Doing this with AI makes you 19% less likely to be right
Your Essential Report from the Frontlines of Applied AI
Westpac, one of Australia’s biggest banks, has halved the cost of a major IT project with AI.
Its team used AI tools to speed up code conversion, testing and documentation while moving an old data warehouse. The work finished more than two years early, at about half its A$70 million budget, the bank said in its annual results this week.
Lawyers in Victoria, Australia, tell a different story. The Victorian Legal Services Board and Commissioner surveyed 2,221 of the state’s lawyers, and 55% now use AI, up from 37% a year earlier.
But only 26% say AI has cut their clients’ fees. The lead researcher says the time it saved was less than expected, and that checking AI’s work may have used up the gains.
So which of your tasks are like Westpac’s, and which are like the lawyers’?
It’s clear to us that:
- AI is brilliant at some tasks and worse than no help at others.
- How hard a task looks is a poor guide to which is which.
- The biggest gains go to the people who know where that line sits in their own work.
Once you know where the line is, you can lean on AI hard wherever it helps. And you stop paying for the checking on the tasks where it doesn’t.
So this week is about finding that line. There’s a prompt below from AI+ Management Consultants. It asks AI for two genuinely different explanations of your evidence, which is one of the tasks AI does surprisingly well. And there’s a case study on the experiment with 758 Boston Consulting Group consultants that showed where AI helped them, and where it made them worse.
Plus the week’s news in applied AI — Japan gives about 180,000 civil servants an AI app to help draft answers for parliament, New Zealand Police investigators get AI interview transcripts they must check, more than 4,000 SK Telecom staff have built their own AI agents, and Aeon puts about 90 AI robots to work in a warehouse.
You can’t tell where AI helps by how hard the task looks. Learn it task by task.
AI’s ability doesn’t follow how difficult a task looks to you. It can excel at work a skilled professional finds genuinely hard, and get things wrong on work that looks routine.
We give consultants two examples. AI is surprisingly good at coming up with other explanations for the same evidence, because it isn’t anchored to your first instinct. But it can quietly get a detail wrong when it summarises one client’s specific numbers. A summary that sounds standard is no guarantee the figures are right.
So treat your sense of where AI helps as a first guess. Each time it surprises you, in either direction, note it.
That’s how you learn which work to hand over, and which to keep.
- Government
Japan gives about 180,000 civil servants an AI app to help draft answers to questions in parliament
Japan’s Digital Agency began offering the app to ministries on 30 October. It helps with the whole job, from checking a lawmaker’s question to drafting a proposed answer.
Nikkei, 30 October 2026
- Policing
New Zealand Police investigators get AI transcripts of their interviews, and must check them before acting
From 2 November, Axon’s Auto-Transcribe turns recorded interviews into text for investigators and prosecutions staff. Axon says it averages 80% accuracy. Police say investigators must confirm what a transcript says before acting on it, and transcripts for court will still be prepared and verified by people.
RNZ, 2 November 2026
- Telecoms
More than 4,000 SK Telecom staff have built their own AI agents, and 1,500 are in regular use
The Korean operator now plans to redesign how whole teams work around AI. In a first test this year, one employee planned, built and checked a service working alongside AI agents. The figures are the company’s own.
SK Telecom, 1 November 2026
- Logistics
Aeon puts about 90 AI robots to work unloading deliveries and stacking outgoing pallets at a warehouse
The robots went into full operation at the retailer’s Kawaguchi hub, north of Tokyo, on 29 October. They recognise products that were never registered in the system and stack them by size and weight. Short of staff, and with wage costs rising, Aeon aims in future to automate 80% of its loading and unloading work.
ABEMA TIMES, 29 October 2026
Leave out client names and anything confidential. Paste your first explanation and the evidence behind it, and change “management consultant” to your own job. The letters in square brackets are the book’s labels; you can leave them out. AI is good at this task because it isn’t anchored to your first instinct. Deciding which explanation the evidence actually supports stays with you.
You are a management consultant [P]. Based on this initial hypothesis [paste hypothesis] and supporting evidence [paste evidence], generate two genuinely different alternative hypotheses that could also explain the evidence [R, S]. Make them meaningfully different, not variations of the same idea [I].
Prompt 3.7 from AI+ Management Consultants, word for word.
Find out where AI helps before you rely on it.
Researchers from Harvard Business School and other universities worked with Boston Consulting Group to test AI on 758 of its consultants.
On 18 realistic consulting tasks, the consultants using AI completed 12.2% more tasks, 25.1% more quickly, and their work was of significantly higher quality.
The researchers also set one task that looked no harder, but sat outside what AI could do well. On that task, consultants using AI were 19% less likely to reach the correct answer than those working without it.
The researchers call this the “jagged frontier”: AI makes you better at some tasks and worse at others, even when they look equally difficult.
Before you lean on AI for a task, try it on a few cases where you already know the right answer.