The March of AI

Yesterday, today, tomorrow and beyond

Tim Barnes

Two numbers, both true

900 million people use ChatGPT every week ChatGPT weekly active users, 2022 to 2026 Weekly active users climb from one million in the five days after launch in November 2022 to 100 million in November 2023, 200 million in August 2024, 300 million in December 2024, 400 million in February 2025, 700 million in August 2025, 800 million in October 2025 and 900 million in February 2026. A dashed segment continues towards one billion. 1 billion 500 million 2023 2024 2025 2026 100M Nov 2023 400M Feb 2025 900M Feb 2026
95% of enterprise generative AI pilots show no measurable profit impact Ninety five of one hundred pilots, and the five that take root A block of one hundred dots. Ninety five turn red and burst. The remaining five turn green, fall to the ground and grow into small plants.
The shape of this talk Five stages in order: Yesterday, Today, Tomorrow, Beyond, then Your question. Yesterday Today Tomorrow Beyond Your question

“No measurable profit impact” is not the same as “failed”. Many pilots had no baseline to measure against.

OpenAI, February 2026; MIT Project NANDA, The GenAI Divide, 2025. Chart figures as stated by OpenAI; the dashed segment is The Information, July 2026, unconfirmed.

75 years in ten moments

Seventy five years of artificial intelligence A timeline whose events are sparse on the left and crowded on the right, so the pace of change is visible. 1950, Turing proposes the imitation game. 1956, the Dartmouth workshop names the field artificial intelligence. 1974 to 1993, two AI winters, shown as a dip in the line. 1997, Deep Blue beats Kasparov. 2012, AlexNet and the deep learning breakthrough. 2016, AlphaGo beats Lee Sedol four games to one. 2017, Google's Transformer paper, the T in GPT. 2022, ChatGPT reaches 100 million users in about two months. 2024, two Nobel Prizes for AI work in physics and chemistry. 2025 to 2026, reasoning models and agents. 1974 to 1993 two AI winters 1950 Turing proposes the “imitation game” 1956 Dartmouth workshop names the field “artificial intelligence” 1997 Deep Blue beats Kasparov 2012 AlexNet, deep learning breaks through 2016 AlphaGo beats Lee Sedol 4 to 1 2017 Transformer paper, the “T” in GPT 2022 ChatGPT: 100 million users in about two months 2024 Two Nobel Prizes for AI, physics and chemistry 2025 to 2026 Reasoning models and agents

Stanford AI Index 2025; Productivity Commission interim report, August 2025; Nobel Prize announcements, 2024.

From rules to agents

Four steps from rules to agents An ascending staircase of four steps. Step one, rules and robotic process automation, follows fixed instructions, as Deep Blue did in 1997. Step two, machine learning, learns patterns from data, as AlphaGo did in 2016. Step three, generative AI, creates content from plain-English prompts, as ChatGPT did in 2022. Step four, agentic AI, plans and acts across steps using tools, as the Commonwealth Bank agentic fraud system does in 2026. 1 Rules and RPA Follow fixed instructions. Deep Blue, 1997. Grandmaster rules, 200 million positions a second, no learning. 2 Machine learning Learn patterns from data. AlphaGo, 2016. Taught itself from millions of games. 3 Generative AI Create content from plain-English prompts. ChatGPT, 2022. 4 Agentic AI Plan and act across steps using tools. CBA’s agentic fraud system, built in three months, 2026.

IBM; DeepMind; Commonwealth Bank newsroom, 2026.

What AI is good at, and where it isn’t

Writing

40% faster, 18% higher quality

MIT, published in Science, 2023

Customer service

14% more productive, with the biggest gains for newer staff

5,179 agents studied

Coding

ANZ engineers 55.8% faster with GitHub Copilot in a controlled trial

Randomised trial, 2024

Science

AlphaFold predicted about 200 million protein structures

DeepMind
The jagged frontier of AI capability An irregular ridge line divides the picture. The shaded region beneath the ridge is inside the frontier, and holds drafting, summarising, code completion and translation. The open space above the ridge is outside the frontier, and holds multi-step reasoning, checking its own accuracy, context-dependent nuance and emotional situations. The line is jagged, not smooth, and nothing marks where it falls. Multi-step reasoning Checking its own accuracy Context-dependent nuance Emotional situations Drafting Summarising Code completion Translation Outside the frontier Inside the frontier

Inside the frontier, 758 BCG consultants using GPT-4 were 25% faster and produced 40% higher quality work.

Outside the frontier, they were 19 percentage points more likely to be wrong.

There is no sign telling you which side of the line you are on.

Noy and Zhang, Science, 2023; Brynjolfsson, Li and Raymond, NBER, 2023; ANZ study, arXiv 2402.05636; Stanford AI Index 2025; Dell’Acqua et al., Navigating the Jagged Technological Frontier, Harvard Business School and BCG, 2023.

What goes wrong

Confident errors
Deloitte Australia partly refunded a A$440,000 government report that contained AI-invented references October 2025
The perception gap
Experienced developers were 19% slower with AI, but believed they were 20% faster METR randomised trial, July 2025
AI for attackers too
CBA suspects about A$1 billion in home loans obtained partly using AI-faked documents UNSW Newsroom, March 2026
Over-reliance
The more people trust AI, the less critically they think Microsoft Research, CHI 2025

Plus bias, privacy, security, copyright and explainability.

AP, Guardian, October 2025; METR randomised trial, July 2025; UNSW Newsroom, March 2026; Microsoft Research, CHI 2025.

Australian finance in action

CBA

  • 80 million events screened daily
  • More than 40,000 scam alerts a day
  • Scam losses down 76%
  • Fraud losses down more than 20%

ANZ

  • Copilot trial 55.8% faster
  • Scaled to more than 1,000 engineers

Suncorp

  • AI claims summary saves 5 to 30 minutes per claim
  • About 1,500 staff using it
  • Won an AI ethics award

Macquarie

  • Gemini Enterprise rolled out bank-wide
  • Aiming for daily use by every employee

CBA and Macquarie figures are company-reported.

CBA newsroom, 2025 to 2026; ANZ arXiv study, 2024; Insurance Business Australia, 2025; Google Cloud, October 2025.

The Australian rulebook

ASIC REP 798
Governance is lagging adoption. Licensed humans remain accountable for AI outcomes.
APRA CPS 230 and CPS 234
Operational risk and information security rules already apply to AI and to AI vendors.
National AI Plan
December 2025. No standalone AI Act: existing laws, plus a new AI Safety Institute.

You can delegate the drafting, not the liability.

ASIC REP 798, October 2024; APRA; Department of Industry, National AI Plan, December 2025.

Skills to build now

1
Problem framing. Ask the right question, not just type a prompt.
4
Domain judgement. When to accept, edit or overrule.
2
Verification. Checking is now the core skill of knowledge work.
5
AI literacy. Safe, effective everyday use.
3
Knowing the frontier. Where AI is reliable and where it is not.
6
People skills. Communication, influence, trust.

39% of workers’ core skills will change by 2030.

World Economic Forum, Future of Jobs 2025; Microsoft Research, CHI 2025.

From telling to doing

The agent loop A circle of four repeating steps: plan, act, check, then adjust, and back to plan. Plan Act Check Adjust

2023: AI told you things.

2026 to 2027: AI does things. It researches, fills in forms, updates systems, drafts and sends.

  • Research and document processing
  • File notes, meeting summaries and drafting
  • Compliance monitoring and alert triage
  • Fraud detection, where CBA’s agentic system is already live
  • Code and testing

The length of task AI can complete has doubled about every seven months since 2019, and faster still since 2024, but at 50% reliability. Financial processes need 99% or better.

The agent can recommend. A person decides.

METR, Kwa et al., Measuring AI Ability to Complete Long Tasks, arXiv 2503.14499, March 2025, and Time Horizon 1.1, January 2026; Commonwealth Bank newsroom, 2026.

Productivity and operating models

How big is the prize? A range from cautious to bullish. Cautious: about one per cent of GDP over a decade, from Acemoglu at MIT. Australia: more than 116 billion dollars over a decade, from the Productivity Commission. Bullish: about seven per cent of global GDP, from Goldman Sachs. cautious bullish Cautious about 1% GDP over a decade Acemoglu, MIT Australia A$116 billion or more over a decade Productivity Commission Bullish about 7% global GDP Goldman Sachs

The ceiling is high, but execution decides where you land.

What the winners do differently

  • 3 times more likely to redesign workflows end to end
  • Aim for growth, not just cost-out
  • Shift people from producing work to reviewing and owning it
  • Make decision rights explicit: faster decisions need clearer accountability

Acemoglu, The Simple Macroeconomics of AI, 2024; Productivity Commission, 2025; Goldman Sachs, 2023; McKinsey, State of AI, 2025.

What organisations should do now

1
Governance first. A named human owner for every AI use.
4
Measure honestly. Set a baseline before you deploy, and count value, not licences.
2
Fix the data. There is no value without clean, accessible data.
5
Invest in people. Reskill at scale and protect entry-level pathways.
3
Pick few, go deep. Redesign a handful of high-value workflows, and kill weak pilots early.
6
Model the reversal. Cost the Klarna scenario into every business case.

McKinsey, State of AI, 2025; ASIC REP 798; MIT Project NANDA, 2025.

Five futures. Which is most likely in ten years?

Five scenarios from slow to fast A spectrum from slower to faster change. One, stall: capability plateaus, or trust and regulation slow it down. Two, modest impact: real but small gains, from Acemoglu. Three, normal technology: spreads over decades like electricity, from Narayanan and Kapoor at Princeton. Four, productivity boom: a broad, measurable surge, from Goldman Sachs and Brynjolfsson. Five, rapid transformation: AI matches humans at most valuable work within years, from AI 2027. slower faster 1 Stall capability plateaus, or trust and regulation slow it down 2 Modest impact real but small gains (Acemoglu) 3 Normal technology spreads over decades, like electricity (Narayanan and Kapoor, Princeton) 4 Productivity boom a broad, measurable surge (Goldman Sachs, Brynjolfsson) 5 Rapid transformation AI matches humans at most valuable work within years (AI 2027)
Stall Modest impact Normal technology Productivity boom Rapid transformation

Narayanan and Kapoor, AI as Normal Technology, 2025; Kokotajlo et al., AI 2027, 2025; Acemoglu, 2024; Goldman Sachs, 2023.

How work may change

19% below trend: workers aged 22 to 25 in AI-exposed jobs, mostly because fewer are being hired, not because people are being fired

From doing to reviewing. People steward, check and integrate.

Team shapes vary widely. 32% of firms expect headcount to fall, 43% expect no change, 13% expect it to rise.

The apprentice problem. If AI does the junior work, how do people learn?

Brynjolfsson, Chandar and Chen, Canaries in the Coal Mine?, Stanford Digital Economy Lab, 2025 to 2026; McKinsey, State of AI, 2025.

Opportunities and risks

Opportunities

  • Small teams with huge leverage
  • Faster science and discovery
  • Personal advice, health and education at scale
  • Much lower cost to serve

Risks

  • Power concentrated in a few AI providers
  • Interconnected agents failing together
  • Skills fading through over-reliance
  • Hollowed-out career ladders
  • Governance lagging capability
“We are past the event horizon.” Sam Altman, OpenAI, The Gentle Singularity, June 2025
Today’s language models are a “dead end” for human-level AI. Yann LeCun, public remarks

Sam Altman, The Gentle Singularity, June 2025; Yann LeCun, public remarks; Stanford AI Index 2025.

15.9 million Australians with unmet advice needs
15,100 financial advisers in March 2026, down from almost 29,000 in 2019. Nearly half gone.
1 in 10 Australians who get professional advice
  • Advisers spend less time on file notes, research and paperwork
  • More time with clients
  • Lower cost per client, so advice reaches people currently priced out
  • Licensed humans stay accountable

Where will AI add most value in your work?

Efficiency Quality New products Risk reduction

Investment Trends, 2025 Financial Advice Report; FAAA, March 2026; Adviser Ratings, 2025; ASIC.

If AI keeps improving at this pace, which skills, decisions and human capabilities become more valuable, not less?

  • Analysis is not judgement
  • A recommendation is not a decision
  • A decision is not authorisation
  • Execution is not oversight

AI is doing more of the work. People will do more of the judging.

The march of AI does not make us less valuable. It changes what we are valued for.

Tim Barnes

ASIC REP 798; World Economic Forum, Future of Jobs 2025; Microsoft Research, CHI 2025; Dell’Acqua et al., 2023.

Appendix: slides moved out of the talk

Each was a slide in an earlier version of the talk. Figures and sources as they appeared.

Why this wave is different

Stanford HAI AI Index 2025.

Global examples, and the caution

OpenAI and Morgan Stanley; JPMorgan via media reports, 2025; Bank of America; Bloomberg and Reuters, 2025.

Everyone is using it. Few are profiting.

McKinsey, State of AI, 2025; MIT Project NANDA, 2025; ABS, 2024 to 2025; National AI Centre SME AI Pulse; KPMG and University of Melbourne, 2025.

AI in the tools you already use

Microsoft Source Asia, 2024; Anthropic, Model Context Protocol, 2024.

Appendix: full sources

Grouped by act, with a reliability tag for each. Where no URL is given here, the source is listed by name and date.

Opening

Yesterday

Today

Tomorrow

Beyond

Close

Press ? for shortcuts