Ethical AI, which relates to model behaviour, and ethical automation, which relates to process design, are separate problems with some logical overlap. These principles, when applied together, form a practice that can address the worries of industries like financial crime compliance that are being innovated by AI: what happens to the people whose jobs the automation touches? 

In FCC, an AI decisioning capability needs to be fair and explainable, sitting inside a screening or monitoring workflow with clear ownership, real human oversight, and a defensible answer to who is accountable when an alert is closed incorrectly. Ethical AI, done properly, is what makes ethical automation achievable.

Two real efforts show what taking this seriously looks like: Singapore's MAS-convened Veritas Initiative, which turned fairness and accountability into assessable methodologies across two dozen banks, and the Financial Stability Board's (FSB’s) Sound Practices for Responsible AI, a global baseline for governance and oversight.

What's the difference between ethical AI and ethical automation?

The FSB's ‘Sound Practices’ assign AI accountability explicitly to the board and senior management, built on the three-lines-of-defence model most compliance functions already run. Veritas takes a parallel approach for its own use cases, pairing dedicated accountability and ethics assessment tools with its fairness metrics, so responsibility for an outcome is assigned rather than assumed. Both are answering the same underlying question: who owns an automated decision, and who answers when it's wrong?

Ethical AI is about the model: is it tested for disparate outcomes, can its output be explained, is its behaviour monitored for drift? Ethical automation is about the workflow the model sits inside: who owns the process, who can override it, what happens when it fails, and does it create excessive job displacement? 

Automation research calls the space between them a “responsibility gap”; nobody who builds or deploys a system makes any single decision it produces, yet someone still has to answer for it. The FSB and Veritas both close that gap the same way, by assigning it to a named person before deployment, rather than discovering it afterwards.

In FCC the gap can have real consequences. A missed sanctions hit or a wrongly closed adverse media alert is a regulatory exposure with a named MLRO or L2/L3 reviewer whose sign-off is on record. Workflows have to make that ownership explicit before deployment, not during an examination.

What does automation need before it can be trusted?

Ethical AI is the foundation ethical automation depends on: a workflow can only be trusted to run with less human touch once the model underneath it is fair, explainable, and matched to the right level of oversight. 

Three principles, each with a working precedent at Veritas or the FSB, show what that foundation looks like in practice:

• Veritas's FEAT framework tested fairness by segment for credit scoring, using metrics built by a consortium of over twenty banks, including DBS, HSBC, and Standard Chartered. If a decisioning process is automated without this discipline a model can score well overall while quietly performing worse for particular name structures or geographies, scaling the bias.

• Veritas built transparency tools for risk decisioning; the FSB pushes institutions toward explainable AI, or compensating controls, wherever a decision's basis is hard to follow. Automation stays accountable only if the human closing it has a defensible reason for the reviewer.

• The FSB's proportionality principle is, in practice, risk tiering – weighing materiality and risk to set oversight levels. Automating every alert type the same way ignores that a sanctions name match, an adverse media hit, and a transaction monitoring alert each demand a different evidentiary bar for closing autonomously, so that tiering has to run across an entire AI programme, not just within a single workflow.

How does ethical AI prevent automation from becoming job displacement?

One case study in the FSB's own Sound Practices document is instructive: an insurer eliminating roughly 400,000 manual processing instances through AI, framed purely as an efficiency gain, with no mention of what happened to the people who did that work.

Neither the FSB nor Veritas’s FEAT framework addresses workforce impact at all, the FSB says so explicitly, which is why job preservation has to be designed for separately, rather than assumed as a byproduct of getting fairness and accountability right. To this end, the measure that matters is what happens to the capacity automation frees up. It should go toward harder, judgement-heavy alerts; the escalations and typologies that need more in-depth investigative reasoning, rather than letting headcount shrink because volume dropped.

A programme that reports an efficiency gain without also reporting where that freed capacity went is applying the same incomplete measure of success the FSB's own case study shows can go unquestioned.

How does Iris 7 put governed, policy-bound AI into practice?

Iris 7, our investigation and decisioning platform, applies the same idea: a principle only counts once it's assessable, with a named owner and a documented control. Its decisioning Agents operate through policy-bound reasoning, executing decisions within the institution’s defined risk appetite and policy limits. 

The platform is designed for human-in-the-loop oversight, where analysts and compliance teams set what the AI is authorised to decide, and human oversight continues in production, where a dedicated governance and quality assurance layer reviews both AI-made and human-made decisions across the full lifecycle, addressing structurally what automation research warns against: a reviewer positioned after a decision is made, with little real power to change it.

Every decision carries a natural language explanation and a structured evidence trail, the same explainability-for-the-reviewer principle Veritas and the FSB both push toward, hardened by model governance, validation, and audit review cycles in global Tier 1 institutions since 2018. 

Key takeaways

• Veritas and the FSB show fairness, accountability, and oversight already operating as assessable industry practice.

• Ethical AI (fair, explainable, monitored models) is the foundation ethical automation depends on in compliance.

• Explainability must be calibrated to the accountable reviewer, not just the model team.

• Job preservation means directing the capacity automation frees up toward harder, judgement-heavy alerts, not treating headcount reduction as the measure of success, an area neither framework yet addresses.

• Iris 7 shows what closing that gap can look like in practice: policy-bound decisioning, human oversight at both policy design and production, and audit-ready explanations built into every decision.

• The gap between leadership enthusiasm and frontline caution is a signal to act on, not a resistance problem.