76% of Australian firms hit an AI failure: what NZ should plan for

Australian results from BSI's AI failure study were reported on 6 November: 76% of Australian respondents had encountered an AI-related incident in the past year, and 88% of those firms reported more than one. BSI released the global results on 20 October, with 75% of businesses reporting at least one failure, so Australia sits almost exactly on the global line. We read the result as a sign that incidents have become a normal running cost of using AI, and any NZ adoption plan needs a budget line for them.

What the study found

BSI commissioned FocalData to survey 1,287 business leaders in nine countries in August 2026, from managers up to the C-suite. Australia was included and New Zealand wasn't. Alma Economics did the economic modelling.

The most common Australian consequence was ordinary lost time. Half of affected firms reported downtime or efficiency losses, against a global 44%. Regulatory or compliance consequences hit 35%, security breaches 31% and direct financial losses 24%. Failures involved internal workflow tools in 34% of cases, customer-facing systems in 21%, and both in 28%, which puts customers within reach of the failure in about half of cases.

The same respondents aren't backing away. In Australia, 81% reported tangible operational benefits, 72% a positive return and 77% plan to expand AI investment. Cost pressure is rising alongside those plans, with 72% seeing AI costs increase year on year and half saying AI spending pulled funding from other initiatives. BSI chief executive Susan Taylor Martin said: "AI failure is becoming a regular occurrence, bringing with it significant costs."

How much weight the numbers can carry

BSI has a commercial interest in the finding. The company's report page promotes its own AI training, ISO/IEC 42001 certification and assessment services. The results are also self-reported, and the public summary doesn't define a failure, so a logged incident could be anything from a poor chatbot reply to a data breach.

BSI's release and the Australian coverage don't state the Australian sample size. An even split of 1,287 across nine markets would give about 143 respondents per country, so we'd treat gaps of a few points between Australia and the global figures as statistical noise.

The headline revenue estimate needs similar care. BSI's modelling puts up to 3.6% of annual market revenue at risk, excluding regulatory fines, and aims for an order of magnitude. That figure describes exposure across a market. Multiplying your own turnover by 3.6% won't give you a forecast.

Why NZ firms should pay attention

The Prospa SME tracker shows NZ small businesses moving quickly into the group BSI is measuring. IT Brief reported that 59% of NZ SMEs now use AI in some form, up from 50% six months earlier, in a survey of 501 decision-makers at firms with fewer than 50 full-time staff. Only 7% have AI integrated across several functions.

BSI's global data shows failures are more common in larger businesses, at 83% against 57% of SMEs. We expect most NZ SMEs sit nearer the lower figure today because their use is light. We also expect their exposure to rise as AI moves from drafting help into quoting, invoicing and customer replies. Prospa found uncertainty about return was the top barrier for non-users, at 34%. A business case that leaves out rework and failure handling will overstate that return.

Detection is the weak point. BSI's consulting estimate draws on an Australian case involving a government-commissioned report with fabricated citations and a refund of 22.3% of the fee. The coverage doesn't name the firm, but the details match Deloitte's report for Australia's employment department, where a University of Sydney academic flagged the errors after publication. A firm that relies on the reader to find the mistake has already paid for the failure in reputation.

The governance figures suggest many businesses would struggle to see a failure early. Globally, only 47% have a process for managing AI incidents. In Australia, 38% know which data sources their AI tools rely on, and 37% restrict unauthorised tools. A failure in an unapproved tool is unlikely to reach anyone who'd log it, so the 76% figure may understate what's happening.

Agents raise the stakes. About 88% of Australian respondents are exploring, piloting or deploying agentic AI, and BSI's model gives agents the highest exposure, at 3.58% of the market each year. Where an AI failure exposes personal information in NZ, the Privacy Act test also applies. A breach that has caused or is likely to cause serious harm must be notified to the Privacy Commissioner and affected people as soon as you're practically able. Meeting that timeframe is hard if nobody knows who owns an AI incident.

Practical next steps

Most of this work fits into normal AI adoption planning and doesn't need a separate programme. We'd start here:

• List every AI tool in use. Include AI features switched on inside existing software, then name an owner for each tool and record the data it reads.
• Define a failure for each use. Examples include a wrong figure in a quote, a customer reply nobody approved, or an agent run that spends more than expected.
• Route AI failures through your service desk. Add a category and severity levels to your existing IT service management process, and log near misses as well as outages.
• Check output before customers see it. Set human sign-off for reports, advice and quotes, and sample routine output on a fixed schedule.
• Test the manual fallback. Downtime was the most common Australian consequence, so switch a tool off for a day and see which work stops.
• Give staff an approved option. Pair any rule against unapproved tools with a sanctioned tool that does the same job, or the use will carry on out of sight.
• Add failure costs to the business case. Allow for rework, staff time and rising usage charges, then compare the estimate with your incident log after a quarter.
• Set agent limits before go-live. Cap spending and permissions, and test how you'd stop an agent mid-task.

BSI's study has no NZ sample, and we don't know of a comparable NZ failure measure. For now, a firm's own incident log will tell its leaders more than any regional survey, and even a simple spreadsheet started this month will show which AI uses are costing more than they return.