AI Customer Support Stumbles: Commonwealth Bank Reverses Layoffs as Calls Surge

AI layoffs backfire: CBA's chatbot plan ends in a u-turn

Australia's largest bank rolled out a customer service bot, cut 45 roles, then reversed the decision within weeks as calls went up, not down. The Commonwealth Bank of Australia (CBA) learned a simple truth: AI needs time, oversight, and human backup - especially in customer support.

If you lead a support team, the signal is clear. Don't cut people before the system proves itself in live traffic, for weeks, with hard metrics.

What happened at CBA

CBA launched an AI chatbot, reportedly called Bumblebee, to handle simple queries and free staff for complex issues. After claiming a reduction of 2,000 calls a week, the bank announced 45 customer service redundancies.

The Finance Sector Union (FSU) disputed the figures and took a complaint to the Fair Work Commission. Members reported call volumes were actually rising, with overtime offered and team leaders pulled onto phones.

Weeks later, CBA canceled the redundancies and apologized to the affected staff. One long-serving employee, Kathryn Sullivan, who helped test the bot, said she had "inadvertently" trained the system that put her job at risk and called for stronger protections.

Despite the backflip, CBA is pressing ahead with a multiyear partnership with OpenAI to improve scam prevention and personalize services. The bank says it will invest in staff AI proficiency to better support customers.

The wider trend: companies are rethinking AI-first support

Klarna claimed its AI could do the work of about 700 customer agents, then shifted course. The CEO now plans to hire more agents and offer a choice: AI for basic issues, humans for a premium experience.

Research from May 2025 shows half of companies plan to drop efforts to replace support workers with AI by 2027. A third of UK tech leaders admit they cut staff for AI and now regret it; 55% say they moved too fast and created talent gaps.

Why AI rollouts trigger more calls (and stress) at first

• False deflection: bots handle easy intent, but misroute edge cases, creating repeat contacts.
• Measurement errors: "reduced calls" on paper, while live queues show rising volume and longer handle time.
• Failure demand: customers call back after poor bot outcomes, increasing total contacts.
• Change shock: new flows, new knowledge, and new tooling slow agents until processes settle.

Playbook: adopt AI without cutting support too soon

• Set a freeze: no staff cuts for 90-180 days after go-live. Review only after stable metrics for 8 straight weeks.
• Start internal-first: use AI for agent assist, summaries, knowledge surfacing, and suggested replies before front-door automation.
• Gate by outcomes: require minimum thresholds for containment quality, first contact resolution, repeat-contact rate, and CSAT before any restructuring.
• Keep human choice: always offer a human path. Consider tiered service like Klarna's plan - bot for simple, human for premium.
• Staff surge plans: pre-approve overtime, cross-train, and keep a flex pool ready for spikes.
• Handoff standards: fast escalation from bot to agent, with full context, no dead ends.
• Hard QA: weekly audits of bot transcripts; label failure demand; ship fixes fast.
• Transparent change: involve reps and unions early; document impacts; align with regulators where relevant.
• Upskill your team: conversation design, prompt writing, AI QA, and data hygiene. Train before you automate.

Metrics that matter (track weekly)

• Containment with quality: issues fully solved by bot without repeat contact.
• Repeat-contact rate within 7 days (by intent and channel).
• Handoff health: escalation rate, wait time to human, and context completeness.
• Agent impact: AHT, occupancy, shrinkage, and burnout indicators.
• Customer effort and CSAT post-bot and post-agent.
• Operational risk: unresolved queues, abandonment, and complaint rates.

Protect jobs while you modernize

People carry your customer trust. Use AI to remove toil, not your bench of experience. Tie role changes to proven results, not vendor pitch decks.

If you're upskilling support teams for AI work - conversation design, prompt quality, and AI-assisted workflows - explore focused training paths here: AI courses by job.

Bottom line

AI can make support faster and more consistent, but early rollouts are messy. CBA's reversal shows the cost of cutting before the data is real. Keep humans in the loop, prove the metrics, then decide what to automate - and what should stay human.