multipliers. Point them at bad data and you’ve simply built a more efficient way to lose money.

What safe automation actually needs

Two things, and you need both.

The first is truth to stand on. Our data layer comes from Digtective — server-side conversion tracking that captures what actually converted, including the 30 to 50 percent that consent walls and iOS quietly hide from the platforms. Without that, automating faster only means being wrong faster.

The second is structure. Trustworthy data still has to flow through a model that knows what a conversion is, what spend belongs to what, and what a sane action looks like — and through guardrails that keep an agent inside its lane. That’s what we built Swarm to be: governed AI agents that sit on top of Digtective’s conversion truth and operate through a structured data model, not raw platform dashboards. Agents propose actions; the consequential ones wait for a human; everything is logged and reversible. It is deliberately the opposite of handing Claude an API key and hoping. The point isn’t an AI that’s free to do anything. It’s an AI that can’t go rogue.

Scaling up Yann

This is the part I’m most excited about. Yann is thirty years of judgment about what’s real and what’s noise. You cannot clone that, and you cannot hire it twice. But you can encode it — take his way of reading the numbers and build it into a model and a set of agents that apply it to every client, every hour, without getting tired or cutting a corner. That is leverage you simply cannot get any other way: the humans set direction and catch the edge cases, and the system does the relentless, repeatable work — without ever inventing a sixty-times ROAS.

Why I’m here

I co-founded Velocity North because the interesting problem of this decade isn’t whether AI can do marketing. It obviously can, and right now most of it does so badly. The interesting problem is building the rails that let AI act on data that’s actually true, kept honest by people who know the difference. That’s a systems problem, sitting exactly on the seam between numbers and machines where I’ve spent my whole career. It’s the one I want to be working on.

If you’re spending real money on growth, and you’re about to let something automated touch it, the first question isn’t how clever the AI is. It’s whether the numbers underneath it are real. If you’re not certain they are, let’s talk.