Most AI business cases signed off in the past three years were not efficiency programmes. They were risk transfers, and in most cases the board did not know it was approving one. Headcount came off the payroll line and was booked as a saving. The judgement those people exercised, the checking, the catching and the quiet correction of other people’s mistakes, did not disappear. It moved onto the organisation’s balance sheet as unpriced liability.
Meta has now shown what that looks like in practice. As The Guardian reports, citing Reuters, the company has scaled back plans to cut staff and expand its use of AI agents after finding employees were spending more and more time cleaning up technical and security errors. Having stripped out middle managers in its 2023 efficiency push, it is now expanding its managerial ranks again. Meta is one of the best-resourced technology companies on earth. If it miscalculated the cost of removing human review, the average ASX 200 company almost certainly has too.
Review capacity was the control
The flaw in most AI business cases is treating human review as overhead. It was never overhead. It was a control, and in many organisations it was the most effective control they had. Middle managers, senior analysts and experienced reviewers were where errors were caught before they reached a customer, a regulator or a production system.
Remove that layer and the error rate does not fall. Detection moves downstream, and every step downstream makes remediation more expensive. A flawed paragraph caught by a reviewer costs minutes. The same flaw caught by a client costs a relationship. Caught by a regulator, it costs a notification, an investigation and possibly a penalty. AI makes this worse because it produces errors at volume and with confidence. Output that looks finished gets less scrutiny than a junior employee’s draft, not more.
Meta’s reversal is the predictable result. The organisation is buying back the control it sold, and paying more for it, because it now needs people who can untangle errors that have already compounded, not just people who prevent them.
Security errors are the expensive kind
The Guardian’s account specifically names security errors. Boards should note that, because security errors have a different cost profile from quality errors.
A poorly drafted report is embarrassing. AI-generated code that introduces a vulnerability, or an agent granted broad system permissions that misconfigures access, creates exposure that sits silently until someone exploits it. In Australia the consequences are no longer discretionary. A breach involving personal information triggers the Notifiable Data Breaches scheme. APRA-regulated entities answer under CPS 234 for the information security capability of their environment, and under CPS 230 for operational risk management, including how they manage material service providers. Privacy Act amendments passed in 2024 require organisations to disclose in their privacy policies when they use personal information in substantially automated decisions that significantly affect individuals.
None of these frameworks accepts “the model got it wrong” as a defence. Your organisation made the decision to deploy the model, and your organisation owns the output.
Liability never left the building
Every enterprise AI contract we have reviewed puts the risk of incorrect output squarely on the customer. Vendors warrant that the service will be available. They do not warrant that it will be right. Courts and tribunals have taken the same view. Air Canada was held liable in 2024 for its chatbot’s incorrect advice to a customer, after arguing, unsuccessfully, that the chatbot was responsible for its own statements. Closer to home, Deloitte Australia partially refunded the federal government in 2025 after a report it delivered was found to contain AI-generated errors, including fabricated references.
The pattern is consistent. When AI output is wrong, the organisation that deployed it pays. If the business case assumed otherwise, or never asked the question, the projected savings were overstated from day one.
Why the risk column was left blank
We have sat in the rooms where these business cases were approved. The risk section was not vague by accident. The savings were certain, immediate and attributable to whoever sponsored the programme. The costs were uncertain, deferred and spread across functions that had no seat in the approval meeting: security, legal, customer operations, and whoever inherited the rework.
Nobody had an incentive to quantify the downside. Vendors certainly did not. Consultants paid on transformation outcomes did not. Executives whose bonuses were tied to cost reduction did not. And boards, under pressure not to fall behind on AI, did not push hard enough to make them.
This is a governance failure, not a technology failure. Audit committees approved a reduction in control capacity without seeing it described that way.
What boards and CFOs should do now
Reopen the business cases. Take every AI programme approved on a headcount or efficiency basis and ask what has happened to error rates, rework volumes and incident numbers in the affected functions since rollout. If nobody is measuring, that is your first finding.
Name the owner of wrong output. For each material AI deployment, a named executive should be accountable when the output is wrong. If the answer is “the vendor” or “the team that uses it”, the accountability gap is already on your risk register whether or not anyone has written it down.
Price the review layer honestly. Real AI ROI is the saving minus the cost of the human judgement still required to make the output safe to rely on. For some use cases that number remains strongly positive. For others it is negative, and the organisation has been subsidising the shortfall through unbudgeted rework and unrecognised exposure.
Treat review capacity as a control. Before approving further workforce reductions linked to AI, require management to show which controls those roles performed and what replaces them. This is the same discipline you would apply to removing a firewall or an approval step in a payments process.
Meta has the margin to reverse course quietly. Most organisations will discover their version of this problem through a breach notification, a regulator’s letter or a client dispute. Boards that ask these questions now will be correcting a business case. Boards that wait will be explaining one.