NEWSROOM

AI systems are beginning to act: approving transactions, adjusting limits, initiating payments, executing trades, negotiating with other systems on behalf of people and companies. When an AI system acts, its output is no longer information. It is an event in the world, with financial, legal and reputational consequences.

A blockchain can prove that a transaction happened and that it was signed by a valid key. It can't prove that the transaction should have happened. That is a decision, and decisions need their own infrastructure.

For most institutions, that proposal is wrong, and not only because migration is expensive and risky. It is wrong because it misidentifies the problem. In most institutions, the main obstacle to better decisions is not the quality of their models. It is the absence of infrastructure between those models and the decisions they inform.

Agents are starting to act. They pay suppliers, schedule transfers, move balances between accounts, initiate payments through Pix and Open Finance...

It happens quietly, every day, in financial institutions. A field arrives empty and the system fills it with a default. A bureau query times out and the pipeline proceeds without it. A financial statement is two years old, but nothing checks its date. The model receives something that looks like a complete input, produces a confident score, and a decision is made.

The first wave of thinking about AI agents assumed a simple structure: a human delegates a task to an agent, and the agent acts. One principal, one agent, one line of authority.

Ask a financial institution how its credit model performs and you will get a precise answer: an AUC, a Gini coefficient, a calibration curve. Ask what that model has contributed to the institution's results and the answer becomes vague.

That gap is the reason so many institutions struggle to explain, audit and measure their AI decisions. The fix is conceptually simple and architecturally demanding: every consequential decision needs a single record, under a single identifier, that connects everything the decision depended on and everything that followed from it. We call it the decision record.

Every AI system produces logs. Almost none produces proof.

Every AI decision system in a financial institution has a human-in-the-loop somewhere. An analyst reviews borderline credit requests. An account manager decides whether to follow a recommendation. A credit committee approves exceptions. A supervisor reverses a decision after a complaint.

Yet in many institutions, that policy doesn't fully belong to them. Part of it is configured inside a vendor's system by the vendor's consultants. Part is hard-coded in software no one in the credit team can read. Part lives in spreadsheets and documents that don't match what the system actually does.

For financial institutions using AI in credit, this right raises a practical question that goes beyond compliance checklists: when a customer asks why, what exactly does the institution need to be able to explain?

The usual approach to explainable AI in financial services follows a predictable sequence. A model is built for performance. It is deployed. Then, when compliance, customers or regulators ask how it works, an explainability tool is attached to it: feature importances, SHAP values, counterfactual examples.

A cooperative system doesn't need to choose between central control and local autonomy. It needs to make the boundary between them explicit.

In most credit operations, the central object is a number. A score of 0.73. A probability of default of 2.4%. A risk rating of B+. Systems compute it, dashboards display it, analysts discuss it. But a number is not a decision. At some point, someone or something has to turn it into an action: approve or not, how much, under what terms, now or later. In most institutions, that translation is the least structured part of the process. It lives in cutoffs, informal ranges and an

That model is under pressure. Regulators and supervisors increasingly ask not whether the framework exists, but whether a specific decision, about a specific customer, on a specific date, complied with the rules in force at that moment. Answering that question for one decision is laborious. Answering it for thousands is often impossible.

Machine learning is probabilistic by nature. A credit model doesn't say a customer will default. It says there is a 3% probability. Language models go further: ask the same question twice and you may get two different answers.

An obsession with accuracy can be counterproductive. Models overfitted to the past create an illusion of control. In practice, that illusion leads to hidden risk exposure and to decisions that look technical but are dangerously dogmatic.

AI decision infrastructure is the layer between a model's output and an institution's action. It turns intelligence into decisions that can be explained, governed, audited and measured. It has six components...

Decision intelligence was right about one thing: decisions are what matter. But insight alone doesn't simplify decisions. Structure does.