By Brad Ferris · 22 August 2026
On 19 August, iTnews reported that Metcash is exploring agentic AI to automate retail ordering: agents that would place routine stock orders on behalf of the independent retailers the A$19.6 billion wholesaler supplies, building on the conversational search it already runs in its Sorted marketplace. The reason given is not labour cost. It is customer loyalty. If the agent that reorders a store's stock sits inside Metcash's systems and knows that store's ordering history, the retailer has one less reason to price the order elsewhere on a Tuesday morning.
That is worth sitting with, because it inverts the way most Australian businesses are currently scoping AI. The default brief points the technology inward, at your own staff and your own process, and the business case is written in hours saved. Metcash is pointing it outward, at the customer's process, and the business case is written in switching costs.
Metcash is not the only Australian example from the past fortnight.
CBA's in-app AI Companion, launched in limited testing in late May, produced a finding its executives clearly did not expect. Michael Baumann told iTnews that roughly half the questions put to the assistant concern spending and saving guidance that branch and call centre staff do not typically handle. The bank built a service channel and discovered a demand signal. Customers were carrying a question they had never had a cheap way to ask, and the assistant made asking cheap enough that they asked it.
Pro Medicus took the same posture and put a number on it. The medical imaging software company reported FY26 net profit up 130 per cent to $265.3 million, alongside $407 million in new long-term contracts across nine US health networks and one in Europe, with chief executive Sam Hupert positioning its Visage platform as the workflow that clinical AI has to run through rather than a product AI would eventually replace. Owning the pipe beat owning the model.
Three different sectors, one shape. The value did not come from removing internal effort. It came from being embedded further into how the customer works.
There is a second reason to take the outward brief seriously this quarter, and it is unsentimental.
Gartner, reported by ARN on 18 August, projects that inference cost per agentic workflow will rise more than fivefold by 2028. Falling token prices are being swamped by the complexity of multi-step agent work, and agentic reasoning already costs at least five times a basic chatbot interaction. Gartner calls it the inference paradox: the unit price of intelligence keeps dropping while the bill keeps climbing, because each task now consumes far more of it.
Apply that to an internal efficiency case. If the return is a fixed number of hours saved in accounts payable, and the cost of running the agent that saves them rises through the contract term, the margin on that project narrows every year you operate it. The case does not fail. It just gets thinner, and it gets thinner in the years when the board expects it to be improving.
Now apply it to a retention case. A retailer who has let an agent learn their ordering pattern for eighteen months does not casually move. The value of that position compounds while the cost curve rises, rather than being consumed by it. Both projects face the same inference bill. Only one of them is building something that grows.
The outward brief is harder, and that is most of why it is less common.
Internal automation can be scoped in a room with your own people. Customer-facing agents cannot. They need clean data about what the customer is doing, a channel the customer already uses, and a real tolerance for what happens when the agent is wrong in front of someone who pays you. That last one is the actual constraint. Most businesses have not decided what an agent is allowed to do on a customer's behalf without a person checking, and that decision is a commercial one, not a technical one.
It also needs a level of commitment that is easy to underestimate. IAG, disclosing its FY27 plans this month, put roughly $200 million against AI, with more than 600 internal activators and 90-plus agents published. Very few mid-market operators will spend at that scale, and they do not need to. But the shape of that investment, spread across many people close to real workflows rather than concentrated in one central team, is the part worth copying.
Take whatever AI work you have queued for the next two quarters and sort it into two piles: work that makes your team faster, and work that makes your customer's job easier.
If the second pile is empty, the strategy is a cost programme. That is a legitimate thing to run, and for some businesses this year it is the right one. But it will not produce the compounding position Metcash is reaching for, and it will be quietly eroded by the same cost curve everyone else is facing.
If the second pile has something in it, the next question is the honest one: does anyone outside your business know it exists, and would they notice if you switched it off?
Where does your business stand? The free AI Scorecard takes three minutes and shows you. If you want a straight steer from a person, book an AI Opportunity Call.