There is one question no technical demo answers: who signs for the decision?
In engineering, the answer was structural — there was an architect, a technical lead, a review log. Every commit left a trail. Every deploy passed through approval. AI introduces a new layer: a system that deliberates. And deliberation without signature is the opposite of governance.
This note is about how to think that boundary in practice.
The language mistake
The industry adopted the term “human-in-the-loop” as if it were a solution. It is not. It is a description. It says a person exists somewhere in the execution — not where, with what authority, or under what consequence.
In a serious operation, “in the loop” means radically different things:
- Prior review — the human reads before the action occurs
- Exception approval — the human is only called when the system flags uncertainty
- Posterior audit — the human never interrupts; they reconstruct what happened
- Co-execution — human and machine decide together, with different weights
Each of these modalities carries a distinct accountability structure. Treating them all as “human-in-the-loop” imports a marketing phrase into the place where we would need protocol.
Three decision levels
Panair works with a simple criterion — it is not ours, it is a synthesis of environments that run AI under legal coverage. Every call to a system is classified into one of three levels:
I. Advisory
The AI proposes. A human decides. The system delivers evidence, options, risk evaluations — but it does not execute anything alone. This is the appropriate level for any decision that: (a) has irreversible consequence, (b) involves third parties not consulted, or (c) requires interpretation of an ambiguous rule.
Most executive applications live here. And this is where the real work is — it is not setting up the AI, it is designing the review interface so the human can decide in seconds without becoming a bottleneck.
II. Conditional
The AI executes within a pre-approved envelope. There are value limits, scope limits, impact limits. Outside the envelope, it automatically escalates to Advisory. Inside the envelope, it operates without case-by-case approval.
It works for repetitive, well-bounded operations: ticket triage, document classification, internal query response, inventory adjustments. It requires an envelope designed by someone who understands both the rule and what happens when the rule fails.
III. Autonomous
The AI decides and executes without human review along the way. Audit is posterior, not prior. Reserved for what is (a) reversible, (b) of low individual impact, and (c) repetitive enough that anomalous patterns are statistically detectable.
Most organizations err by trying to start here. Autonomous is the last level, not the first. Systems reach it after months proving behavior in Conditional.
The boundary is not technical
Note what these three levels do not say: nothing about the model, the architecture, the vendor. The boundary between them is institutional, not technical.
Who, ultimately, signs for the harm caused by a decision? This question resolves most of the doubts:
- If no one signs, no one can operate autonomously
- If signature climbs to the C-level on every exception, the conditional envelope is poorly calibrated
- If signature descends to operational levels without support, what is happening is dilution of responsibility — not distribution of it
AI governance is the engineering that makes signature possible. Logs, trails, recorded motivation for every decision, reproducible evidence. Without it, any of the three levels is theater.
What changes in 2026
Three concrete changes make this discussion more urgent than it was two years ago:
First: models with long context windows have shifted what counts as “an action.” Before, “a decision” was a prompt and a response. Today, it is a sequence of calls that can include reading internal bases, writing to systems, communication with third parties. The envelope must cover the entire sequence, not the call.
Second: regulators have started demanding an auditable description of the decision process. In Brazil, the topic entered public consultation across more than one agency. In the European Union, it is an obligation. Operating without a traceable trail will cost fines, not just clients.
Third: the model has stopped being the bottleneck. In 2024 the discussion was about accuracy. In 2026 it is about who approved what, when, and based on what evidence. The question has migrated from inference to accounting.
How we open a conversation
When a company calls us to discuss AI governance, the first question is never about technology. It is about what is already being decided by systems no one has classified. There is almost always more than expected — and almost always those systems sit in Autonomous without having passed through Advisory or Conditional.
The correction is not to stop everything. It is to reclassify. And to design the interface so each level can operate at the speed its decision category allows — Advisory fast enough not to become a bottleneck, Conditional auditable enough to withstand a stress test, Autonomous transparent enough that an exception does not become a crisis.
AI does not change who answers for the decision. It changes how much work is required for that answer to be honest.