Design Principles for AI That Actually Works

Good design is useful. Good design is honest. Good design makes a product understandable. Good design is concerned with the long term. Good design is as little design as possible.

Now strip out the word design and replace it with AI.

Most of what we read about AI fails most of those tests. It is about more, not less. It is about commercial efficiency, not human good. It is about now, not the long term. It is about scale of output, not honesty of output.

Dieter Rams wrote his principles in the seventies. He was talking about Braun radios and shelving systems. Sixty years on, his principles are still the closest thing the design world has to a constitution.

Reading a recent Gates Foundation grant call for AI tools that strengthen charitable giving, something shifted. We had read a lot about AI. We had read almost nothing about what it is actually doing for cancer patients, refugees, or people watching their coastline disappear. Almost nothing about the work being built by people who got into this to make something better, not to make money.

So we went looking. Three projects, in particular, stand up to Rams' tests.

NHS EDITH: AI as a workforce response

In February 2025, the NHS launched the world's largest AI mammography trial. It is called EDITH and it is running across thirty hospitals with around 700,000 women. Britain is short of radiologists: a 30% shortfall reported in 2023, projected to hit 40% by 2028. Every screening mammogram needs two specialists. The maths does not work. EDITH is testing whether AI, working alongside one human reader, can match the accuracy of two. Early evidence from a Swedish trial published in The Lancet suggests it can.

This is AI as an answer to a workforce shortage, not a replacement for expertise. The clinical reader stays in the loop. The AI absorbs the volume that would otherwise crush the system. Useful. Honest. Concerned with the long term.

UNHCR Jetson and DRC Foresight: AI built around what it should not do

A long way from a UK hospital, the same kind of work has been going on for years. In 2011, the Dollo Ado region of southeastern Ethiopia was receiving around 2,000 Somali refugees a day. The team on the ground had been resourced for nothing like it. UNHCR built a project called Jetson out of that experience. Ten variables, including rainfall, commodity prices, conflict incidents, and historical movement, that in combination predict where displacement is likely to happen and when.

A separate tool called Foresight, built by the Danish Refugee Council, can model displacement up to three years out. The Foresight team deliberately built it so it cannot predict exactly where refugees are going.

UNHCR has chosen to keep most of its work internal. Both teams were worried about the same thing: that information shared too widely could be weaponised by politicians or border agencies against the people the tools were built to protect.

Build the thing. Decide what it will not do. Hold the line.

That is restraint as a design principle. Rams would have approved.

What these projects have in common

Three projects. Three problems. All of them built around the same shape:

It is also, almost word for word, how we work with our clients.

We call it the Contour Methodology. Map the terrain first. Narrow in on what is specific. Then, only then, ask the actual question, with the evidence chain visible and the boundaries between fact and inference clear.

Most people skip the first two steps. Someone posts "you are an expert in X, now solve Y" and presses send, which is the most efficient way ever invented to produce confident-sounding nonsense. Skill downloaded. No understanding.

A small tool for thinking the right way around

We built something small to demonstrate this. A free tool, Briefing in Contours, that does one thing. It walks you through three short stages, the way we would walk through a client engagement, and outputs three structured prompts you can paste into any AI tool in sequence.

Stage one tells the AI to understand the environment. Do not answer yet. Do not write a report. Confirm you have looked at the right places.

Stage two narrows in. Given what you found, what are the three or four real considerations? What are we explicitly not solving?

Stage three is the actual question, asked with all the context the AI has just built for itself.

Three exchanges instead of one sprawling mess. Less token waste. Less hallucination. More of the answer you were actually looking for.

Less, but better.

It is not a magic prompt. It will not replace proper context engineering. But in the space of two or three uses, it shows you why context matters, which is the point.

The bigger pattern

The bigger thing the Gates call surfaced is that the good stuff is already out there. The radiologists' shortage being met. Displacement being anticipated. The careful, deliberately limited tools built by people who would lose sleep if they got it wrong.

They just do not shout as loud as some.

Worth knowing about, next time someone tells you the technology is only good for one thing.

Try Briefing in Contours, or score your current AI approach against the patterns associated with success and failure.