What AI can decide for you — and what it can’t
In the piece on session autonomy we settled a simple thing: given a task, AI lays out its own path to the goal — it breaks the task into intermediate goals, decides the order, and recognises on its own when it’s finished. If it didn’t do that, every job would be an endless loop until a human said “stop.” What AI does not do is set itself a goal out of nothing — it doesn’t choose what to occupy itself with in the first place, doesn’t set its own agenda from scratch. The direction and the limits come from a human; inside them, AI is self-reliant.
We also split actions there into three classes — reversible (does them itself), irreversible (only with human consent), and forbidden (never). That was the question “what is it allowed to do.”
This text goes one floor deeper and asks something harder: what AI can decide for you — not “carry out,” but “settle.” Because the more companies let AI into real work, the more often the tempting line comes back: “our system will make the decisions for you.” Worth breaking that down before we agree to it.
As always, one caveat up front: we don’t write this as experts who know how you should set up your company. We show how we ourselves see it from our side — from everyday work with multi-agent systems. What to do with it stays your decision.
A decision is not just the right choice
The first reflex: if the model can predict a good answer, it can decide. And that’s where the mistake sits. A decision is not just the outcome — it’s also whose it is. The model will compute what would be a good choice, often accurately. But it won’t be the one who answers for that choice — who carries the weight when it goes wrong, who stands before the client or the regulator. A decision with no one behind it isn’t a decision — it’s a prediction.
In a company that difference is the difference between a tool and a risk: a model that “decided” and has no way to answer for the result quietly shifts the responsibility onto whoever let it in.
Two layers of every decision
To see what can really be handed to a machine, it helps to split a decision into two layers.
The rule layer — is it legal, in line with good principles, does it harm no one. This layer the model, on the basis of the data it was built on, recognises fairly well; it can point out “this isn’t allowed.” That’s the part you can largely automate.
The character layer — what this particular company would decide, with its priorities, its risk appetite, its style. This layer can also be partly read off — from history, from what the company has so far accepted and what it rejected. On that basis a system can reproduce your pattern.
And now the practical crux: where both layers agree, and the matter is routine and fits the pattern — it can safely be handed to the system. Classifying a typical ticket, a first-pass reply along a known template, splitting tasks the way it’s always done. Where the layers diverge — or the matter is new — the decision goes back to a human.
Why the human stays right at the edge
Here’s a point that sounds abstract but is very concrete.
A company’s history can only answer the question “what did we decide before.” And the whole reason a decision is needed at some fork is that it isn’t there yet. If the matter fit the pattern completely — you wouldn’t be deciding, it would be a lookup from a table, automatic. Every real decision is by nature a little beyond what the system already knows.
That’s why the human doesn’t sit “everywhere” — they sit exactly where predictability ends. And it isn’t a matter of gathering more data. The more data, the better the system guesses the routine — but the fork, by definition, escapes the pattern. It’s structural, not temporary.
It flips the usual question. Not “can AI decide” (on the routine — yes), but: which decisions are in-pattern and cheap (let it), and which are new, expensive or irreversible (the human takes them on). And most importantly — the real job of a good system isn’t to “decide everything,” but to know when it’s out of its depth and pass the matter up. The value is in recognising the edge, not in the choice itself.
What it means in practice
You don’t have to make philosophy of it. It’s enough to set three things:
- Split decisions, not just tasks. What’s routine and reversible — the system does itself. What’s irreversible or high-stakes (money, a client, data, a live system) — needs a human. That’s a boundary, not a slogan.
- Build an escalation threshold. The system should be able to say “this goes beyond what I know” and pass the matter on — instead of manufacturing a ruling in a confident tone. A quiet, confident shot at a new matter is more dangerous than an honest “I don’t know.”
- Keep the last word where the stakes are real. Not because the system can’t be trusted — because when something is irreversible, someone has to be able to answer for it.
In the end
“Fully autonomous AI that will decide for you” sounds like saved time. In truth it sells the one thing not worth handing over: the fact that behind a decision stands someone who can be held to it.
A mature system isn’t the one that decides everything. It’s the one that calmly takes the routine — and hands you the edge. Let the routine flow on its own. The fork stays yours, because only yours it can be.
MafiaAI — a team of people and AI agents building tools, websites and solutions, including AI deployments run locally and privately, so your data stays with you. More: t8.pl