Fire 40% or Grow 40%? The Three Vectors of AI — Notes from Production

Imagine AI just freed up 40% of your team's time. Congratulations — now comes the decision that defines everything else. In a recent interview on Arpa Talks (in Spanish), Pep Martorell — deep-tech investor, partner at InvivoAI and Esade faculty — puts it without anesthesia: "You can basically do two things. One: if 40% of the time is freed, I cut 40% of the people. It's easy, and I capture the value through lower costs. And that's where it ends. The other, more sophisticated and complex: turn that freed time into more business, more clients, more product. Instead of reducing from below, grow from above."

The counterintuitive part is what is actually happening: Martorell cites companies that reversed layoffs "because AI costs were perhaps not what they seemed," and recent studies correlating AI intensity with headcount growth. It's the radiologist paradox: algorithms have read scans well for years, and yet there are more radiologists than ever — because automating the task multiplied the demand for diagnostics. The economist Jevons described it in 1865 with coal: when something becomes more efficient, we consume more of it, not less.

But the 40% is just the entrance. The most useful part of the interview is the map Martorell draws to explain himself — "they are three different things, and if we mix them up we'll get dizzy," he warns the interviewer, walking through them one at a time: AI is changing companies along three distinct vectors. (Hold on to that "one at a time" — there's an objection worth raising later.)

Vector 1: process automation

"It's what we've always tried to do with every new technology," he says — adding the nuance almost nobody admits: "there are a lot of people talking, and then implementing less, because this is genuinely hard."

Here are our receipts, because this vector is our daily bread. When we automated document validation for a US insurtech, the system ran side-by-side with human analysts for two weeks until it reached 98% agreement with their answers — and the outcome wasn't firing analysts: the expert stopped reading entire documents and started validating flagged findings. Their judgment — the truly expensive part — concentrated where it matters. Same with the quoting engine that turns a pasted free-text list into a formal quotation: the sales team didn't shrink — it started answering the same day and handling volume that used to be impossible. Capturing value "from above," as Martorell puts it.

Vector 2: the agentification of work

Here the interview gets serious. An agent, in Martorell's definition, is an AI that "doesn't just answer your questions — it has a will of its own to do things." And the mindset shift it demands: "stop being an executor of sequential tasks and become an orchestrator of agents. Besides your human team, you will start to have an agentic team working 24x7."

His firm lives it: at Invivo Partners, the answer to "how many are you?" is "15 + 1" — fifteen people and Mother, their agentic colleague. And here is the documentary gem: Luis Pareras, the managing partner, published the most honest technical paper we have read about an agentic colleague in production (link in the sources). It details the full architecture: an actor–critic–boss loop running on three models from different labs, because "a single model cannot effectively criticize itself: it shares its own priors and blind spots." It details the two-pass fine-tuning — first the sector corpus, then the partner's taste ("AIA likes the companies I would like, before I have read the deck"). It reveals that Mother has a constitutional file with her goals — and that she once answered "no" when he tried to rush a document. She was right.

We reached the same conclusions through practice: our sales copilot listens to the video call and whispers cues in ~3 seconds from a private knowledge base — and its most important rule is not what to say but when to stay quiet. And we use consensus panels across models from different families for hard decisions, for the same reason as Pareras: different models' errors don't overlap. When two practitioners who have never met converge on the same architecture, it usually means something real lives there.

Vector 3: disruption

The most demanding one: "it's not doing things faster than before; it's doing things you couldn't do before." Martorell is honest — today this is only clearly visible in science, and it will reach business later. His difficulty scale orders the whole map: automating (we've done it for decades) < agentifying (it will take effort) < disrupting (harder still).

A necessary objection: the vectors only separate on the whiteboard

Martorell explains them separately so nobody gets dizzy — it's an expository device, and as a taxonomy for organizing decisions it works: automating is a project with an ROI, agentifying is a mindset change, disrupting is a strategic bet. But it's worth saying out loud what the separation hides: in reality, the three vectors happen together and feed each other.

Our own work proves it. Which vector does a sales copilot that listens to live video calls fall into? It automates the meeting minutes and the data lookups (vector 1), it is an agentic colleague that decides when to speak and when to stay quiet (vector 2), and it gives a salesperson their company's complete memory in real time — something that previously didn't exist at any price (vector 3). One system, all three vectors at once. Mother, Invivo's agentic colleague, likewise: she triages decks, she is the team's "+1", and she changes the economics of investment analysis to the point where Pareras argues it will alter the structure of venture capital itself. The vectors aren't three boxes: they are a gradient where each one enables the next — automation generates the data and the confidence that make the agent possible; the agent running 24x7 accumulates capabilities that end up being disruption.

And there is something more uncomfortable: the vectors don't ask for permission. They are not something your company "does or doesn't do" — they are something happening to your market. If you don't automate, your competitor multiplies capacity and demand routes toward them: the radiologist paradox also works at the industry level. The six-month window cuts both ways — if you don't open it, someone is opening it over you. Not adopting isn't a neutral state; it's occupying a position on someone else's timeline. Pareras says it about his own peers: "most have not metabolized the arrival of AI yet; they are about to be forced to." Forced — not invited.

The six-month window — and the eternal-pilot trap

The pragmatic manager's question remains: why build today what a vendor will sell packaged a year from now? Martorell gives the most complete answer we've heard. First: "the cost of failing is practically zero... and building pilots takes you through a process of maturing your thinking about the business that is priceless. The mere act of doing it raises questions about your business you had never asked." Second: "in certain businesses, being 6 or 12 months ahead of the competition is brutal. Everything I've done in that time, nobody can take away from me." Pareras says it even more bluntly in his paper: his system's first three months were embarrassing, by month six it matched his analysts, by month eight it exceeded them — "the early mediocrity is the price of admission, and most of your competitors are going to refuse to pay it. That is your window."

But the window has one condition, and Martorell names it plainly: "What you cannot do is become an eternal innovator of pilots that never ships to production." To show where the trap lives, he uses a case that became famous: the CTO of a major retail chain publicly shared that two people on his team built, in one weekend, an AI search engine for their online store — and it worked better than the one from their contracted vendor. Spectacular. So do you cancel the vendor's contract? That's where the real debate begins, Martorell warns: to actually replace it you need a plan B if it fails, someone who signs terms and conditions, a team to maintain it, cybersecurity, and updates at the pace a vendor guarantees. The weekend produces the demo; none of that comes included. Gartner put a number on the trap: over 40% of agentic AI projects would be canceled over unclear business value or insufficient risk controls.

The synthesis we take home — and that we defend case by case: the six-month advantage is only cashed in if the experiment ships to production. Experiment cheaply, mature it with feedback from the people who know your business best. Then scale: bring in an owner, the security, and everything production demands.

One practical closing note: that second half doesn't have to be done by the same people who built the pilot. When an experiment proves it deserves to live, leaning on experts is simply sensible — it's the shortest path to cashing the window without paying the full learning curve. That is, case by case, the work we document.