‍

‍

‍

Dirk J. Primus is an associate professor at UNSW's Australian Graduate School of Management, where he researches strategy, innovation, and technology adoption. He sat down with Dave Kenyon from OAHI's Evolving Workforce podcast.

‍‍‍

The short answer

Most enterprise AI rollouts do not fail because the technology is weak. They fail because nobody assigned ownership of the layer between the technology and the people using it. Buying licences is a procurement decision. Deciding which workflows change, who is accountable for AI-assisted output, and where human sign-off still applies is a management decision, and most organisations skip it entirely.

That gap, more than any limitation in the AI itself, is what decides whether adoption spreads across a business or stalls out team by team.

Why the same tool produces two different outcomes

I spend a lot of my time talking to people living this problem in real time, mostly through my MBA students, many of whom sit at a fairly senior level in their own organisations, and through the practitioners I work with directly. So I get a fairly honest view of what is happening on the ground, not just what makes it into a board pack.

Here is the pattern I keep seeing, in Australian organisations and elsewhere.

A business rolls out enterprise access to Claude, Copilot, or ChatGPT. Licences are deployed. A pilot runs somewhere. Leadership believes the organisation is “doing AI.” Six months later:

• One team has genuinely changed how it works. The tool is embedded in daily workflows and the output quality has visibly improved.
• Another team, with access to the identical platform, is still using it to draft the odd email and nothing more.

If the underlying technology is the same, the difference has to be organisational, not technical. Deploying a thousand Copilot licences means nothing unless someone actually uses them. That is not a technology failure. It is a diffusion problem, and diffusion is a social process, not a technical one.

Think in layers, not in tools

The easiest way I explain this to my students, and I know it sounds a bit simplified, is to picture an organisation as a layered cake.

The technical layer sits at the base. This is what IT is responsible for: the systems of record and systems of engagement, and how they connect. Most AI strategy work stops here. It treats “we bought the licences” as equivalent to “we changed how work gets done,” and those are not the same thing at all.

The management layer sits above it. This is where someone decides where information flows, who acts on it, and when. This is the layer that actually determines whether adoption succeeds, and it is the layer most AI strategies never address.

Why business units drift apart

If different business units are left to work out their own rules in isolation, what I call organisational slices, you end up with pockets of sophisticated use sitting right next to pockets that never move past drafting emails. There is no shared understanding of where the line sits between the two.

This does not mean the answer is to connect every layer to every other layer and standardise every workflow around the AI. Sometimes the right managerial decision is to keep certain layers deliberately separate, so a mistake or a shortcut in one part of the business does not spread into a function where the consequences are more serious, anything that ends up in front of a client, a regulator, or an auditor. The skill leaders need is not “integrate everything.” It is knowing which connections create value and which create risk.

The two failure modes I see most often

Treating AI adoption as an IT problem, and delegating it entirely downward, tends to produce one of two outcomes.

1. AI theatre. This is the direct descendant of what used to be called innovation theatre. Conversations at board and middle management level stay performative. AI, engine, licence, and deployment get mentioned enough times in a meeting to signal progress, without anyone being held accountable for whether adoption actually changed a workflow.

2. Silent breakdown. The technology works exactly as designed, but the interaction between the human and the system breaks down because nobody redesigned the workflow, the review process, or the accountability structure around it.

Neither shows up cleanly in a project status report. Both show up a year later as wildly inconsistent quality of AI-assisted work across teams that were handed the identical tool.

What good looks like

The organisations getting this right are not the ones with the most sophisticated AI use cases. They are the ones where someone above the technical layer, typically a CHRO, a COO, or a senior operations leader, is acting as a systems thinker and an orchestrator across the slices.

That person can answer, in plain terms:

• What changes about a manager's accountability when their team starts using AI to draft client-facing material
• Where human sign-off still needs to sit
• Where a consistent standard needs to exist across the business, rather than a dozen local interpretations of “using AI responsibly”

Three things to do deliberately, not leave to chance

1. Name who owns the connective layer between teams, not just who owns the licence budget.
2. Measure actual adoption and behaviour change, not licence counts. A rollout should be judged on whether the work genuinely changed, not whether everyone logged in once.
3. Resist pushing AI into every layer of the business at once before anyone has worked out where the higher-risk boundaries sit, particularly anywhere with compliance, safety, or regulatory exposure attached to the output.

The practical takeaway

An AI strategy document is not the same thing as an operating model. Before the next wave of licences goes out to another division, the more useful question for a CEO or CHRO to ask is not “does this technology work,” but:

Who owns what happens when one team's use of AI touches the next team's accountability?

That single ownership gap, far more than any limitation in the underlying AI, is what decides whether adoption compounds across a business or stalls out team by team.

Dirk J. Primus is an associate professor at UNSW's Australian Graduate School of Management, specialising in strategy, innovation, and technology. These reflections are adapted from his conversation on the Evolving Workforce podcast, where he spoke with host Dave Kenyon of OAHI about AI strategy and the workforce.

‍

‍