Organisations are not waiting for AI to mature before deploying it at scale. The gap between how fast the technology is being adopted and how carefully it is being governed is widening, and the consequences are beginning to show up in places most executives are not yet looking.

According to Stanford University’s 2025 AI Index Report, the proportion of organisations using AI in at least one business function jumped from 55% in 2023 to 78% in 2024, a 23-percentage-point leap in a single year. The World Economic Forum’s Future of Jobs Report 2025, which surveyed more than 1000 employers representing 14 million workers across 55 economies, found that 86% of those employers expect AI and information processing technologies to transform their business by 2030, and that 40% plan to reduce their workforce in roles where AI can automate tasks.

Dr Lynn Gribble, an Associate Professor in the School of Management and Governance at UNSW Business School, has spent years studying how technology reshapes organisations and the people within them. In seeking to cut costs through AI adoption, she believes most organisations are focusing on efficiency gains rather than on the harder work of redesigning how they operate. And the potential costs of this misalignment are accumulating in ways that will be difficult to reverse.

Swedish fintech Klarna, for example, became one of the most widely studied examples of what that misalignment looks like in practice. Between 2022 and 2024, Klarna eliminated approximately 700 positions, primarily in customer service, and replaced them with an AI assistant that, at its peak, handled up to three-quarters of all customer interactions. By early 2025, satisfaction scores were deteriorating on complex service interactions, and CEO Sebastian Siemiatkowski acknowledged the company had gone ‘too far’, admitting cost had been ‘a too predominant evaluation factor’ – and began rehiring human staff.

The efficiency argument is the wrong starting point

The dominant conversation around AI in business has centred on the efficiency dividend: doing more with less, faster. For A/Prof. Gribble, that framing is both a distraction and a trap. When organisations focus narrowly on time, they are likely to expect people to just do things more quickly. Instead, it is about now that some time is freed up, how can this person add more value rather than just doing more? Organisations and people need to look beyond optimising within an existing structure and ask the bigger questions, including whether the structure itself still makes sense.”

“When we just talk about the AI dividend, we’re really not changing anything, and we need to really move to that understanding,” she said. “How could we reconsider or reimagine our whole business so that AI becomes another resource? What is our strategy? How are we going to differentiate ourselves?” she asked.

Learn more: How AI is changing work and boosting economic productivity

“We’ve seen the failings of companies that have adopted and rolled out AI, but nobody has had the skills to really check it (or they haven’t checked it). That can be so damaging to business and really difficult to recover from. The smarter companies are going slowly on this, rather than quickly. Yes, people are saying, ‘We’re yet to see the real dividend or the real benefits.’ But the issue is because they’re just focused on that efficiency argument, instead of focusing on rethinking their business and using AI for genuine competitive advantage.”

A study by researchers at MIT and Stanford found AI tools increased worker productivity by 14% in a customer support setting, with the largest gains among less experienced workers. Where organisations simply hand out access and move on, gains are marginal, and risks rise. The efficiency argument, A/Prof. Gribble argues, actively draws attention away from the harder question of how work itself should be redesigned.

Routine is just the beginning

The conventional view of AI automation focuses on the transactional: processing invoices, managing bookings, and routing customer enquiries. A/Prof. Gribble pointed to automated expense management in systems (such as SAP) as an example of an implementation that, in some cases, has created more friction than it has resolved. “If there’s anything out of the ordinary, it often creates a loop where it could be five times more onerous,” she said. “Putting AI in, in these kinds of circumstances, may not actually make things better for me as a client.”

A more consequential shift comes as AI capabilities extend beyond routine processes. An important and emerging trend is where AI systems are capable of exercising something that resembles judgement: aggregating data, assessing context, generating recommendations and, in some instances, taking action.

A/Prof. Gribble offered an example from medical practice administration, where AI agents now book appointments, follow up on test results and direct patients to human staff only when clinical complexity demands it. “There is a future where agents are talking to agents and the human is potentially not in that loop at all,” she noted, describing it as a matter of serious concern when applied more broadly, particularly where financial transactions or identity verification are involved.

The distinction between automating tasks and automating judgment matters for how organisations think about accountability. A task can be audited, but judgement, once embedded in a system without transparent logic, is harder to interrogate. As AI takes on more of the supervisory and coordination functions that management layers once performed, A/Prof. Gribble said the question of who is responsible for outcomes and who has the knowledge to assess them becomes harder to answer.

The hidden cost of institutional knowledge walking out the door

The large-scale management layoffs that might be expected to accompany structural change are beginning to emerge in some organisations, particularly in the technology sector. Elsewhere, A/Prof. Gribble said the change is subtler: roles are not filled when vacated; responsibilities are redistributed upward or absorbed into AI workflows; or teams are restructured without formal announcement. What is rarely counted in these arrangements is what leaves with the people.

A/Prof. Gribble drew on her earlier research into job loss and retrenchment during the economic rationalism wave of the 1980s and 1990s, and drew direct parallels with what she is currently witnessing in the world of work. When organisations shed experienced managers in pursuit of efficiency, they do not simply cut salary costs; they remove decades of institutional knowledge, such as informal understandings of how a business works, judgements about when to deviate from procedures, and important relational knowledge that makes processes run smoothly.

Subscribe to BusinessThink for the latest research, analysis and insights from UNSW Business School

“There could be a reason that’s never been passed on why we do something in a particular way, because it’s just been accepted,” she reflected. “That could be bad, but it could be largely good, because we might know that checking every loan that we don’t approve for three things could give us 5% more people, or 5% less people, or whatever. Those things are really, really useful.”

A Gartner study of 350 global enterprises found that 80% had reduced staff as part of AI adoption, but that there was no meaningful correlation between the scale of those cuts and financial returns, meaning organisations had sacrificed institutional knowledge for no measurable gain. Gartner has since predicted that by 2027, half of all companies that attributed headcount reductions to AI will rehire staff to perform similar functions, its absence only becoming visible once the systems designed to replace it begin to fail.

Cutting jobs for AI under false pretences?

There is an additional and timely pressure point that A/Prof. Gribble highlighted. In the current global geopolitical environment, shaped in part by the oil crisis in the Strait of Hormuz, she said this could be a precursor to the kind of economic conditions that historically drive job losses, giving organisations a ready justification for cuts they may have been planning anyway.

“People are going to turn to AI as a way of telling themselves, ‘Well, it was time to rationalise the business anyway, because now we’ve got AI,’” she observed. The risk, in her view, is that economic pressure and the efficiency narrative will combine to produce a return to the pattern she researched three decades ago: fewer people doing more work, with quality declining beneath a surface of apparent productivity.

This risk is compounded when AI systems are introduced without the organisation having first captured and codified what those departing managers knew. Rather than inheriting the organisation's institutional wisdom, the AI just captures surface knowledge: the written procedures, the recorded exceptions, and the formal outputs – while A/Prof. Gribble noted that the knowledge source (humans) that made those outputs meaningful has left the organisation.

Research published in the Journal of Political Economy, for example, offers a rigorous theoretical basis for this concern. The framework, developed to study AI’s effects on the knowledge economy, found that the defining characteristic of knowledge work is that production know-how is predominantly tacit – developed through repeated observations of practical successes and failures, and therefore inherently embodied in individuals rather than capturable in documented systems.

Furthermore, economists at the Dallas Federal Reserve have proposed that AI may automate codified knowledge (the book-learning kind) but not tacit knowledge derived from experience, meaning AI can substitute for entry-level workers who possess formal knowledge (but not yet experience), while complementing experienced workers whose tacit knowledge cannot be replicated. The implication for organisations that use AI-driven restructuring to remove experienced middle managers is stark: what is being automated away is the very layer of experiential judgement the organisation most depends on and least knows how to replace.

Accountability in an AI-intermediated organisation

When an AI system produces an output (such as a report, a recommendation or a decision) and that output travels through a chain of review without being meaningfully interrogated, the quality controls that once existed in human relationships between managers and their teams dissolve.

“The real question of who is the quality control and where is the quality controller becomes key to actually how we bring AI into a business,” A/Prof. Gribble said. In a pre-AI workflow, a manager reviewing a subordinate’s work could rely on accumulated knowledge of that person’s diligence and judgement. In an AI-mediated workflow, that trust relationship is disrupted: “You don’t know I’ve done the research. You don’t know that I’ve checked whether those links are the right links, or that’s the right person, or whatever the case may be. And people have too high a trust.”

"When we just talk about the AI dividend, we’re really not changing anything, and we need to really move to that understanding"

LYNN GRIBBLE

A/Prof. Gribble argued that AI should be placed on the organisational chart with a defined role, clear responsibilities, boundaries and transparency, in the same way that executive functions are defined. “Put AI on your organisational chart, and formalise what AI will do, just like you know what your chief customer officer or your chief marketing officer does. You will see what AI does in our business, so it’s transparent and clear to customers, and it’s clear to everybody in the business.” Without that clarity, she suggested organisations are likely to realise gains that are efficient but not necessarily effective.

Reimagining structure for an AI-enabled workforce

A/Prof. Gribble’s five-to-ten-year view is that AI will be a visible, accountable component of the organisational hierarchy: a tool not just deployed by individuals, but a resource managed at the enterprise level. “I think that in five years, the corporate structure will have AI very clearly on it, just like we formally define roles held by humans in the organisational chart. We will see AI very clearly and how it is used in business,” she predicted.

The HR function, she argued, is well positioned to own that transition, in a role it had not yet broadly assumed. A/Prof. Gribble explained: "HR manages human resources, and they shouldn’t treat this as just a computer. It doesn’t just go to asset management or to IT management. So who should be managing this resource? Actually, HR should be.”

Some organisations are already moving in this direction. Moderna, for example, announced it was merging its technology and human resources departments into a single function, appointing a Chief People and Digital Technology Officer, in part due to its partnership with OpenAI and the technology's influence on the workforce. Moderna’s people chief, Tracey Franklin, said she is redesigning teams across the company based on what work is best done by people versus what can be automated, with roles being created, eliminated and reimagined as a result.

Five key takeaways for business leaders and practitioners

The changes to organisational structure underway now (such as the erosion of coordination roles, the loss of institutional knowledge, and the diffusion of accountability) are accumulating without the visibility of a formal restructuring announcement. For executives and senior managers, A/Prof. Gribble said this has significant implications.

1. AI efficiency framing needs to be replaced with a redesign framing. Organisations that use AI primarily to do the same things faster accrue short-term productivity gains at the cost of strategic differentiation. Those well-positioned in a decade will now be asking how AI changes what they should be doing, not merely how quickly they can do what they already do.

2. Second, AI needs governance architecture, not just deployment. Placing AI on the organisational chart (defining its role, boundaries, and accountability relationships) is the mechanism by which organisations retain control over quality, maintain transparency with customers and staff, and preserve the human oversight that prevents governance failures of the kind A/Prof. Gribble described.

3. Institutional knowledge needs to be actively managed before it walks out the door. As experienced managers leave and their roles are absorbed into AI workflows, organisations need strategies to capture and codify the knowledge they held. Without that, the AI inherits surface-level knowledge of an organisation, rather than its real intelligence.

4. The humans in the loop need to be capable of the oversight function they are being asked to perform. A review step that exists on paper but is routinely bypassed because the reviewer lacks the expertise or the time, or because the incentive to engage is a liability.

5. The skills divide is real and widening. Organisations that invest in structured AI capability development and genuine competence (and not just providing access) will retain the human judgement that makes AI output trustworthy. Those that do not will produce outputs that appear well-formed to the uninformed, but shallow and potentially flawed to those who know the difference.