Overview

Arjun Wijaya spent twelve years building risk models for a regional bank in Kuala Lumpur before landing an AI governance role at a global insurance group. His first week on the job, a product team showed him a claims-routing model they had already deployed. "We benchmarked it against our old process," they told him. "Accuracy is up fourteen percent." Arjun asked one question: "Fourteen percent accurate at predicting what?" The room went quiet. Nobody had written down what the model was actually optimizing for - and nobody had asked whether that objective matched what the company believed was fair to policyholders.

Why Values Come Before Frameworks

Most organizations jump straight to picking an ethics framework - a checklist from a standards body, a set of principles from a government white paper. That instinct is understandable. Frameworks feel concrete. But a framework borrowed from somewhere else only works if it aligns with what your organization actually believes.

Think of it like a building code. Building codes exist because societies have already agreed that buildings should not collapse on their occupants. The code translates that shared value into specific rules about load-bearing walls and fire exits. If you have not agreed on the underlying value - safety matters - no building code will save you.

The same logic applies to AI ethics. Before you adopt any framework, you need to articulate your organization's actual values: what you believe about fairness, about whose interests you serve, about acceptable risk, about transparency with customers. Only then can you evaluate which frameworks help you act on those values.

The Four Frameworks You Will Encounter

Four ethical frameworks dominate the AI governance conversation. You do not have to pick just one. Most mature organizations blend elements of all four, depending on the decision at hand.

Consequentialism

The consequentialist view asks: what outcome does this produce? An action is ethical if it creates more benefit than harm, measured across everyone affected. In AI terms, this means running impact assessments before deployment. You estimate benefits to users, risks of harm to affected populations, and net effect. The challenge is measurement. Whose benefit counts, and how do you weigh a five percent improvement in speed against a two percent increase in erroneous decisions affecting a protected group?

Deontology

The deontological view says certain actions are right or wrong regardless of outcome. You do not deceive users. You do not discriminate on protected characteristics. You give people the right to contest automated decisions. These are rules, not trade-offs. Many AI regulations - including the European Union's AI Act - are heavily deontological: they list prohibited uses and required rights, regardless of whether a particular deployment would produce net-positive outcomes.

Virtue Ethics

Virtue ethics asks: what would a trustworthy, responsible organization do here? It focuses less on rules or outcomes and more on character. Would we be comfortable if this decision were reported on the front page of a newspaper? Would we explain it openly to the customers it affects? This framework is particularly useful for novel situations where rules have not yet been written. It asks teams to reason from character rather than looking for a loophole.

Care Ethics

Care ethics centers relationships and context. It asks: who is most vulnerable in this situation, and what do we owe them? This framework pushes back against one-size-fits-all policies. A customer who is digitally literate and financially stable needs different protections than a first-generation smartphone user with limited English. Care ethics insists you design AI systems with your most vulnerable users in mind, not just your average user.

Values Alignment in Practice

Values alignment is not a one-time ethics workshop. It is an ongoing organizational practice. Arjun's insurance group learned this the hard way. They ran a half-day values session and produced a two-page principles document. Eighteen months later, a product team launched a pricing model that technically violated two of those principles. Nobody had read the document. Nobody had built the principles into the product development process.

The organizations that get this right treat values alignment as an operating system, not a document. Here is what that looks like in practice.

Translate values into testable criteria

"We believe in fairness" is not testable. "Approval rates for identical risk profiles should not differ by more than five percentage points across demographic groups" is testable. Before any AI system goes to production, your team should be able to name the specific metrics that would tell you the system is acting in accordance with your stated values.

Build values checkpoints into the development lifecycle

Add a values review at three specific gates: when a use case is first proposed, when a model is ready for user testing, and when a system moves from pilot to production. Each review asks the same core questions: Who is affected? How could this harm someone? What values are we trading off against each other here, and is that trade-off justified?

Create a dissent channel

People inside your organization will notice when a system feels wrong before the metrics catch it. Arjun's team created a simple internal form - think of it as an ethics incident report - where any employee could flag a concern about an AI system. Concerns are reviewed monthly by a cross-functional group. In the first year, eleven concerns were submitted. Four led to meaningful changes. Two were escalated to leadership. The remaining five were reviewed and closed with documented reasoning. The act of reviewing and closing matters as much as acting on the serious ones.

The Hardest Problem: Values in Conflict

The real skill in ethical frameworks is not applying one value - it is navigating the moments when two legitimate values pull in opposite directions.

Consider privacy versus safety. A mental health app wants to use AI to flag users who may be at risk of self-harm. Flagging those users may save lives. But it also requires processing deeply sensitive personal data in ways users may not expect. Both safety and privacy are real values. The framework cannot tell you which wins in every case. What it can do is force you to be explicit about the trade-off, document your reasoning, and build in review mechanisms so you revisit the decision as your understanding improves.

>
Ethics is not the absence of tension. It is the practice of making your tensions visible and your reasoning defensible.

The organizations that struggle most with AI ethics are those that pretend the tensions do not exist - that they can be "fair and efficient" or "transparent and profitable" without ever hitting a moment where those values point in different directions. They will hit that moment. The question is whether they have built the organizational muscle to work through it deliberately.

Building the Muscle: Practical Steps

If you are starting from a weak position - a principles document nobody reads, no review process, no dissent channel - here is a three-step sequence that works in ninety days.

Month one: Values inventory. Convene a small group across legal, product, operations, and customer-facing teams. Ask: "What do we actually believe about how AI should treat our customers?" Document the answers, including the disagreements. Disagreements are data. They tell you where your organization's values are genuinely unsettled.

Month two: Framework mapping. Map your stated values to the four frameworks above. Ask: "Which framework does this value primarily reflect?" This exercise reveals gaps. Organizations often find they have strong consequentialist instincts but weak deontological commitments - they measure outcomes well but have not drawn any firm lines about what they will never do.

Month three: Process integration. Pick one current AI project and run it through a values review using the criteria you just developed. Do this with the actual product team, not as an audit imposed from outside. The goal is to demonstrate that values alignment makes better products, not to catch teams doing something wrong.

Key Takeaways

• Values before frameworks. Borrowing an ethics framework only works if it maps to what your organization actually believes. Articulate your values first.
  - Four frameworks, one toolkit. Consequentialism, deontology, virtue ethics, and care ethics each capture something real. Use them together, depending on the situation.
  - Testable criteria are essential. Translate abstract values like "fairness" into specific, measurable thresholds before any system is deployed.
  - Build values into the process, not a document. Ethics checkpoints at proposal, testing, and production stages prevent the principles-on-paper problem.
  - Create a dissent channel. Employees notice problems early. Give them a safe, structured way to surface concerns before issues become incidents.
  - Values in conflict are normal. The goal is not to eliminate tension but to make trade-offs visible, deliberate, and documented.
  - Start with one project. A ninety-day inventory, mapping, and integration sequence turns abstract commitment into practical organizational habit.