Where internal sustainability risk expertise is weak, banks often become over-reliant on the apparent “knowledge” of the model. Large language models are then asked to interpret broad sustainability information, decide what matters and produce risk conclusions. Agentic workflows can amplify this by letting one model-generated step trigger another, building chains of reasoning that appear sophisticated but may rest on unstable foundations.

An agentic workflow is risky for several reasons.

First, sustainability risk is a domain with less standardised LLM training signal than traditional financial analysis. The boundaries between relevant and irrelevant information are more ambiguous, and terminology is often inconsistent across sectors and sources.

Second, general-purpose LLMs are not built to reflect a bank’s internal materiality logic, sector approach or risk governance expectations unless those are imposed externally.

Third, output variability matters more than many innovation teams assume. In marketing, variation may be acceptable. In regulated credit environments, it is a governance issue.

The alternative to unconstrained AI is deterministic design.

A deterministic sustainability risk workflow does not ask the model to “understand everything”. Instead, it breaks the problem into smaller controlled steps. Sector and activity mapping can be rule-based. Materiality logic can be framework-driven. Evidence collection can be structured. Risk factors can be assessed against predefined criteria. AI can then be used only in carefully bounded areas where interpretation is truly required.

This design changes the role of AI from decision-maker to constrained analytical component.

That has major advantages. It improves reproducibility. It reduces hallucination risk. It makes traceability easier. It allows simpler and cheaper models to perform well. And it anchors the output in the bank’s own methodology rather than in opaque model behaviour.

In practice, this often means that the strongest system is not the one with the most “agentic intelligence”, but the one with the best architecture. In other words, the quality of the framework matters more than the apparent sophistication of the model.

This is especially important when institutions lack internal sustainability risk expertise. In that situation, a bank should be extremely cautious about delegating judgement to generative systems. Where the bank itself cannot define the answer quality standard, it cannot safely rely on the model’s implied judgement either.

The right objective is not AI theatre. It is defensible automation.

In sustainability risk assessment, that means building a system where the same input leads to highly consistent outputs, where the role of AI is limited and explainable, and where the logic can withstand internal challenge from risk, compliance, model governance and audit.

In short, the strongest automation is not the most creative. It is the most controlled.

Other articles in the series: