Curadoria de IA: When Humans and Agents Decide Together

Decisions don't fail just for lack of data. They fail because they happen outside the right time. In MCI, this has an operational name: decision latency. In consultative operations, delays in Comparison are associated with drops in pT points per business day — intent has a half-life. Curadoria de IA solves this: it brings the human in at the right moment, with the right context.

SLA as a flow regulator

In MCI, SLA doesn't exist to "nag people." It exists to protect the flow. The rule is straightforward: an expired SLA shouldn't generate a human alert by default — it should generate an automatic action, preserving continuity. Notifying people is the last option, not the first reaction. Instead of "queue + pressure," you create "state + response." The cycle operates with four SLAs (first response, client response, human response, and total cycle), and each, upon expiring, triggers a system action before summoning a human.

Curadoria de IA in 3 human decisions

Curadoria de IA is not an extra task. It's an input pattern. It works because it doesn't add work — it removes the work that no one should be doing: reconstructing what has already happened. It boils down to three decisions.

1. Confirming the context. Checking if the system's reading is coherent with what is at stake. Confirming takes 30 seconds when the Bandeja de Contexto is good; it takes 15 minutes when there is no Bandeja de Contexto. The human can agree, disagree (and correct the system), or complement — and each response feeds the learning process.

2. Validating the stall. Understanding the true reason for the delay: technical risk, internal policy, price objection, hidden decision-maker. This is where the human adds the most value. IA detects that the Confidence Score has dropped; only the human notices that "I'll think about it" means "I need to convince my partner." Validating the stall combines human experience with system telemetry — and produces precision that neither would generate alone.

3. Directing the next step. Choosing the action that moves the decision with integrity — the right question, the right proof, the right meeting, or a conscious pause. Each direction is recorded and tracked: if a "technical session with IT" unstuck the cycle, the system learns that this action works for "Estudiosos" (Studious profiles) with a technical stall.

What consumes time isn't curating. It's forgetting.

Tempo de Transbordo: the KPI of operational truth

If there is one KPI that reveals the operational truth of MCI, it is this:

Tempo de Transbordo = time between the system signal and the human assumption of the decision.

This indicator doesn't measure individual effort. It measures organizational capacity and system design quality: routing, specialist availability, Curadoria de IA maturity, and clarity on what should (or shouldn't) go to a human. When it's high, it points to an undersized team, concentration of decisions in few specialists, loose escalation criteria, or a low-quality Bandeja de Contexto. It answers a question CEOs often ask with little evidence: "Do I need to hire more people or organize the system better?". Measured as median and P90, it distinguishes lack of capacity from unnecessary over-escalation.

Curadoria de IA as continuous learning

Curadoria de IA is learning in production. Every act validates or corrects the IA's reading, refines stall hypotheses, and improves the next Bandeja de Contexto. The effect is compound: the more cycles are curated, the better the system becomes; the better it becomes, the less Curadoria de IA is needed — because the IA starts correctly solving a larger portion of situations, and the human enters only where judgment is irreplaceable.

This effect is what makes MCI economically sustainable at scale: the cost per decision falls over time, it doesn't rise. IA handles the repeatable. Humans handle the responsible. Curadoria de IA is the quality pact between the two.