Friday, 7 August 2026 • By Timothy, CPA — Managing Director, Professional Financelink (PFL)

Stylised illustration of a steeply rising curve and a flat line running side by side across the same chart, with the gap between them shaded

AI Is Getting Smarter Faster Than Your Productivity Is Improving. That Gap Is a Budgeting Instruction.

Australian labour productivity grew 0.3% over the year to March. Meanwhile AI is solving open mathematical problems. The most useful explanation for that gap tells finance leaders exactly where to point their AI spending — and it isn't where most of it is going.

Two facts about the middle of 2026, sitting side by side, that don't seem to belong in the same year.

The economist Noah Smith opened a widely-shared essay on 26 July with the first one, and it's worth confirming because it sounds implausible: a Harvard mathematician used Anthropic's Claude Fable 5 to construct a counterexample disproving the general form of the Jacobian Conjecture, a problem about polynomial maps that had stood since 1939 — 87 years. Terence Tao, widely regarded as the finest living mathematician, spent the following days publishing his own detailed mathematical reconstruction of the proof, and separately shared an AI conversation showing how he worked through an alternative version himself. Around the same fortnight, a separate group of researchers used AI to help crack "unclonable encryption" — a six-year-old open problem in quantum cryptography.

The second fact is less dramatic and, for this readership, more useful: Australian labour productivity fell 0.6 per cent in the March quarter and grew just 0.3 per cent over the year to March. Hours worked rose 2.2 per cent over the same year. As the Productivity Commission's Deputy Chair, Alex Robson, put it in the June bulletin, "in aggregate, we are working harder and longer, but we are not working smarter." We are now 0.1 per cent below where productivity sat in March 2023.

Reading those two facts side by side is more useful than treating the first as a curiosity, which is how it's usually discussed. Explaining the gap correctly tells you where to spend.

Three Explanations, and Why the Third One Matters

Smith's essay is what got me thinking about this properly, and the competing explanations he lays out for the gap fall into three camps.

Friction. The capability is there; adoption, governance and organisational plumbing are the bottleneck. On this view the productivity gains are coming, just later than the demos suggested. This is the comfortable explanation, and it's partly true — around 12 per cent of Australian businesses reported using AI in 2024–25, which is not the profile of a technology that has finished arriving.

Diminishing returns to intelligence itself. The AI researcher François Chollet has argued that intelligence isn't an unbounded quantity like height but "a conversion ratio, with an optimality bound." His analogy: pushing it higher isn't like building a taller tower — it's like polishing a ball that's already close to perfectly round, where there's only so much rounder it can get before further effort barely shows. Related arguments from Arvind Narayanan and Sayash Kapoor point to "irreducible error": in many real-world tasks, human performance is already close to the theoretical limit, and no amount of additional intelligence gets you much further.

We're measuring the wrong thing. This is Smith's own position, and it comes with a practical instruction attached. His argument isn't that AI will out-think a person at any single task — it's that the real payoff comes from pairing human-style judgement with what a computer already does that no human can: hold vast amounts of data, never tire, run the same process a thousand times over. As he puts it, the mistake is "arrogantly privileging the kind of mental tasks we humans happen to do especially well" over the tasks we're actually bad at.

+0.3%

Australian labour productivity growth over the year to March 2026, against 2.2% growth in hours worked. Output is rising; effort is rising faster.

−0.1%

Non-market sector labour productivity over the year to March 2026 — the part of the economy where much NFP, care and community services activity is measured.

Three Places the Gains Actually Sit

Smith's argument gives three concrete reasons AI could still deliver large productivity gains even if intelligence itself is hitting diminishing returns. Each one translates directly into a finance function.

Replicability. Skilled human judgement is capped by how many experienced people exist and how long it takes to train the next one. AI capacity isn't capped that way — it scales the way a data centre scales, built out of whatever you're willing to reinvest in it. What that means operationally is not "a smarter analyst" but "the same check, run on everything." Finance functions have spent a century designing around scarcity of attention: we sample, we set materiality thresholds, we test 30 transactions and extrapolate. Those are all workarounds for not having enough hours. The gain here isn't a better judgement on one invoice, it's the same adequate judgement on all 40,000 of them.

Distributed tacit knowledge. Smith's example is Zeiss, which makes glass so precise that even its own staff can't fully explain how they do it — the expertise lives in thousands of small, individually-held habits rather than a manual, which is why a competitor can't just hire a few engineers away and replicate it. His point: AI doesn't need to be brilliant to close a gap like that; it just needs to sit across enormous volumes of process data and notice which small things are working, letting a slower-moving organisation catch up to a well-run one much faster than before.

Every finance team has its own version of this. The reconciliation that only works because someone knows which three suppliers always code themselves wrong. The month-end sequence that exists nowhere except in the order one person does things. The claim rejection that gets fixed in forty seconds by whoever has seen it two hundred times. That knowledge is real, it's valuable, and it walks out the door with resignations. Capturing it — and this is the unglamorous, unexciting work — is where the compounding is.

Pattern extraction beyond human intuition. Smith's third argument centres on what he calls "cloud laws" — patterns real and exploitable enough to matter, but too tangled for any one person to fully grasp or hand on to somebody else. The finance-scale version is modest but real: patterns across your own data that nobody would spot because they're spread across too many dimensions. Which combinations of service type, region, roster shape and client cohort actually drive margin. Which early signals precede a claim being rejected. Not insight in the strategic sense — correlation at a resolution humans don't have the working memory for.

Note: The scenarios in this post are based on real experiences — mine and those shared by colleagues across the sector. Details might have been changed and modified slightly to protect confidentiality, and mostly used 1st person perspective for convenience.

What This Means for Next Year's Budget

If the third explanation is right, there's a real risk that a meaningful share of AI budgets are pointed the wrong way — not from bad choices, but because the obvious use case and the highest-value one usually aren't the same thing.

The typical business case funds a smarter version of something a person already does well — a better first draft, a faster answer to a question someone could have answered. Those tools are pleasant, they demo brilliantly, and they produce exactly the kind of diffuse, unmeasurable benefit that shows up in nobody's numbers. It is not an accident that aggregate productivity hasn't moved while AI use has climbed.

The alternative is to fund the boring three: coverage instead of sampling, capture of process knowledge that currently lives in people's heads, and pattern detection across volumes nobody reads. None of those is exciting in a management reporting pack. All of them have denominators, which means all of them can be proved.

There's a specific warning here for this readership. The non-market number above is where much community services, care and NFP activity is measured, and organisations in this position often face the sharpest version of the problem: rising compliance load, rising hours, funding indexed below cost growth, and the least slack to run experiments. The margin for a nice-to-have AI tool is essentially zero. Which is an argument for being more selective, not for standing still.

A note on AI and sensitive data: the applications described here — full-population transaction review, process knowledge capture, cross-dimensional pattern analysis — involve your most sensitive records. Confirm in writing whether any vendor trains its models on customer inputs before that data goes anywhere, and include the AI features already embedded in software you use, not just tools you deliberately adopt.

The Uncomfortable Version

I'll be direct about the part I find hardest. If Chollet is right and intelligence really is subject to diminishing returns, then a lot of what's being sold on the promise of a step change won't deliver one, and some organisations will spend three years finding that out. If Smith is right, the gains are real but they're in the places nobody wants to fund, because "we captured how Julie does the month-end close" is a terrible slide.

Either way, the instruction for a finance leader is the same, and it's unglamorous: stop budgeting for a productivity leap, and start budgeting for coverage, capture and measurement. The organisations that will look transformed in three years won't be the ones that bought the best model. They'll be the ones that spent 2026 turning tacit process knowledge into something reproducible, while everyone else was waiting for the step change.

The aggregate numbers say the step change hasn't arrived. The individual numbers — the ones inside your own organisation, if you've been measuring them — are the only ones that tell you whether it's arriving for you.

Is your AI spending pointed at the things humans do badly, or the things they do well?

Coverage, process capture and measurement are where the compounding is — and they're the hardest to get funded without a clear case. PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations.

Talk to PFL →

Timothy, CPA is Managing Director of Professional Financelink (PFL), providing senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations. 20+ years in finance leadership across NFP, NDIS and SME.

Sources

Tomorrow: the week's AI news wrap. Sunday: finance reads across NFP, NDIS, aged care and childcare.

Search This Blog

Finance Intelligence