Commonwealth Bank cut 45 call-centre roles on the strength of a voice bot, then reversed the cuts two weeks after the union asked for evidence. The 2,000-calls figure was never shown to be wrong. It was a gross count offered for a net decision, and the bank already knows the fix: it reconciles its own profit figures every half-year.
On 7 August 2025, the Finance Sector Union posted an update for its members at the Commonwealth Bank of Australia. Most of the post asked members for their stories. One sentence in the middle held the whole dispute: "Despite us seeking more information, CBA has failed to provide evidence to substantiate its claim that the voice bot has reduced call volumes by 2000 per week."
Two weeks later the bank reversed the 45 redundancies the voice bot was supposed to have made possible, and apologized. What its statement leaves out matters. It does not say the 2,000 was wrong, and nothing in the public record I could read shows that it was. What the bank conceded was something else, and the distance between the number it offered and the thing it got wrong is the part of this story that matters to anyone who builds, buys or signs off on a system like this.
According to Information Age, an AI system began handling the bank's inbound customer inquiries in June 2025. On 29 July the bank confirmed to the ABC that it was cutting 45 call-center jobs. It had briefed the union the week before, which the ABC reported was "the first time a bank had informed the union its job cuts were due to AI." Against the 2,400 contact-center staff Information Age attributes to the bank, 45 roles is under 2 percent.
The bank's public explanation was general: "Our investment in technology, including AI, is making it easier and faster for customers to get help, especially in our call centres." The 2,000 figure entered the public record in two places. On 31 July Information Age attributed it to unnamed reports: "Reports suggest the AI chatbots have diverted 2,000 calls per week." On 7 August the union published it as the bank's claim. In the coverage I could read, I found no statement in which the bank put the number on the record itself, and no published evidence for it from anyone.
On 7 August the union said it had filed two disputes, one of them "escalated to the Fair Work Commission". On 21 August the bank reversed the cuts and told the affected staff they could stay, be redeployed or take the exit payment. The union's next hearing was listed for 25 August. The reversal came before any ruling, so the record is a dispute before the Fair Work Commission, not a decision by it.
The 7 August post lists what the union asked for and did not get. Beyond the evidence for the 2,000: "The bank is also refusing to provide its health and safety risk assessment, or even explain how roles were selected for redundancy."
What the union had instead was its members. The same post says: "To strengthen our cases, we need evidence from you." It offered a sentence to complete, "My job has been impacted by offshoring or AI because…", and the ABC later reported that the union "was continuing to collect accounts from members". By 21 August the accounts had a shape: "Call volumes were in fact increasing and CBA was scrambling to manage the situation by offering staff overtime and directing Team Leaders to answer calls." And: "CBA continually refused to be transparent about call volumes, so we recently took them to the Fair Work Commission."
It would be easy to say the union counted the calls. It didn't, and it couldn't have: members can't see a bot's logs or a queue's weekly totals. What they could see were symptoms. Google's site reliability engineering book names four signals to watch in any user-facing system (latency, traffic, errors and saturation) and defines the last as "How "full" your service is." Overtime offers and team leaders pulled onto the phones are what saturation looks like in a call center: the system is spending its reserve. Nobody needs access to telemetry to notice that. It is real evidence about the load the people were carrying.
Both sides' accounts can hold at once. As the union reported it, the claim was that "the voice bot has reduced call volumes by 2000 per week." That sentence can name two different quantities.
One is gross: the bot handled 2,000 calls a week that would otherwise have gone to a person. That is a counterfactual, and only the system's owner holds the records to compute it. The other is net: the calls reaching people fell by 2,000 a week. That is an observation, and anyone on the floor can check it against their shift. The union's 21 August post reads the claim in the net sense, a bot that "had led to a reduction in call volumes", and reports that members called it "an outright lie". On the net reading, the members' reports contradict the claim. On the gross reading, they neither confirm it nor refute it.
Now the bank's statement, given to Information Age on 21 August: "CBA's initial assessment that the 45 roles in our Customer Service Direct business were not required did not adequately consider all relevant business considerations and this error meant the roles were not redundant." And: "We have apologised to the employees concerned and acknowledge we should have been more thorough in our assessment of the roles required." The union's account of what management admitted is more specific: "Management has admitted they didn't properly consider that an increase in calls would continue over a number of months."
Read together, those are concessions about the net: total calls, over months. It says nothing against the gross. A bot can take 2,000 calls a week while the calls reaching people rise by more, from a seasonal peak, a fraud wave, a product change, or callers the bot failed who rang back. A count of what the bot took cannot tell those apart, and a roster is built from the net.
Made-up round numbers show the shape. A line takes 30,000 calls a week in the month before the bot arrives. Three months later demand is 32,500 a week, and the bot resolves 2,000 of them. People now handle 30,500 calls a week, 500 more than before the bot. The bot's count is true, the floor is busier than it was, and a roster cut to fit the count is short.
The count invites a second check, and it is simple division. Two thousand calls a week across 45 roles is about 44 calls per role per week, or about nine per working day. Is that what one of those roles handled? The answer depends on the handle time of the calls the bot absorbed, the hours the roles worked, and what else they did. Every one of those inputs sat with the bank, under the heading of the union's third request: how the roles were selected.
The bank's chief executive had described a different use for this kind of tool. At the 2024 annual general meeting, according to Information Age, Matt Comyn called a report that a chatbot would replace many of the 2,400 contact-center staff "perhaps… a little ahead of itself," and suggested chatbots might be most useful for surges in inquiries during a mass fraud or other "times of particular concern." A tool that absorbs surges trims the peaks. Cutting base roles needs the ordinary weekly load reaching people to fall and stay down, and by the union's account of management's admission, it did not.
On 13 August 2025, six days after the union's post and eight days before the apology, the bank released its full-year results. The front page of its ASX announcement puts two profit figures side by side: statutory net profit after tax of $10,133 million and cash net profit after tax of $10,252 million.
Cash profit is the bank's own construction. The Profit Announcement calls it "management's preferred measure of the Group's financial performance," built by excluding items the bank considers unrepresentative of its ongoing performance. The standard figure, prepared under the accounting standards, sits beside it. A footnote sends the reader to the "reconciliation between statutory and cash NPAT", where each excluded item is listed with its amount. The document adds that the excluded items are calculated consistently with prior periods and "do not discriminate between positive and negative adjustments."
That is how self-defined numbers are expected to travel in Australian markets. The corporate regulator, ASIC, set it out in Regulatory Guide 230 in 2011: the standard figures "should be given equal or greater prominence", and the company's own figures should "be explained and reconciled to the IFRS financial information", "be calculated consistently from period to period", and "be unbiased and not used to remove 'bad news'." The guide also asks for a label that is not "apt to cause confusion" with the standard figure. In the United States, Regulation G has required since 2003 that a registered company publishing a non-GAAP measure also present "the most directly comparable" standard measure and "a reconciliation (by schedule or other clearly understandable method)" of the differences, even when the number is made public "orally, telephonically".
None of this governs what a bank tells its staff about a roster, and I am not arguing that it should. The point is narrower. The remedy for a self-defined number is well known, and this bank applies it every half-year for its shareholders.
A deflection count is a self-defined number. The system's owner decides what counts as deflected and does the counting. A bare count has a line for every call the bot took and none for anything that came back: it discriminates, by construction, between positive and negative adjustments, which is the "bad news" problem the ASIC guide names. And "reduced call volumes" is a label apt to be confused with the standard figure, because the standard figure is call volumes: the calls offered to people, the calls they answered and the hours those calls took, week by week.
The reconciliation is a short table. Start with the calls people handled before the bot. Add the growth in underlying demand. Subtract the calls the bot resolved. Add back the repeat calls from people the bot did not resolve, who rang again and reached a person. The result is the load the roster has to carry, and it is the table a decision to cut 45 roles should have rested on. Had the bank put it in front of the union, the dispute would have been about the rows of a table rather than whose account to believe.
The case implies a minimum evidence standard for any claim that software has replaced a group of jobs. It has three parts.
The union asked for the second and third, and for the health and safety risk assessment. The first is the one the argument turned on.
For the teams who build and run these systems, the same standard turns into practice:
The reversal did not end the story. The union's site lists later rounds at the bank: 300 jobs cut in February 2026, and another 274 proposed in July 2026. I have not read what drove them and don't claim they involved AI. But the question in the 7 August post will be asked again, at this bank and others, whenever a count of what software handled is offered as the reason people are no longer needed. The answer that would settle it is a reconciliation to the calls that still reach a person. This bank knows how to produce one. It publishes one twice a year for its shareholders.
Part of Where the Number Came From, on how a published number is a fact about the way it was measured: Stanford says 12% to 66%, but 12% of what? · The safety score is a fact about the test rig · Tracing the 2026 AI-failure statistics to a primary
Sources:
Chain of Consciousness keeps a tamper-evident, dated record in hash-linked entries. Record what “deflected” meant when you counted it and each weekly reconciliation against the calls that still reached people, and a changed definition can no longer pass for a trend.
pip install chain-of-consciousness npm install chain-of-consciousness
Or start without installing anything: Hosted Chain of Consciousness.