The Redundancy That Wasn't

What the Commonwealth Bank's AI reversal teaches Australian contact centres about reliability

In July 2025, the Commonwealth Bank of Australia told 45 customer service workers their roles were no longer needed. The reason was a new AI voice bot. The bank said the technology had reduced call volumes by around 2,000 calls a week, which left a smaller team to handle the harder enquiries and made the 45 positions surplus to requirements. It was a clean, familiar story: automate the routine, keep a lean human core, book the saving.

By 21 August, the story had fallen apart. The Finance Sector Union had taken the matter to the Fair Work Commission, arguing the bank was not being straight about what was actually happening on the floor. Call volumes were not falling. They were rising. CBA was offering staff overtime to cope, and directing team leaders back onto the phones to clear queues. The bank conceded that its assessment "did not adequately consider all relevant business considerations, and this error meant the roles were not redundant." The 45 workers were offered their jobs back, redeployment, or the redundancy they had originally been given. The union called it a massive win.

The easy headline wrote itself: AI failed. It is the wrong headline, and for anyone running a contact centre it is a dangerous one, because it points to the wrong lesson. The voice bot did roughly what a voice bot does. It absorbed a share of simple, high-frequency queries. What broke was not the model. What broke was the operating model wrapped around it.

The bot worked. The design didn't.

Strip the emotion out of the CBA case and you are left with a design failure that has almost nothing to do with the quality of the AI.

The bank made a common but consequential assumption: that the calls left over after automation, the residual, were roughly the same shape as the calls before automation, just fewer of them. Automate 30 per cent of the volume, the thinking goes, and you can cut something like 30 per cent of the capacity. It is arithmetic, and it is wrong.

When you automate the routine, you do not shrink the queue proportionally. You change its composition. What remains is denser, harder, and more variable: the ambiguous cases, the escalations, the customers the bot could not satisfy and who now arrive at a human already frustrated. On top of that, a bot that partially fails does not simply drop the call. It often generates a new one. The customer hangs up, stews, and rings back, this time angrier and this time counted twice. The residual is not a smaller version of the original workload. It is a more demanding one, and it can grow even as raw volume appears to fall.

CBA cut capacity as if it were removing the median call. The work that was left was made of exceptions. That is the whole story in one sentence.

Reliability is a discipline, not a feature

This is where the contact-centre conversation needs to borrow from a field that has thought harder about failure than most: high-reliability organisations. Aviation, nuclear operations, emergency medicine. These are environments where the average case is handled almost without thought and the entire discipline is built around the exceptions, because the exceptions are where people get hurt.

High-reliability thinking has a few habits worth naming, because CBA violated most of them.

The first is a preoccupation with failure. Reliable operations assume the edge cases are where the risk lives and they instrument for them relentlessly. CBA appears to have instrumented for the happy path, the raw call count, and treated a falling headline number as proof the system was working. It was measuring the wrong thing with confidence.

The second is sensitivity to operations, a fancy phrase for knowing what is actually happening on the floor. The people who knew the queues were growing were the workers and their team leaders, the same people being made redundant and then quietly asked to do overtime. The signal was present. The organisation was not set up to hear it, or chose not to.

The third is deference to expertise. In a reliable operation, when the person closest to the work says the system is not coping, that carries weight regardless of what the dashboard says. CBA had the dashboard say one thing and the floor say another, and it went with the dashboard until a union and a tribunal forced the correction.

None of this is an argument against automation. It is an argument that automating a contact centre is a reliability problem, not a cost problem, and the two are solved very differently. If you treat it as a cost problem, you ask how many humans you can remove. If you treat it as a reliability problem, you ask a harder and more useful question first: where does human judgement remain load-bearing, and what happens to the system when the automation hits its limits? You map the tail before you touch it. CBA cut first and discovered the map afterwards.

The number that matters is the one you keep

IBM offers the cleaner articulation of the same principle, and it is worth holding up next to CBA precisely because IBM got the framing closer to right.

IBM's internal HR assistant, AskHR, resolves something like 94 per cent of routine human resources queries without a person. On its face that is a spectacular automation rate, the kind of figure that tempts a leadership team to cut the HR function to the bone. But IBM has been explicit about the other end: the remaining 6 per cent, the cases involving nuance, exceptions, and ethical judgement, still need a human, and that 6 per cent is not a rounding error. It is the part of the work where getting it wrong is expensive, and where a wrong answer delivered confidently is worse than no answer at all.

The instinct is to celebrate the 94 per cent. The discipline is to obsess over who owns the 6 per cent. That is the number that determines whether your operation is reliable, because that is where the failures that damage customers and reputations actually occur. Automation moves the human workforce up the value chain toward exactly these cases. An organisation that treats the residual as trivial, as CBA did, has automated its way into a more fragile operation while congratulating itself on efficiency.

This is not an Australian problem, but it is an Australian story

The pattern is global. Klarna is the most-cited example internationally, and for good reason. In 2024 the Swedish fintech deployed an OpenAI-based assistant it claimed could do the work of 700 agents, cut roughly 1,000 customer service roles, and paused hiring for the better part of a year. Within months, customer satisfaction had fallen far enough that the company began rehiring, with its chief executive conceding the strategy had leaned too hard on cost-cutting and that the quality of human interaction was worth investing in. The broader data tells the same story from a distance: staffing firm Robert Half found that roughly a third of US hiring managers had rehired for a role they had eliminated because of AI, with finance leading the reversals. Analytics firm Orgvue found that among leaders who had made redundancies specifically because of AI, more than half later judged the decision a mistake.

So the CBA case is not unusual in substance. What makes it a distinctly Australian story is the speed and visibility of the correction.

In the United States, where union density is low, these reversals tend to happen quietly, months later, through attrition and rehiring that never gets attached to the original decision. In Australia, the industrial architecture works differently. A union with real coverage in financial services took the bank to the Fair Work Commission within weeks, forced disclosure of what was actually happening to call volumes, and turned a quiet operational stumble into a public admission of error. For an Australian contact-centre leader, that is the material point. Here, a botched AI deployment does not stay an internal operations issue. It becomes an industrial one, a reputational one, and a matter of public record, quickly. The cost of getting the operating model wrong is higher and it arrives faster.

The honest complication

It would be tidy to present CBA purely as a company that trusted its technology too much. The truth is messier, and the mess is instructive.

The union's central accusation was not really about the bot. It was that the bank was, in its words, dressing up job cuts as innovation, using the AI narrative as cover for a more conventional cost exercise. That charge gains weight from a detail that sat alongside the redundancies: reporting that CBA was hiring for similar customer-facing roles offshore, in India, even as it declared the Australian roles surplus. The bank was, at the same time, deepening its AI ambitions through a high-profile partnership with OpenAI.

Take all of that together and a more uncomfortable reading emerges. The failure was not simply that leadership believed the AI hype. It is that headcount decisions were being made ahead of, and somewhat independently of, any rigorous understanding of the operation. AI became the available justification for a workforce decision that had other drivers. That is the deeper hazard, and it is not solved by better technology. It is solved by refusing to let a compelling narrative, whether that narrative is automation or offshoring or anything else, substitute for an actual analysis of where the work lives and who is equipped to do it.

Stronger, or just smaller

There is a striking postscript to the CBA episode, and it comes from the same source as the original decision.

In May 2026, nine months after the reversal, CBA chief executive Matt Comyn used a contributed piece in the Australian Financial Review, and the stage at the bank's first Accelerate AI conference, to speak with unusual directness about AI and work. He said plainly that AI will remove jobs across the economy and that employers have a duty to help their people adapt rather than offer false comfort. Pretending every role can be preserved, he argued, does not protect anyone, it simply leaves workers surprised later. And he drew a distinction that could serve as the thesis of this entire article: the difference between using AI to build a stronger organisation and using it to simply make the organisation smaller. The better question, he suggested, is not how many people you can remove, but whether the technology is making the business stronger or merely stripping out cost.

It is a genuinely good distinction. It is also, almost precisely, the lesson of his own bank's 45 roles.

That tension is worth sitting with rather than scoring points off. The most charitable and probably the most accurate reading is that the CBA episode taught the organisation something, and that Comyn's later framing is what the lesson sounds like once it has been absorbed at the top. The bank has backed the words with money: a $90 million Future Workforce Programme running over three years, built around reskilling pathways, internal mobility, and early visibility for staff into how their roles are changing, developed, notably, in consultation with the unions it had recently been fighting. Comyn's line that the people who will matter most are those who combine customer understanding, risk judgement and the ability to direct AI systems is exactly right, and it is the same point as the 6 per cent at IBM: the human role does not disappear, it concentrates into the work that is hardest and highest-stakes.

The test, of course, is not the speech. It is whether the next deployment maps the tail before it cuts, or whether the dashboard wins again. Stated philosophy is easy. The 45 roles are a reminder that execution is where reliability is actually proved or lost.

There is early evidence of what that next deployment looks like. In July 2026, CBA detailed an AI orchestration agent, co-developed with Microsoft over two years, that now sits at the front of its retail customer support. The design is telling: the agent interprets a customer's intent and routes it dynamically, sending straightforward queries to conversational AI while handing sensitive conversations to a human specialist, with AI supporting in the background rather than replacing the person. The bank reports that around 84.6 per cent of self-service messaging interactions were resolved end-to-end in May 2026, which means the architecture has a built-in answer to the question the voice bot episode left open: who owns the other 15 per cent. Routing to humans by design, rather than discovering the need for them after the redundancy notices, is precisely the shift this article argues for. The test now moves with the technology, because CBA intends to extend the platform into voice and, eventually, across the entire bank. Whether headcount decisions follow the design discipline, or run ahead of it again, is the part worth watching

Re-skilling is necessary. It is not sufficient.

Comyn's answer to displacement, and it is the standard corporate answer, is re-skilling. It is the right instinct and it deserves support. But it is worth being honest that re-skilling, even done well, does not dissolve the social consequences of this transition. It addresses the individual capability question and leaves several harder questions largely untouched. A reliability mindset, which is built around the failures that averages conceal, should be the first to say so.

Consider what remains even in the optimistic case, where a displaced worker is retrained to equivalent skill and equivalent pay.

There is a timing gap. Re-skilling is never instant and never universal. Between the displacement and the new role sits a period that falls hardest on the people least able to absorb it: older workers, those in regional areas, people with caring responsibilities who cannot simply relocate or study full-time on short notice. "Re-skilled to the same level" is an average that hides a very uneven distribution of who actually makes it across.

There is an identity cost. A contact-centre role is not only a wage. It is a routine, a team, a source of competence and standing. Even a lateral move to an equivalent job carries a real psychological cost that a pay comparison does not capture. People lose the quiet value of being the person who knew how things worked.

There is the erosion of the entry rung. Contact centres have long been on-ramps, first jobs, second-chance jobs, the place people step onto before climbing. Automate the routine tier and you do not only displace today's workers, you remove the ladder for the next cohort. The re-skilling conversation is almost always about incumbents. It rarely accounts for the people who now never get the entry role at all.

There is geographic concentration. In Australia, contact centres are often meaningful regional employers. Displacement does not spread evenly across the economy, it lands on particular towns and suburbs, which turns an operational decision into a community one.

And there is trust. The CBA episode is, at bottom, a story about what happens when change is experienced as something done to a workforce rather than with it. Even an excellent re-skilling program fails socially if people do not trust the intent behind it. Process legitimacy matters as much as the outcome, which is precisely why Comyn's emphasis on transparency and on bringing the unions along is not a soft add-on. It is part of the engineering.

The mature position is neither "AI will retrain everyone" nor "AI will destroy work." It is that re-skilling is a necessary response to a real problem and an incomplete one, and that the incompleteness is not an argument for delay. It is an argument for designing the transition with the same rigour you would apply to the technology itself.

What reliability-first deployment actually looks like

None of this counsels cautions about AI in the contact centre. Deployed well, conversational automation is genuinely transformative, and the organisations that hang back will lose to the ones that move. The argument is narrower and more practical: sequence the reliability work before the headcount work, not after.

In concrete terms, that means a few things. Instrument the tail before you cut it, so you understand what the residual workload actually looks like once the routine is automated, not what you assume it looks like. Define exception ownership explicitly, so there is a named, resourced answer to the question of who handles the cases the automation cannot, before those cases start arriving. Measure the composition of the queue, not just its volume, because a falling call count can hide a rising concentration of hard work. And treat human expertise as a designed component of the system, deliberately positioned where judgement carries the load, rather than as a residual cost to be trimmed to the last defensible unit.

CBA had the technology working. What it lacked was the reliability discipline to know what the technology left behind. The 45 roles were never redundant. They were load-bearing, and the bank discovered that the expensive way, in public, with a union and a tribunal supplying the correction that its own instrumentation should have caught first.

The lesson for Australian contact centres is not that AI is risky. It is that automation without reliability engineering is risky, and that the difference between the two is a discipline you can build before the mistake, or one the market will teach you afterwards.

The work that survives automation is not the work that remains. It is the work you deliberately design for. Defining those roles, before the redundancy notices, not after the reversal, is where OpsArchitecture helps.

Built on Rigor. Engineered for Scale.