Australia's biggest bank cut 45 customer service roles on the strength of a claim that its new AI voice bot was reducing call volumes. The people answering the phones said the opposite. Weeks later, the bank apologized and reversed the decision.
In July 2025, Commonwealth Bank of Australia told 45 people in its Customer Service Direct team that their roles were no longer needed. The reason given was technology. The bank had introduced an AI "voice bot" to take simple inbound calls, and it said the bot was reducing the number of calls that needed a person. Fewer calls reaching people, it followed, meant fewer people.
The people still answering the phones saw something different. According to the Finance Sector Union, call volumes went up after the bot arrived, not down. Managers were offering overtime to keep up. Team leaders were being pulled onto the phones. The union took a dispute to Australia's Fair Work Commission.
On August 21, 2025, the bank reversed the decision. In a statement reported by the Bendigo Advertiser, it said its initial assessment "did not adequately consider all relevant business considerations and this error meant the roles were not redundant." It added: "We have apologised to the employees concerned and acknowledge we should have been more thorough in our assessment of the roles required." The 45 staff were offered a choice: stay in their roles, move to another job inside the bank, or leave.
Notice what the bank did not say. It did not say the voice bot was broken. It said the assessment was wrong. A claim about call volumes became a decision about people, and the claim did not survive contact with the people who could see the call queue.
That is the part worth studying, because it has nothing to do with banking. A new system goes in. A number comes out that says the system is working. The number is convenient, it fits the plan, and nobody puts it next to the raw figures and asks whether it is true. The people who finally checked were not analysts. They were the staff working the overtime.
Every business has a version of this number. "Conversions are up since the redesign." "The new campaign is driving most of our sales." "Organic traffic is flat, so SEO is fine." When the number comes from an AI, it arrives faster, sounds more certain, and is just as easy to leave unchecked.
AI tools that answer questions about your data are getting good at sounding right. That is exactly the problem. A fluent paragraph about your traffic looks the same whether the figures in it are correct, out of date, measured over the wrong period, or made up.
Four questions catch most of the damage before it reaches a decision.
The bank's mistake cost it a public apology and a Fair Work dispute. A wrong number acted on in a marketing budget costs money every day until someone notices, and often nobody does.
In Askalytics, every answer names its sources and its date window, every figure is checked against your own data, and every AI Insight gets a second opinion.
This is why Askalytics does not let an answer stand on its own. Every answer names the sources it read and states the exact date window it was measured over. Every number in an answer is checked against your own data before you see it, and if a figure cannot be traced back, the answer says so. Every figure in an AI Insight links to the query that produced it, so "can I see the working" is always yes. And every AI Insight gets a second opinion from a second AI on a different provider, the Confidence Check. If the two disagree, you see it.
There is a quieter lesson in the bank's story too. The problem was found by people who happened to be watching. AI Sentinel watches your traffic, conversions, search and ad data every night, so a metric that moves when nobody is looking still gets noticed.
The bank's mistake was not using AI. It was acting on a claim about AI without checking it against the numbers underneath. The fix is not more confidence from the machine. It is the habit of asking where the number came from, and tools that make that question easy to answer. Read how we keep AI honest, or see what a checked answer looks like.
The bank did not say the bot failed. It said its assessment of the roles was wrong: it "did not adequately consider all relevant business considerations." According to the Finance Sector Union, call volumes rose after the bot was introduced, which is the opposite of what the decision assumed.
Four things: which data it came from, what date window it covers, what it is being compared with, and whether you can see the query or calculation behind it. If any of those is unknown, the number is a claim, not a measurement.
Every answer names its sources and states its date window. Every figure is checked against your own data before you see it, and anything that cannot be traced back is flagged on the answer. AI Insights, and Ask AI answers that use a credit, also get a second opinion from a second AI on a different provider. An answer can still be wrong, but you can see how it was checked.