By Brad Ferris · 29 August 2026
On 27 August, iTnews reported that Woolworths had lifted the covers on its internal AI work at its full year results. Chief executive Amanda Bardwell said an employee support tool called Team Assist fielded 7,000 staff enquiries a week across the financial year, with the need to escalate to a human removed in nine cases out of ten.
What makes that worth reading twice is how small it is. There was no transformation narrative attached to it, no claim about how much of the workforce is now AI-enabled, no headcount reduction. One tool, one queue, one rate.
Team Assist can be measured because the queue was there before the tool was. Employee enquiries arrive, get logged, get resolved or escalated. That gives you a denominator, 7,000 a week, and a numerator, the one in ten that still needs a person. The measurement falls out of the workflow rather than being commissioned afterwards as a study.
The second tool Bardwell disclosed has the same property. Woolworths said an AI marketing tool cut production of the weekly shopper catalogue from a week to a few hours. Elapsed time is a crude measure, but it is a measure, and it was being tracked before anyone introduced a model to the process.
The two tools still in trial follow the pattern. Woolworths' head of SAP, Murtuza Imdadali, described a training chatbot that answers questions about the company's promotions app so staff stop logging support calls when they get stuck on a screen, and a second system that checks whether product dimensions in the inventory record are wrong. That second one is the more instructive of the pair. Imdadali said it returns a confidence rating of low, medium or high, and the team is still working out whether a high-confidence flag should be auto-corrected or sent to a person as a workflow item. That is a bounded task, an existing exception queue, and an open question about where the human sits. It is a design conversation, not a vision statement.
Two days before the Woolworths story, McKinsey published its State of AI global survey for 2026. Roughly four in ten respondents, 37 per cent, report that AI has contributed positively to their organisation's EBIT, essentially unchanged on the year before. The Register reported that the survey ran across 1,719 respondents and that only 6 per cent qualify as high performers, meaning they attribute at least 5 per cent of EBIT to AI.
The flatness is the finding. Adoption climbed over the year. Attribution did not move. Two years into broad deployment, most organisations still cannot connect the spend to the earnings line.
The usual explanation is that the technology underdelivers. A simpler one fits the data better. A great deal of what has been deployed sits across work that was never counted. A licence rolled out to 400 people, used unevenly, on tasks with no recorded volume, no cycle time and no error rate, produces no evidence either way. The tool may well be working. Nobody can say by how much, because there was no before.
It is worth being precise about the limits. Woolworths gave no cost figure and no return figure for any of these tools. A deflection rate is an operational measure, not a financial one. Ninety per cent of 7,000 enquiries a week is work that did not reach a support team, and whether that converts into money depends entirely on what that team did with the time instead.
So it is not a full business case. It is still more disclosure than most Australian companies offered this reporting season, and it is specific enough that someone could argue with it, which is the test.
None of this depends on Woolworths' scale. The mechanics are the same in a business with one shared inbox.
Pick the queue you already have a record of. The support address, the accounts inbox, the questions your operations manager answers twice a day, the exceptions your warehouse supervisor fixes by hand. Then count it for a fortnight before you build anything: how many items, in what categories, how many needed real judgement and how many were the same question with a documented answer.
That fortnight is the whole exercise. It gives you the denominator, and it tells you which slice is worth automating first, which is almost always the repeat question that already has a written answer somewhere nobody can find. Route everything else to a person by default and leave it there until the evidence says otherwise.
The reason to work in that order is not process discipline for its own sake. It is that without the before number, a tool that is genuinely working looks identical to one that is not, and the first thing to go in a tight budget round is the line item nobody can defend.
Woolworths published one rate, on one queue, for one year. Most operators could produce the equivalent inside a quarter. The 37 per cent will move when more of what gets deployed looks like Team Assist, and less of it looks like a licence rollout with a launch email.
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.