Essay
A full day of discussion about AI foundations, agents, enterprise risk, data readiness and the future of work kept returning to the same organisational question: how do we hold direction while the terrain keeps moving?
On 4 August 2026, I attended 6DAI Melbourne in my Beyond Blue capacity, alongside three peers.
It was good to spend the day comparing notes with them and to see some former colleagues across the room. The event was framed around the future of technology, innovation and work, and the program covered a lot of ground.
The morning began with infrastructure, privacy, trust and the foundations needed to support AI. It moved through intelligent decision-making, agentic systems and enterprise use cases. The afternoon turned to hidden risk, autonomous enterprises, the future of work, data readiness, governance and the emerging shape of human-agent workforces.
Those are different topics, but they did not feel separate by the end of the day.
The same concerns kept resurfacing in different forms: governance has to keep pace with adoption; human oversight has to be real rather than ceremonial; technology teams enable the work but cannot own every consequence; unmanaged AI use is already an organisational behaviour problem; operating models and roles will need to change; and digital, data and AI capability can no longer be treated as optional specialist knowledge.
None of those observations is especially novel on its own.
What mattered was the convergence.
We are trying to govern a capability that is changing quickly, appearing through multiple channels and moving into ordinary work before most organisations have settled on a destination.
By the afternoon, one thought kept returning:
Enterprise AI needs a compass, not just a roadmap.
Roadmaps still matter.
Organisations need to sequence work, choose platforms, fund capability, test use cases, develop controls and decide what comes next. Without that discipline, AI becomes a collection of disconnected pilots, vendor promises and local experiments.
But a roadmap is not the strategy.
A roadmap assumes the destination and terrain are understood well enough to plot a route. Enterprise AI gives us neither with much confidence. Models change. Vendor capabilities change. Staff find new uses before formal programs catch up. A safe and useful application in one part of an organisation may be inappropriate in another.
The road keeps changing while we are already travelling on it.
A roadmap can tell us what we planned to do next. It cannot, by itself, tell us whether that next step is still the right one.
That is the role of a compass.
These are not implementation questions. They are questions of organisational direction and judgement.
The sessions on trust, intelligent decision-making and agentic systems made the human question difficult to avoid.
“Human in the loop” is now used as though the presence of a person resolves the governance issue.
It does not.
A person who lacks the subject knowledge, context, time or authority to challenge an AI-generated output is not exercising meaningful oversight. They may be present in the workflow, but they are still likely to confirm what the system has already produced.
That is not a control. It is a rubber stamp with a human name attached.
Human oversight becomes meaningful when the person can understand the decision, recognise when confidence is weak, ask for evidence, override the output and stop the process when necessary.
The human also needs to know that this is their responsibility. Accountability cannot remain ambiguous simply because a model contributed to the answer.
A skilled human in the loop can improve judgement. An underprepared human in the loop can create false assurance.
The discussion about adoption, use cases and enterprise risk also returned to shadow AI.
I am not convinced that the underlying phenomenon is new.
People have a need. The approved environment does not meet it quickly or easily enough. They find a tool that does. Adoption moves faster than governance, support and visibility.
That is shadow IT.
AI changes the consequences and the evidence trail. Prompts can contain sensitive information. Outputs can be difficult to trace. Models and terms can change without much notice. Generated content can enter decisions, documents and customer interactions without a clear record of how it was produced.
Increasingly, the tool may also do more than store or process information. It may recommend, rank, generate, route or act.
So shadow AI deserves attention. But treating it as entirely separate from shadow IT risks forgetting what the behaviour is telling us.
Shadow use is often a signal that demand has outrun supported capability.
The response therefore cannot be limited to prohibition or another policy reminder. Organisations need to understand the demand, provide supported pathways that are useful enough to choose, set boundaries people can follow and observe enough of the environment to know where unmanaged use is occurring.
The later sessions on the future of work and data readiness reinforced another point: the workforce question is much broader than teaching people how to use a particular assistant.
Digital, data and AI literacy should now be assumed organisational capabilities.
That does not mean everyone needs the same level of expertise. Capability should be role-based. A clinician, procurement specialist, executive, analyst and software engineer will encounter different decisions, risks and responsibilities.
But a baseline should be common.
People should understand when an AI tool is appropriate, what information should not be entered, why an output may be wrong, when verification is required, when use should be disclosed and when a concern must be escalated.
Those are no longer specialist questions.
They are part of competent work in a digital organisation.
The baseline also has to extend beyond safe tool use. Employees will increasingly need to frame work clearly, provide context, judge the quality of outputs, manage exceptions and know which decisions should not be delegated.
That is what allows people to facilitate work with AI rather than merely receive whatever the system produces.
This changes the obligation on leaders. We cannot declare AI strategically important while leaving staff to build capability through personal experimentation, vendor webinars or whatever tool happens to be available.
If the capability is expected, it has to be supported.
This is the point that is often missing from the enterprise AI conversation.
Organisations cannot introduce systems that generate, recommend, route and increasingly act, then assume the existing operating model will absorb them without change.
If employees are expected to frame work, supervise outputs, exercise judgement and intervene when confidence is weak, those responsibilities must be designed into roles. Workflows, decision rights, hand-offs, escalation paths and measures of performance all need to reflect the new division of labour.
You cannot automate parts of the work and leave the accountability model untouched.
This is not simply a headcount exercise. It is organisational remodelling: deciding what the system should do, what people must continue to own and how evidence moves between them.
In some roles, less time may be spent producing a first draft or processing routine work. More time may be spent supplying context, testing assumptions, managing exceptions, interpreting consequences and maintaining the human relationships that the system cannot carry.
Technology teams can provide supported platforms, controls and evidence paths. They cannot decide organisational purpose, role accountability or acceptable consequences alone.
That shift will not happen because a tool has been licensed. It requires deliberate design and a workforce capable of operating within it.
The future workforce is not simply the existing organisation with AI added. It is an operating model in which work, authority and evidence are redistributed - and must be governed accordingly.
The useful part of the compass metaphor is not that it predicts every turn.
It does not.
It keeps purpose, limits, accountability, skilled judgement, organisational design and observable evidence in view while the organisation learns.
This is where governance becomes operational. Not a policy written once. Not a committee that approves the first use case. Not a training module that can be marked complete.
Governance has to remain connected to how the capability is actually being used, how work and authority are changing, what outcomes the system is producing and where the original assumptions have stopped holding.
I have argued before that AI does not create a new domain of trust. It changes the behaviour of systems and organisational arrangements that already exist.
The same is true here.
Enterprise AI does not require us to abandon roadmaps, delivery disciplines or technology foundations. It requires us to stop mistaking them for direction - and to redesign the organisation where the work itself is changing.
At the end of the day, what stayed with me was not a particular product, demonstration or prediction. It was the way the different sessions kept returning to the same underlying challenge.
AI is moving from a specialist technology conversation into ordinary organisational life.
The question is no longer whether organisations will use it.
The question is whether they can remain competent, accountable and capable of changing course as they redesign work around it.
A roadmap tells teams what comes next.
A compass helps the enterprise decide whether next is still the right direction.
Context: this essay was prompted by my attendance at 6DAI Melbourne, The Future of Tech, Innovation & Work, on 4 August 2026. The interpretation and conclusions are my own.
