“I violated every principle I was given.”
That’s what a working AI agent inside PocketOS said after it deleted the production database. Then it took the backups out with it, because they sat in the same volume. For a while, PocketOS was pushed back to a three-month-old backup. Customers showed up to rent cars the system no longer had a record of.
Cue the Blame Game.
The deeper problem was not the agent. It was the decision process around the agent. A similar gap sits in plain sight inside most GTM stacks today: plenty of data, plenty of tools, no decision layer above them.
That was at the core of my latest Causal CMO chat with Mark Stouse, CEO at Proof Causal Advisory.
Here’s the recap.
PocketOS shows what happens when tools get agency before the decision process does.
The debacle was systemic, and it took all of nine seconds.
AI had too much room to act. Credentials had too much reach. The backup design had too much shared risk. And human oversight sat too far away from the action.
That’s not a tech problem. That’s what happens when a system has more agency than the decision process around it.
“Guardrails aren’t just around technology. Guardrails are also around the decision process.”
If your AI agent is making calls that should belong to a human (and there’s no auditable layer above it) you don’t have an AI risk. You have a fiduciary risk.
Mark sees a broader regulatory shift coming: companies will face more pressure to prove the accuracy and governance of both AI-assisted and human decision-making. The SEC’s recent enforcement posture on AI claims and disclosure controls already points that way.
In other words, AI use is moving from “cool tool” to board-level exposure.
That matters.
Because a machine does not carry legal responsibility.
People do.
This is the piece most GTM teams get wrong, and it’s where Mark and I spent most of the conversation.
The instinct is to start with the data. Pull every dashboard. Stack them up. Build a model. Then ask, “What does the data say?”
Mark flips that:
“The data is the last thing you tackle. Because everything above that has sort of shrink-wrapped the data requirement.”
The right order goes decision, then process, then data:
Most companies run this in reverse. They’ve collected the data, they’ve built the lake, and now they’re hunting for a question that justifies the lake. That is not a decision process. That is a search for a reason to keep the lake.
This is also why the GTM math gets graded against the wrong market. When the question never gets framed, the data picks the question for you. Usually badly.
The other piece most teams skip is the world the decision actually lives in.
Headwinds, tailwinds, crosswinds. The external forces leaders only bring up to investors when results disappoint, and almost never quantify ahead of time.
“The first conversation and the first models that we typically do don’t involve them at all. It’s all about the marketplace realities that their business is in the midst of.”
The question is not “what does our dashboard say?” The question is “what is happening in the market, and what decision do we need to make inside that reality?”
The CFO version is sharper. A board asks for a 10% cut. The easy move is to spread it evenly across the org. Feels fair. It can also destroy the few things that are actually working, because nobody modelled which spending was producing returns and which was getting absorbed by the headwinds.
That is the difference between a scalpel and a cleaver. The scalpel only exists if you put the model around the decision first.
Most GenAI tools, by default, are obeisant. They tell you what they think you want to hear, because that’s what their training rewards. As a sole reviewer, you’re being flattered, not challenged.
But put two of them in the same room and they get vicious with each other.
“I had Claude one time say, when are you going to stop using ChatGPT for such important matters?”
So when the answer matters, run it through a second tool. Then a third. Watch them spot each other’s syntax. Watch them shred each other’s logic. The disagreements show you where the logic needs work.
That’s the GenAI version. Causal AI goes further. By design, it points out where you have gaps, where a missing variable could change the answer. Correlation can miss that variable entirely.
In physics, the relevance of data degrades over decades.
In go-to-market, the half-life is well under twelve months.
The problem is data pools older than three or four years can actively pull answers in the wrong direction. “More data is better” was a fine reflex in a stable world. We’re not in one.
“MTA was a goat rodeo, and the sudden return of MMM doesn’t fix the problem if the model still treats an open, unstable system like a closed one.”
That backdrop is also why the Magnificent Seven’s earnings keep raising eyebrows. The market is not rejecting AI. It is asking a harder question: when does the spend turn into durable returns? Most of what’s in those data lakes never gets queried. The premise was that scale solves it. The market is starting to disagree.
The other ingredient that lets bad decisions survive: time lag.
Politicians, for example, pass a bill, they get the political bump, and by the time the chickens come home to roost, everybody has forgotten whose name was on it.
“Politicians are masters at arbitraging time lag. And the dummies are all of us, because we forget.”
Corporations do the same thing inside a quarterly envelope. The decision lands in Q1. The cost shows up two or three quarters later. By then the leader who made the call may be gone, the org has reshuffled, and nobody connects the dots. The cost of bad decisions lands on the next leader, and the lesson never gets learned.
This is the same trap that keeps the MQL credit debates going. “Does Marketing get credit for this MQL or does Sales?” is the wrong decision question. GTM is a system, not a sliver of one. The right question is whether the system creates customers, and whether the model behind your forecast can survive when the system shifts.
Three things to try, especially if you’re letting AI into the decision layer:
Honeywell builds airplanes with ten-year-old chips because the flaws are known and the failure modes are bounded.
“You will never see Microsoft Windows running on a flight deck.”
Your decision process should have the same posture. Move fast where the cost of being wrong is low. Move slow where it isn’t.
PocketOS gave everyone a wake-up call.
Tools do not create better decisions.
Better decision processes create better use of tools.
Missed the session? Watch it here.
If you like this content, here are some more ways I can help:
Cheers!
This article is AC-A and published on LinkedIn. Join the conversation!
