In July 2025 I wrote a post about ChatGPT and financial models that ended up as one of the highest-rated posts in the history of r/accounting. The argument was narrow: careless AI output is dangerous when a lender or a board will rely on the number, and the professional who reviews it never stops being necessary. That argument has held up better than I expected, partly because the evidence for it now comes from outside accounting entirely.

On 28 August 2026, Andrew Ng sat down for a thirty-eight minute interview on the Silicon Valley Girl podcast under the title The Biggest Opportunities in AI Aren't Where You Think. It has been watched more than 850,000 times. Ng co-founded Coursera, built the founding Google Brain team, has taught AI to roughly eight million people, and has no commercial interest in talking down the technology. His message sits between the doom story and the hype story, and it maps onto accounting work more precisely than most AI commentary does.

The point is what his analysis implies for an Australian practice or finance team deciding what to do in the next twelve months.

Why Task-Level Analysis Changes the Answer

The framing Ng pushes back on is the job apocalypse, the idea that AI takes over entire occupations and leaves people with nothing to do. His counter is the task analysis that economists such as Erik Brynjolfsson at Stanford and Andrew McAfee at MIT have produced. They break each job into its component tasks and ask which of those tasks AI can handle.

The number Ng cites is that AI could do perhaps 30 to 40 per cent of the tasks in many jobs. He is explicit that this is not the same as replacing 30 to 40 per cent of the people, and he is equally explicit about the part that is easy to skip. The remaining 60 to 70 per cent of the work does not sit there unchanged. It becomes the part where human value concentrates, because it is the part the machine cannot reach.

For accountants, the split is easy to see once you stop looking at job titles. Bank reconciliations, data entry, formatting, pulling the same figure out of three systems, drafting a first cut of a management report, and producing a set of statements that a reviewer then rebuilds by hand are all task-shaped. They are repetitive, they follow defined rules, and they are exactly where AI-assisted automation performs well.

The other side of the split is where the money actually is. Deciding whether a revenue assumption is credible. Knowing that a client's September is always slow because of how their industry works. Judging whether a covenant breach is a real risk or a timing artefact. Explaining to a lender why the downside case still services the debt. None of that is a task with a defined input and output, and none of it is what AI systems are good at.

Anyone who has watched an AI tool produce a confident, tidy and subtly wrong answer already understands the difference from the receiving end. The tool did not fail at the tasks. It failed at the judgment.

Software Engineering Went First, and Hiring Went Up

The most useful part of Ng's argument for anyone outside technology is what he says about software engineers, because they are the profession that has had longest exposure to AI and the clearest evidence base.

Writing code is the thing AI models are unambiguously excellent at. If the job-replacement thesis were correct, software engineering would be the first occupation to empty out. Instead, Ng points to job openings in software engineering rising, and to the engineers he knows being busier than they have ever been. What happened is that the work broadened. Developers took on more of the surrounding pipeline, moved from narrow tickets towards end-to-end ownership, and used the speed gain to do more than they did before.

Read it as a forecast for accounting and finance roles. The pattern is that AI absorbs the mechanical portion of a role, the role expands to include the parts that were previously split across several people or simply left undone, and the total volume of work grows. In a mid-sized Australian firm, the tasks that get absorbed first are the ones nobody enjoys and nobody invoices for directly. The capacity that frees up either goes into advisory work and better analysis, or it goes nowhere, and that choice is a management decision rather than a technology outcome.

The comparison also sets a realistic expectation about speed. Software engineering has had excellent AI tooling for years and the profession is still figuring out how to organise around it. There is no version of this where a finance function is transformed in a quarter.

The Context Advantage Is the Durable Part

Ng's explanation for why human judgment survives is specific, and it is the part I would hand to any accountant worried about their position.

He describes the human advantage as contextual. Years of client contact, an accumulated understanding of how a particular business actually operates, the history of decisions and their consequences, the relationships that make a difficult conversation possible. That context is not stored anywhere an AI model can retrieve it, and he does not see a mechanism on the horizon that would let a model extract it. His conclusion is that the value of that contextual judgment rises as the mechanical work gets cheaper.

For an Australian practice, the context advantage is unusually concrete. It includes the client's trading pattern across a year that does not follow the calendar. It includes positions taken on treatment of items and the reasoning behind them. It includes the informal knowledge that a particular supplier is being paid late and why that matters to the cash forecast. It includes knowing which figures in a set of accounts an experienced reader would challenge first.

None of that can be prompted out of a model, because it was never written down. It lives in the professional's head and in the client relationship. AI changes the cost of producing an output and does not change the scarcity of knowing whether the output is right.

This is also the honest limit of the technology, and it is worth saying plainly because much of the marketing in this space does not. A system can be built to compute correctly and to refuse to present a number that fails a check. It cannot be built to know that a client will lose their largest customer in March. The professional supplies that, and the professional remains accountable for it.

His Own Finance Team Stopped Waiting for Engineers

The detail in the interview that should get the most attention from finance people is almost an aside. Ng describes the finance team at his own organisation writing scheduled scripts that open files, check consistency and alert the team when something looks wrong. That work would previously have sat in an engineering backlog.

Consider what that means. His finance team identified a repeated checking task, automated it, and now has a process that runs on a schedule and reports exceptions. That is the shape of useful AI adoption in finance. It is a small, boring, unglamorous process that removes a manual check from someone's week and makes the result visible.

It also matches what the finance work actually needs. Most of the value in a finance function comes after the statement is produced: knowing that it is internally consistent and that the assumptions behind it still hold. Those are checks, and checks are exactly what a scheduled script does well. They are also the first thing to be skipped when a month-end is under pressure, which is when they matter most.

Ng extends the same point to hiring. For marketers, recruiters and operations people in 2026 and 2027, his bar is whether the candidate can build with AI. The distinction is worth taking seriously. Using is asking a tool for an output and hoping. Building is turning a piece of work you understand into a process that runs reliably, that checks itself, and that someone else can review.

The Training Problem Nobody Wants to Raise

There is a section of the interview that matters more to accounting than it first appears. Ng, who has spent his career building AI education companies and has taught AI to millions of people, says models are poor tools for learning. The reason is mechanical. Learning happens in the effort, and a model that supplies the answer removes the effort.

For a profession built on an apprenticeship model, that is a live problem. A graduate accountant currently learns by building the reconciliation, chasing the difference, finding the misposting and understanding why it happened. The intuition that lets a senior accountant look at a set of accounts and sense that something is wrong is the accumulated residue of hundreds of those small struggles. If the software produces a reconciled result and the junior never sees the working, the intuition does not form, and the profession gets a cohort that can operate tools without being able to tell when the tools are wrong.

Ng also notes that the university system is roughly two years behind the technology, which matches what most accounting graduates would say about their degrees. The practical response is to be deliberate about which work juniors do by hand, and why. The tasks that build the judgment worth having are the ones a firm should protect, even when they can be automated, because automating them also automates away the training. The reasoning behind a decision is what transfers, and the reasoning is exactly what gets skipped when the output arrives finished.

What This Means for a Practice in 2026

The temptation with a macro argument is to nod along and change nothing, so it is worth translating into specific decisions.

The first is to be honest about which tasks in your week are mechanical and which rely on judgment, because the two need opposite strategies. For the mechanical tasks, the goal is to remove them from people's time, and the tools to do that are now within reach of a finance team without a software project. For the judgment tasks, the goal is to protect the time they need, because that is the work the client is actually paying for and it is the work that gets compressed when a month is busy.

The second is to stop treating AI adoption as a decision about a product. The organisations getting value took a repeated, well-understood piece of work and rebuilt it as a process with a defined input, an automatic check and a clear owner. That work is unglamorous, and it is where the return is.

The third is to be sceptical in a specific way. The question to ask about any AI tool in a finance context is what it checks before it shows you a number, and whether it can produce an answer that is internally inconsistent without telling you. A tool that cannot fail loudly will fail quietly, and quiet failures in finance are the expensive kind. That test does not depend on which model is behind the tool or how impressive the demonstration looks.

The fourth is to treat the data question as a first-class decision at the start of the work. Client financial information carries confidentiality obligations, and the practical question of what may be sent to a third-party model, and what may not, deserves a written answer before it becomes an incident.

The Discipline Behind the Tools

If there is one correction to the popular reading of Ng's interview, it is that the opportunity sits in what gets built on top of the models. He says this directly. The models are improving quickly and are accessible to everyone, including your competitors. The advantage sits in what gets built on top of them, in the workflows and tools and vertical applications that solve a specific problem for a specific kind of user.

That is a harder message than it sounds, because it moves the bottleneck away from technology and towards clarity. The limiting question is whether you know precisely which piece of work is worth rebuilding, what the correct output looks like, and who is accountable when it is wrong. Those questions belong to people who understand the work, which in finance means accountants rather than engineers.

This is the point where the argument in my 2025 post and the argument in Ng's interview converge from opposite directions. He arrives at it from labour economics and says the human portion of the work becomes more valuable. I arrived at it from broken financial models and said the same thing about review. Both end up in the same place. The professional's value is in defining what correct looks like and being answerable for it, and the machinery around them exists to make that cheaper to achieve and harder to fake.

I have spent the time since that 2025 post building the machinery, and I have written about what I got wrong in the process. The short version is that the bottleneck was the scaffolding around the models. Anyone who reads Ng's interview and concludes that accountants should buy a subscription has drawn the least useful lesson from it. The useful one is that the tools to build are now available to people who understand the problem, and the people who understand the problem are already in the profession.

Related reading

• Why ChatGPT isn't replacing Accountants anytime soon
• AI Agents for Australian Businesses: Transforming Operations
• ChatGPT vs Copilot vs Gemini for Excel Work: Which One Wins?