Derisking AI by design: How to build risk management into AI development