AI and High-Income Work: Who Captures Value?

After looking at Australia’s higher-income occupations and their employment trends, I don’t think the main question is simply, “Will AI take these jobs?” The six groups I examined—chief executives, financial dealers, mining engineers, anaesthetists, surgeons, and judicial and other legal professionals—all had more people employed in 2025 than in 1996. The more useful question is: when AI changes the work, does the value go to the company selling the AI, the business using it, or both?[3]

There is an important distinction behind the income numbers. The ATO’s latest top-one-percent threshold is $429,531 in taxable income for 2023–24, not a salary cutoff. Among the six occupation groups, anaesthetists’ median taxable income was above that line, while their median salary-or-wage income was much lower. Surgeons had a high average taxable income, but a median below the threshold. A job title on its own does not tell us who is in the top 1% or what part of their income came from wages.[1] [2]

The AI supply side is powerful, but not settled

The companies building leading AI models and the infrastructure behind them have an early advantage. Stanford’s 2026 AI Index reports that industry produced 91.2% of the notable AI models released in 2025. OpenAI, Google and Alibaba were among the leading contributors. It also estimates that Nvidia chips account for more than 60% of AI computing capacity. That is evidence of concentration in model development and computing supply. It is not proof that one company controls the whole AI market.[4]

There is competition too. Stanford found four leading model providers clustered within 25 points on one human-rated leaderboard in March 2026. The leading closed model was only 3.3% ahead of the best open model on the report’s comparison. Those measures do not show market share, but they do suggest that a provider’s lead can be challenged. For an AI company, the commercial opportunity is real; so are the costs of keeping up and the risk that customers can switch to another model.[4]

The first revenue can come from access fees, subscriptions and pay-for-use services. The Federal Reserve’s 2026 survey of 734 US financial executives found that early company spending was generally modest and often went to software subscriptions, services and training rather than new hardware or large internal systems. In other words, suppliers can earn revenue when a firm starts using AI, before that firm has proved a lasting productivity gain.[5]

Pricing is changing as well. BCG describes a shift in some enterprise AI services from flat-fee subscriptions toward charges linked to how much computing a customer uses. That makes the cost of AI more visible to the adopter. It also means a supplier’s growth in usage can become a customer’s variable operating cost. The return depends on what the extra use actually produces.[8]

The adopter has to turn access into business value

For a business using AI, buying access is only the starting point. The return comes when the work changes in a way that matters: more customers served, shorter waiting times, better products, fewer mistakes, or more completed work from the same specialist team.

The Federal Reserve survey offers a useful clue about where firms are seeing value. Executives reported AI-attributed revenue per employee growth of about 0.6% in 2025, with larger gains expected in 2026. The reported productivity gains were above 2% in high-skill services and finance. The strongest links to revenue growth were developing or improving products and reaching customers more effectively. Cutting labour costs was not a consistent driver across the companies surveyed.[5]

That is relevant to the Australian occupations in the previous analysis. AI may help a hospital clinician spend less time writing notes, a financial team review more information, or an engineering group evaluate more options. The business case is not automatically fewer specialists. It may be that the same specialist team can do more useful work, reduce backlogs, or offer a service that was previously too expensive to provide.

There is an example in health care. Stanford’s 2026 AI Index reports that some AI note-taking tools were associated with physicians spending up to 83% less time writing clinical notes, alongside reported reductions in burnout. The report also warns that rigorous evidence for clinical AI beyond certain tools remains limited. Saving documentation time is a promising use case; it is not, by itself, proof of better health outcomes or a reason to remove clinical oversight.[4]

Australia has a large potential prize, but it is not guaranteed. The Productivity Commission’s 2025 report says an estimated gain of more than 2.3% in how efficiently Australia combines labour, equipment and technology over the next decade could translate into about 4.3% labour-productivity growth and around A$116 billion in additional GDP in 2023–24 prices. The Commission stresses that there is considerable uncertainty. This is a whole-economy scenario, not a forecast of what any one business will earn.[7]

Adoption is also uneven. The Productivity Commission cites the National AI Centre’s first-quarter 2025 tracker: 82% of firms with 200–500 employees reported adopting AI, compared with 33% of firms with 0–4 employees. Large Australian businesses were investing in customer interactions, internal processes, manual-task reduction and new or improved products. For smaller firms, the opportunity may be significant, but limited time, skills and confidence can make it harder to turn a tool subscription into a changed way of working.[7]

The workforce change may begin inside the job

The evidence so far points more clearly to task changes than to a broad collapse in employment. In the Federal Reserve survey, executives expected AI to reduce total employment by less than 0.4% in 2026. They expected routine clerical work to account for a smaller share of jobs by 2028, with some movement toward technical roles such as engineers, data analysts and scientists. These are executive expectations from a US survey, not a measured outcome for Australia.[5]

AI exposure is not limited to lower-paid work. Anthropic’s analysis of its own AI-use data found that US workers whose tasks were more exposed to AI tended to be more educated and had average earnings 47% higher than workers with no observed task coverage. The researchers found no systematic increase in unemployment among highly exposed workers since late 2022, but did find suggestive evidence that hiring of younger workers had slowed in exposed occupations. The data describes one AI platform’s use and cannot be treated as a direct measure of Australian job losses.[6]

Stanford’s AI Index also reports that US developers aged 22 to 25 saw employment fall nearly 20% from 2024, while older developers’ headcount continued to grow. That pattern is worth watching, but it does not prove AI caused the decline. A more cautious reading is that productivity gains can arrive alongside changes to who gets hired and which tasks are used to train people for senior roles.[4]

That raises a question for high-income work. If AI takes on more research, drafting, documentation or routine analysis, businesses may keep demand for experienced people while needing fewer junior staff for parts of the old pathway. If that happens, the issue is not only how many jobs disappear today. It is also how people gain the experience needed to become tomorrow’s specialists and leaders.

Two value propositions, moving at different speeds

For AI companies, the proposition is to sell useful capability repeatedly: through access, subscriptions, usage charges and the computing needed to deliver the service. The early revenue can be visible and recurring. But provider success depends on managing computing costs, keeping models useful and giving customers a reason to stay as alternatives improve.[4] [8]

For businesses adopting AI, the proposition is different. It is to turn that capability into better operating results. The early evidence suggests that product improvement and reaching customers can matter more than simply cutting staff. This can allow both sides to benefit: the supplier earns from providing the service, while the adopter earns from serving more people or creating more value. But if the AI bill grows and the workflow stays the same, the supplier may capture the clearer return.

My practical test is simple: choose one workflow, record how long it takes, how much rework it creates, how many customers it serves and what quality looks like. Then trial AI with a person still accountable for the result. Count the full cost, including the subscription, usage, training, integration and checking. If the tool only adds seats and prompts, the business has bought access. If it improves output, service or revenue after all costs, it has an adoption case.

The 30-year occupation data does not show that AI has already reduced these high-income workforces. It shows that the selected groups grew, even as technology and business conditions changed around them. The next shift may be less about wiping out entire professions and more about changing how work is divided, where junior roles begin, and who owns the extra capacity created by AI.

References

More writing · Previous: Article 1 · Next: Article 3 — Choosing a career in the AI era · About Gokul