Foundation First. AI Second.

AI has created a strange temptation. When a system is slow, fragmented, expensive or frustrating, we increasingly ask what artificial intelligence could do to it before asking whether the system itself has been properly designed.

AI can accelerate a sound system. It can also accelerate confusion. A model cannot clarify a purpose an organisation has never defined. It cannot repair ownership that no one has assigned, create trustworthy data from records nobody has stewarded, or make an incoherent workflow coherent simply because it now moves faster. Technology enters a structure that already exists, and it inherits much of what that structure gets right—and what it gets wrong.

AI inherits the system beneath it

It is easy to treat AI as a layer that sits above the organisation: choose a model, connect some data, automate a task, measure the savings. But the visible technology is only one part of the system. Beneath it are decisions about purpose, data, process, authority, incentives, accountability and human judgement.

If those foundations are weak, automation can make the weakness harder to see. A confusing process may become a faster confusing process. Poor data may be reproduced at scale. Unclear responsibility may become even less visible when people begin saying that “the system” produced the result. A workflow that should have been redesigned may simply acquire an AI step and become more complicated.

This is why the first question should not be, What can AI do here? It should be, What are we actually trying to achieve, and is the system beneath the technology worthy of being scaled?

The foundation is not merely technical

Good AI foundations are often described in technical language—data quality, security, model performance, architecture and integration. Those things matter, but they are not enough. The deeper foundation includes governance, provenance, ownership, measures of success, human oversight and the trustworthiness of the process that produces the data in the first place.

NIST’s AI Risk Management Framework treats trustworthy AI as a socio-technical problem rather than a model-quality problem alone. Its core places governance across the wider work of mapping, measuring and managing risk throughout the AI lifecycle. The OECD’s updated AI Principles likewise emphasise accountability, traceability and systematic risk management across the lifecycle of an AI system. In Australia, the national framework for AI assurance in government explicitly focuses on foundations across government rather than treating assurance as a purely technical exercise.

The pattern is consistent: reliable AI does not begin with the model. It begins with the system of responsibility around the model.

Automate administration. Augment intelligence. Preserve judgement.

This has become one of the clearest Renovaré principles for the use of AI. Automation is valuable when it removes repetitive administration, reduces avoidable friction, helps people retrieve information, organises material or performs work that does not require moral or professional judgement at every step.

AI is also powerful as an augmentation tool. It can help us see patterns, compare options, summarise complexity, test assumptions and work through information at a scale that would otherwise consume enormous amounts of time. But augmentation is different from abdication. Where decisions materially affect people, rights, resources, safety, opportunity or institutional trust, someone still needs to understand the decision, own it and remain answerable for it.

Automate administration. Augment intelligence. Preserve judgement.

Purpose before capability

The Purpose question changes an AI project immediately. Instead of beginning with the available tool, begin with the outcome. What problem are we solving? Who is being served? Which burden are we trying to remove? What would success look like if AI did not exist? And is the use of AI actually necessary, or have we been attracted to the technology because it is available?

This matters because new capability creates its own pressure. Once an organisation knows that something can be automated, the existence of the capability can begin to feel like a reason to automate it. But can is not the same as should. A technology decision is still a stewardship decision about time, money, data, attention, authority and the people who will live with the result.

Structure before scale

AI also forces us to look at structure. Where does the information come from? Who owns it? How is quality checked? Which rules govern access? What happens when systems disagree? Who can override an automated output? How is an error detected, corrected and learned from? If the answers are vague before deployment, scale will not make them clearer.

The attraction of AI is often speed, but speed magnifies whatever has already been designed. A sound structure can use that speed productively. A weak structure can use it to distribute mistakes, confusion and hidden assumptions more efficiently. Foundation First is therefore not a demand for perfection. It is a demand for enough order that we know what we are scaling.

Stewardship before speed

The Stewardship question asks what has been placed in our hands. In an AI system that may include personal information, institutional knowledge, intellectual property, employee time, public money, customer trust, professional authority and decisions that affect people who may never know how the system works.

Good governance is sometimes presented as the thing that slows innovation down. I think that misunderstands governance. Properly designed, governance is the structure that makes responsible scale possible. It tells us what must remain visible, who must remain accountable, which risks require human review and what should happen when the technology behaves differently from what we intended.

Trust is not created by declaring that an AI system is trustworthy. It is earned through the quality of the foundations around it: truthful data, clear ownership, competent oversight, meaningful measures, traceability and the willingness to stop or redesign a system when it no longer serves its purpose.

The long view

Every automation teaches an organisation something about how work should be done. Over time, those choices become habits, dependencies and infrastructure. That is where the Legacy question enters. If we operate this system for five or ten years, what will accumulate? Better knowledge and clearer processes? Or opaque dependencies that no one fully understands and few people know how to challenge?

The legacy of AI will not be produced by models alone. It will be produced by the systems into which we place them: the rules we write, the data we preserve, the responsibilities we assign, the judgement we retain and the practices we allow to become normal.

Foundation first is not an argument against AI. It is an argument for building AI on something worth amplifying.

Sources & further reading

National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), and the Generative Artificial Intelligence Profile.

OECD, AI Principles, updated in 2024—particularly the principles on robustness, accountability, traceability and lifecycle risk management.

Australian Government Department of Finance, National framework for the assurance of artificial intelligence in government.

THE STEWARDSHIP QUESTION

Where are we using technology to compensate for a system we have never properly designed?