What happened
It was reported that in June 2025, Commonwealth Bank introduced an AI voice bot to handle customer calls. The following month, the bank announced that 45 positions in its Customer Service Direct team were no longer needed. Another 45 customer messaging roles were also being considered for removal.
The bank said the technology had reduced call volumes by approximately 2,000 calls a week. With fewer calls reaching employees, it argued, fewer people were needed to answer them.
But staff were experiencing something quite different.
According to the Finance Sector Union, call volumes were increasing, employees were being offered overtime, and team leaders who would not normally answer customer calls were being asked to help manage the workload.
The union challenged the redundancies through the Fair Work Commission, questioning whether the bank’s claims about the AI system reflected what was actually happening.
On 21 August 2025, Commonwealth Bank reversed its decision. The bank acknowledged that its original assessment had failed to consider all relevant business factors and that the 45 positions were not, in fact, redundant.
Affected employees were offered the opportunity to remain in their roles, move into other positions or accept voluntary exit packages. The bank apologised and said it would review its processes.
The future of the additional 45 proposed job cuts remained unclear at the time.
What an auditable version would have shown
The central question was whether the AI system had genuinely reduced the amount of work employees needed to do.
The bank pointed to fewer calls. The union pointed to increasing workloads, overtime and managers answering phones.
Both could be true in different parts of the operation. A bot might handle thousands of simple enquiries while leaving employees with more complicated calls that take longer to resolve. Counting calls alone would not establish how much work had disappeared.
An auditable system would have tracked the number of calls received, how many the bot resolved, how many required human assistance and how much time employees spent dealing with the remaining enquiries.
A MetricRecord could preserve those measurements over time, while a ConductRecord could help establish how the bot handled individual interactions.
That evidence could then be considered alongside staffing levels, overtime, customer outcomes and the actual workload before decisions about redundancies were made.
The public reporting does not establish what internal data CBA retained or exactly how its original assessment was calculated. What the bank acknowledged was that its assessment had not adequately considered all relevant business factors.
Where the gap was
The AI bot was not necessarily the problem. It may have been successfully handling some customer enquiries.
The problem was what the organisation concluded from its performance.
Headlights identifies scope drift in the way a measure of AI performance became part of the justification for removing jobs. A reduction in calls reaching a particular team was treated as evidence that the same amount of human labour was no longer required.
But fewer calls do not automatically mean less work.
A five-minute enquiry handled by a bot is not equivalent to a complicated customer problem that takes an employee half an hour to resolve. Nor does a successful automated interaction necessarily mean the customer received the help they needed.
There is also a human behaviour worth examining. Once an organisation invests in technology intended to improve efficiency, it can become tempting to interpret early results as proof that the investment is delivering the expected benefits.
That does not establish what happened inside CBA’s decision-making process. It does, however, explain why independent checks matter when performance figures begin influencing decisions about people’s livelihoods.
The bank eventually acknowledged that its assessment was wrong. By then, employees had already been told their jobs were at risk.
Headlights’ behavioural interpretation is not a finding made by the Fair Work Commission or the bank.
What governance should have looked like
Before announcing redundancies, CBA should have been able to demonstrate that the AI system had reduced the need for human work over a meaningful period.
That assessment needed more than a headline figure about call volumes. It should have included the number and complexity of enquiries reaching employees, time spent resolving them, overtime, customer wait times and whether the bot was actually resolving problems.
A MetricRecord could preserve the performance data behind those conclusions and allow the figures to be checked against the underlying operational records.
The bank could also have required an independent review before AI-related performance claims were used to justify workforce reductions. That review should have considered evidence from employees and their representatives, particularly where their experience contradicted management’s assessment.
A sustained period of reliable results would provide a stronger basis for workforce planning than an early estimate following a technology rollout.
The lesson is not that AI cannot reduce the need for certain jobs. It is that an organisation needs to establish that the work has genuinely disappeared before deciding the people doing it are no longer required.
In this case, Commonwealth Bank made that decision, then acknowledged it had been wrong.
Failure Pattern: Commonwealth Bank announced 45 redundancies after introducing an AI voice bot and reporting reduced call volumes. The Finance Sector Union challenged the decision, citing continued workload pressures. The bank subsequently acknowledged that its assessment was flawed and reversed the redundancies.
Governance Principle: AI performance claims should not be used to justify workforce reductions without reliable evidence of sustained changes in workload, customer outcomes and staffing requirements. The people affected should have a meaningful opportunity to challenge the assumptions.
The reference implementation of MetricRecord and ConductRecord is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- Commonwealth Bank Reverses Job Cuts Decision Over AI Chatbots (Bloomberg)
- CBA reverses AI-driven job cuts, admits error (ACS Information Age)
- CBA to keep roles it had reviewed after installing AI bot (iTnews)
- Commonwealth Bank backflips on controversial axing of dozens of jobs for AI chatbot (Yahoo Finance AU)