When organizations talk about AI governance, the conversation often turns quickly to policies, risk frameworks, approved tools and compliance. But one of the most important lessons from the ABC’s use of AI in regional news is much more practical:
Good AI governance happens inside the workflow.
Laura Gartry, Newsroom Innovation Lead for ABC News, recently shared how the ABC approached its local news briefings project — one of ABC News’ first audience-facing uses of generative AI.
What makes the example compelling is not the sophistication of the technology. It is the operating model built around it.
Start with the problem, not the AI
Regional ABC newsrooms produce important local journalism every day: council updates, community announcements, local sport and stories that matter specifically to their communities. But much of that journalism historically existed primarily as radio news bulletins.
The challenge was clear: how could the ABC make that journalism more accessible through its website and app without creating an unsustainable additional workload for already-small regional newsrooms?
The ABC already had ABC Assist, its internally developed AI tool trained using its archive and style guide, but rather than searching for somewhere to deploy AI, the organization started with an audience problem.
That distinction matters. Too many AI initiatives begin with:
“We have AI. What can we automate?”
A stronger governance question is:
“What outcome are we trying to achieve, and what role should people, AI and systems each play in achieving it?”
In the ABC’s case, AI assists with reformatting existing journalist-written radio scripts into a digital format.
It does not create the underlying journalism. That boundary is deliberate.
Human-in-the-loop is more than putting an approval button at the end
The phrase human-in-the-loop is now common in discussions about responsible AI, however simply requiring someone to click “approve” after an AI completes a task is not necessarily meaningful human oversight. The ABC example demonstrates something much stronger.
After ABC Assist reformats the material, a journalist checks and edits the output. A local editorial leader reviews it using their knowledge of the community. It then passes through a sub-editing process before publication.
There are at least three sets of human eyes involved. Even more importantly, the workflow helps those humans make good decisions.
Journalists can compare the original and AI-assisted versions side by side, making changes immediately visible.
The AI prompts have been progressively refined to recognize issues such as legally sensitive stories, preserve important local context and avoid summarizing quotations in ways that could alter meaning.
Daily and weekly editorial quality feedback loops then feed back into the system.
This is not simply human-in-the-loop. It is a governed Human-AI workflow.
The process becomes the control
This points to an important shift in how organizations should think about AI governance.
Policies might say:
- AI outputs must be reviewed
- Sensitive information must be protected
- High-risk decisions require human oversight
- AI use must be transparent
- Errors must be reported.
But those statements only become operational governance when they are translated into the process itself.
- Who performs the review?
- At what point?
- What are they reviewing?
- What information do they need?
- When must something be escalated?
- What evidence is retained?
- Who is accountable for the final decision?
- How does feedback improve the AI system?
This is where the process becomes the operating contract between people, AI agents, automation and systems.
Instead of relying on employees to remember an AI policy, the workflow helps enforce the policy through the way the work is performed.
Human judgment belongs where context matters
One of Gartry’s strongest observations concerned local knowledge. A misplaced town name, an incorrectly spelled local surname or a small change in wording may appear trivial to an AI model. To a regional audience, it may immediately undermine credibility. People closest to the work understand these nuances.
They know which stories require extra care. They recognize when wording subtly changes meaning. They know what happens when breaking news disrupts the normal morning workflow.
This illustrates why deciding where human judgment must remain is one of the most important AI governance decisions an organisation can make.
The answer should not simply be “everywhere”. Nor should the objective automatically be to remove humans wherever technically possible.
The better question is:
Where does human context, accountability, experience or judgment materially improve the reliability of the outcome?
That is where human involvement becomes a control rather than an inefficiency.
Governance should create useful friction
AI promises speed, however sometimes the safest and most effective workflow deliberately slows things down.
- An additional review.
- A comparison between the source and AI-generated output.
- A supervisor approval.
- An escalation when confidence is low.
- A requirement to provide evidence before proceeding.
These controls create what is described as “useful friction.” The objective of good process design should therefore not always be to eliminate friction.
It should be to eliminate unnecessary friction while deliberately inserting useful friction where risk, accountability or judgment requires it.
This becomes particularly important as AI moves beyond generating content and begins performing operational tasks.
An AI agent might eventually:
- review a supplier’s insurance certificate;
- assess a compliance requirement;
- analyse an incident report;
- prepare a customer response;
- evaluate evidence;
- recommend an approval;
- complete part of a business process.
The governance question is no longer simply whether employees are allowed to use AI.
It becomes:
What is the AI allowed to do within this process, and where must responsibility return to a human?
AI adoption is a social system
Gartry also highlighted a problem many organizations are beginning to encounter. AI initiatives are often treated primarily as technology implementations.
A platform is selected. A pilot is announced. Employees receive training. Leaders then wonder why adoption remains inconsistent. AI adoption depends heavily on trust.
Employees want to know:
- Do I understand what this system is doing?
- Do I trust the safeguards around it?
- Who is accountable if it gets something wrong?
- Can I challenge the output?
- Can I report unexpected behavior?
- Will my feedback actually change the system?
- Does AI make my work easier or simply introduce another task?
These are governance questions, but they are also organizational design questions.
The success of AI therefore depends not only on the capability of the model but on the organization’s capability to use it responsibly.
Involve the people closest to the work
Perhaps the most important lesson from the ABC case is that the workflow was not designed solely by an AI team. Regional journalists, AI specialists, product teams and editorial leaders were deeply involved in testing and refining it. Months of trials examined accuracy, legal sensitivity, workload sustainability and the realities of working inside a busy newsroom.
That employee involvement was not simply change management. It was risk management.
The people closest to the work often know where a process will fail long before a technology team, consultant or executive can see it.
They know the exceptions. They know the shortcuts people take. They know where judgment matters. They know what information is missing. And they know whether a new workflow will survive contact with a busy Monday morning.
Bottom-up innovation therefore becomes both a risk management strategy and a value creation strategy.
Governance should enable innovation
There is sometimes an assumption that governance slows AI innovation.
The opposite can be true.
When employees understand what AI can do, what it cannot do, where the boundaries sit and who remains accountable, organizations can innovate with greater confidence.
Governance provides the guardrails that allow experimentation to happen safely. That means moving beyond static AI policies towards operational AI governance.
Organizations need to define:
- Purpose — What outcome is AI helping achieve?
- Boundaries — What may the AI do, and what must it never do?
- Human judgment — Where must a person review, decide or approve?
- Accountability — Who owns the outcome?
- Evidence — What needs to be recorded?
- Escalation — What happens when something unexpected occurs?
- Feedback — How does experience from the workflow improve the system?
- Transparency — Who needs to know AI was involved?
Once those decisions are embedded into everyday processes, AI governance stops being a document sitting somewhere in the organization. It becomes how the organization works.
The future belongs to organizations that can govern Human-AI work
The competitive advantage of AI may ultimately have less to do with who has access to the most powerful model. Most organizations will have access to increasingly capable AI. The differentiator will be the ability to integrate those capabilities safely and effectively into real work. That requires more than technology.
It requires process design, clear accountability, organizational trust, feedback loops and deliberate decisions about where human judgment belongs.
As the ABC example demonstrates, the most important innovation may not be the AI itself. It may be the workflow surrounding it. Because as AI becomes another participant in how work gets done, governance cannot sit outside the process. It has to become part of the process.
And that is where human-in-the-loop becomes much more than an AI safety principle. It becomes a fundamental design principle for the future of work.
