By 30 June 2026, every non-corporate Commonwealth entity in Australia had to have a Chief AI Officer. Not a working group, not a centre of excellence. A named senior leader.
That deadline did something useful. It forced a few hundred organisations to answer a question most private companies still dodge: who is actually accountable for AI here? In my experience the honest answer inside most businesses is either “nobody” or “the CIO, by default, on top of everything else.”
Neither works for long.
What the Australian Public Service model gets right
The APS design has one feature worth copying immediately: it splits the role in two.
The AI Accountable Official owns risk and governance. Compliance, assurance, the assessment gates, what gets stopped. The Chief AI Officer owns adoption. Where AI gets applied, how staff build capability, which opportunities get funded, how the organisation keeps pace. GovAI’s own description is explicit that the CAIO focuses “less on technical detail and more on opportunity, transformation and system-wide coordination.”
That split matters because the two jobs pull in opposite directions and collapsing them into one person produces a predictable failure. Give the role to a risk-minded leader and everything routes through an approval queue until the business stops asking. Give it to an enthusiast with no assurance counterweight and you get the Deloitte outcome: confident output, no verification, a public correction.
Most Australian businesses do not need two executives. They do need two named responsibilities, and they should not sit with the same person on a large AI footprint.
How is a CAIO different from a CIO?
Here is the distinction I use with boards.
| CIO / CTO | Chief AI Officer | |
|---|---|---|
| System behaviour | Deterministic. Same input, same output. | Probabilistic. Output quality varies and degrades. |
| Value source | Installing and running capability reliably | Redesigning the process the capability sits in |
| Main constraint | Budget, integration, technical debt | Trust, process clarity, change capacity |
| Failure mode | Outage, breach, project overrun | Silent wrongness that nobody checks |
| Success measure | Availability, delivery, unit cost | Cycle time, cost per outcome, adoption depth |
The row that catches people out is the failure mode. A CIO’s disasters announce themselves. An AI programme’s worst failure is a system that quietly produces plausible, wrong output for six months while everyone reports green. Detecting that requires a different instinct and a different measurement regime, which is the real argument for a distinct role rather than a distinct title.
There is a second argument, and it is the one I find more persuasive. Almost all AI value in Australian business right now is trapped by process, not by technology. National AI Centre data has SME adoption at 43 to 44% in the December 2025 to February 2026 quarter, and 82% of large enterprises running at least one AI system. Yet only 7% of businesses have embedded AI into their actual products or services. Ninety-three percent are using it for internal productivity. That gap is not a model problem. I have written about why that 7% figure is Australia’s real AI problem, and the short version is that nobody owns the redesign work. The CAIO is the person who owns it.
Does your business need one?
Four triggers. Any two of them and I would start the conversation.
Scale. More than a handful of AI systems in production, or AI spend crossing roughly $250,000 a year. Below that you have tools. Above it you have an estate, and estates need an owner.
Consequence. AI touching customer-facing decisions, pricing, credit, hiring, safety, or anything covered by the Privacy Act’s automated decision provisions. Consequence raises the cost of getting governance wrong faster than scale does.
Duplication. Two or more teams independently building the same capability. This is the cheapest signal to detect and the most reliable indicator that coordination is missing. If marketing, finance and operations have each stood up their own agent framework, you have already paid for the role three times over.
Procurement pressure. If you sell into Commonwealth or state government, expect to be asked who holds AI accountability in your organisation. The APS obligations flow downstream through contracts.
If none of these apply, do not create the role. A part-time or fractional arrangement covers the governance and prioritisation work for a fraction of the cost, which is the case I made in when a fractional CAIO makes sense.
What should the role actually own?
Job descriptions for this role are usually terrible. They read like a list of technologies. Here is what the mandate should contain.
A process inventory, not a tool inventory. The CAIO’s foundational asset is a map of how work actually flows through the business, with volume, cycle time and cost per unit against each step. Without it, every AI decision is a guess about where value sits. This is unglamorous and it is the thing nobody does. I have argued before that the process inventory is the moat nobody maps.
A funded portfolio with kill criteria. Every AI initiative gets a hypothesis, a measurement, and a date at which it either scales or stops. The absence of kill criteria is why 95% of pilots stall in pilot.
Capability building. Not a vendor training day. A plan for who in the organisation needs which skill at what depth, and how that gets built given Australia’s AI specialist shortfall of up to 60,000 people by 2027. Buying that talent on the open market at $130,000 to $170,000 for an AI engineer is possible but not scalable for most businesses.
Cost accountability. Token spend, per-agent unit economics, and the link from consumption to outcome. Ninety-eight percent of FinOps practitioners now manage AI spend, up from 31% two years ago, and 73% of agentic projects run over budget. Somebody senior needs to own that number before finance discovers it.
The assurance relationship. Not owning assurance, but being answerable to it. The CAIO proposes, the accountable official gates, the board sees both.
What to measure
Refuse the vanity metrics. Number of AI tools deployed, staff with a licence, prompts run per week: all of these move without value moving.
Measure four things:
- Cycle time on the specific processes AI touched, before and after.
- Cost per completed unit of work in those processes, with AI spend included.
- Share of AI spend attributable to a named business outcome. Industry benchmark is that only 39% of organisations can attribute any measurable impact at all, so this is a low bar to clear and a revealing one.
- Governance currency. Proportion of production AI systems with a current impact assessment and a working decision log.
If the first two do not move within two quarters, the portfolio is wrong. If the fourth does not hold above 90%, the estate is drifting and you will find out the hard way.
The uncomfortable part
The Chief AI Officer title has become a way for organisations to look like they are taking AI seriously without changing anything. I have seen the role handed to someone with no budget, no process authority, and a dotted line to three executives who outrank them. That version of the job cannot succeed and everyone involved knows it within a quarter.
The role works when it comes with three things: authority to redesign a process, a budget that does not require begging, and a direct line to the person who can stop a project. Without all three, appoint nobody and be honest that AI is a set of tools you have not decided to take seriously yet. That is a defensible position. Pretending otherwise is not.