When a system does something unfair — quotes one customer double what it quotes their neighbour, quietly declines the people it was never trained to recognise, buries the one disclosure that actually mattered — we go looking for the person who was unethical. We rarely find one. What we find instead is a decision, made months earlier at a whiteboard, by someone who wasn’t thinking about ethics at all. They were thinking about latency, or cost, or getting the thing shipped.
That’s the uncomfortable idea this post is about: most of the ethical outcomes a system produces were designed in, not acted out. Ethics isn’t only a property of people. It’s a property of the structure those people built — and the structure is the architect’s job.
Three ways to be right
Applied ethics has three classic lenses, and it’s worth two sentences each because they map surprisingly cleanly onto how we build.
- Consequentialism — outcomes are all that matter; act to produce the best results. The catch is that someone has to define “best”, and for whom.
- Deontology — some things are simply your duty regardless of the outcome. This is where professional codes and regulation live: rules you follow because they’re the rules.
- Virtue ethics — less about rules or outcomes, more about character. What a good practitioner does when nobody is checking.
Most ethics conversations in tech stop at the first two, because outcomes and rules are the ones you can audit. The third — the discretionary, nobody’s-watching one — is the one architecture quietly runs on, and it’s the one I want to come back to.
The architecture is the organisation’s operating logic
Ross, Weill and Robertson made the point years ago that enterprise architecture defines the logic by which an organisation operates and evolves. Read that back slowly. It means the values an organisation actually lives by are not the ones on the poster in reception — they’re the ones encoded in its systems, its data flows, and the defaults nobody questions.
Epley and Kumar reach the same conclusion from the other direction: you get an ethical organisation by designing for it, not by exhorting people to be good. A system that makes the fair path the easy path produces fair outcomes from perfectly ordinary people. A system that makes fairness expensive produces unfair outcomes from good ones. The behaviour follows the structure far more reliably than it follows anyone’s intentions.
Which is why “we take ethics very seriously” is not an architecture. It’s a hope.
Nowhere is this sharper than in automated decisions
A model that prices, ranks, scores or approves is making thousands of ethical decisions an hour, and it’s making every one of them exactly the way you built it to. Feed it training data that carries a historical bias and it will launder that bias into a number that looks reassuringly objective — a score, a premium, a yes/no — with all the human judgement stripped out and none of the human accountability.
In a regulated industry this is no longer hypothetical. Regulators are actively studying bias in automated decision-making: in the UK the FCA has been weighing whether algorithmic fairness needs new rules, while the EU’s AI Act has already written obligations for high-risk systems into law. And “the algorithm did it” has never once been an acceptable answer to a regulator, a court, or a customer. The choice of what the model optimises for, what error rate you tolerate, and — the quiet one — on whom you tolerate it, is an architectural decision through and through. It just doesn’t look like one, because it arrives dressed as a metric.
Where the architect actually stands
Regulation gives you a floor: the deontological duty of the rules you must follow, and the consequentialist stick of bad outcomes getting punished. That floor matters. But it is the floor. Between “not illegal” and “actually fair” is a wide gap, and it’s made of discretionary choices no rule reaches:
- whether the fair option is the default, or the thing someone has to fight for;
- whether the trade-off you made got written down, or quietly forgotten;
- whether you raised the concern about the data, or noticed you could and moved on.
That gap is where virtue ethics lives, and it’s the part only the architect can supply — because the architect is the one holding the pen while those choices are still cheap. After the boxes are drawn, fairness becomes a retrofit. Before, it’s a line in a diagram.
It’s the same question this blog keeps asking
At bottom this is the question the whole blog circles back to: what are you optimising for. Fairness, privacy and accountability are non-functional properties like any other — you either design for them, or you find out later that you didn’t. Security people already treat their concern as an -ility to be built in rather than bolted on; ethics deserves exactly the same seat at the table, for exactly the same reason.
And like any architectural decision worth the name, an ethical trade-off deserves to be written down. An architecture decision record isn’t only for “we chose Postgres over Mongo”. “We accepted this residual unfairness because the alternative was X, and here is who signed it off” is precisely the kind of reasoning the next person needs to inherit — and precisely the kind that evaporates if you don’t capture it. A documented, owned, reviewable ethical decision is a world apart from one that simply happened.
It shouldn’t be a surprise that structure carries values. We already know structure leaks: an organisation’s shape ends up stamped on its systems whether anyone intended it or not. Its values leak the same way. The architecture quietly encodes whose interests it serves, long after everyone has forgotten deciding.
The narrower, harder point
None of this makes the architect the ethics police, and it doesn’t pretend every fairness question has a clean answer — plenty don’t. The point is narrower and harder: you cannot delegate this to a code of conduct or a compliance checkbox, because by the time it reaches them the decision is already set in the diagram.
The reassuring part is that a lot of these calls — how much risk to accept, on whom, and whether it’s worth the cost to remove — turn out to be more measurable than they look. There’s a discipline for putting numbers on exactly that, but it’s a thread for another day.
For now, just this: the next time you draw the boxes, remember you’re deciding more than how the system works. You’re deciding how it treats people. And that choice is a great deal cheaper to make on the whiteboard than to apologise for later.