For most of the twentieth century, expertise was a fortress. Entry was controlled by credentialing bodies, reinforced by information asymmetry, and defended by the sheer time it took to acquire specialised knowledge. A patient could not diagnose themselves. A small business owner could not navigate a complex merger without an M&A attorney. A first-generation college student in a township outside Johannesburg had no realistic path to the kind of strategic advice that corporate boards paid handsomely for. The asymmetry was not incidental. It was the economic structure of knowledge work itself.
Artificial intelligence is dismantling that structure, not uniformly, not without friction, and not without creating new concentrations of power alongside new forms of access. The question economists and policymakers are now wrestling with is whether the erosion of expertise barriers will primarily benefit workers and citizens, or whether it will concentrate gains further up the economic chain and compress wages for the professionals who once held the keys.
"The key question isn't which jobs can be automated, but which tasks within a job are automated, and who captures the value created by that shift." - David Autor, MIT
MIT labour economist David Autor has spent two decades mapping how technology reshapes the task composition of work. His insight is precise and consequential: automation does not hit jobs as monolithic units. It hits tasks. And the economic consequences depend entirely on which tasks are automated, whether they are complementary to remaining human tasks or substitutive for them, and whether the productivity gains that result flow to workers, firms, or consumers.
Stanford professor Erik Brynjolfsson frames the same question through the lens of institutional choice. AI's distributional effects are not technologically predetermined. Policy, market structure, and firm-level decisions about how AI is deployed will determine whether the technology expands opportunity or concentrates it. The machine does not decide who benefits. We do.
I. The Economics of Task-Level Disruption
To understand AI's labour market effects, it is necessary to abandon the old question, "will AI take my job?", and replace it with a more granular one: which tasks within my job can be automated, which tasks does automation make more valuable, and where does the bundle of remaining human effort sit in the value chain?
Autor's framework distinguishes between substitution and augmentation. When AI automates a task that was previously a bottleneck, the cognitive heavy lifting that justified a professional's hourly rate, the effect is substitutive. The professional can produce the same output faster, which in competitive markets tends to compress prices and, eventually, wages. When AI automates the lower-value scaffolding around a high-judgment task, the research, the formatting, the routine correspondence, the effect is augmentative. The professional is freed to do more of what only they can do, and productivity and wages can rise.
The critical variable is market structure. In markets with significant barriers to entry, augmentation tends to benefit incumbent professionals. A seasoned litigator who can now do in an hour what once took a day does not immediately face new competition from unqualified entrants. The barrier was never the research, it was the judgment. But in markets with lower cognitive barriers and higher volume, where competition already operates on thin margins, the same productivity gains accelerate the race to the bottom. If AI allows any reasonably literate person to produce a serviceable marketing copy brief, then the marginal rate for that brief drops toward zero. The professionals who were earning a living from it face a structural re-pricing of their output.
Brynjolfsson adds a further dimension: the winner-takes-most dynamics of digital platforms mean that AI does not just redistribute tasks, it can consolidate entire markets. When the best AI-assisted legal research tool is marginally better than the second-best, every law firm migrates to the same tool, and the productivity gains accrue primarily to whoever owns the platform. This is not new as a mechanism, but AI accelerates it because the quality differential between a well-trained model and a poorly-trained one can be enormous, and replication costs are near zero.
The wage effects of AI are not uniform across a profession. They are radically different depending on which tasks you perform, at what volume, in what market structure, and whether you own equity in the technology or simply operate it.
II. The Commodification of Expertise
What Is Being Unlocked - and For Whom
The most consequential economic event in the short-to-medium term AI story is not the replacement of low-wage manual work. It is the commodification of high-value cognitive expertise that was previously inaccessible to most people and most businesses on the planet. This is simultaneously the most democratising and the most economically disruptive aspect of the current wave.
To identify which domains face the greatest near-term commodification pressure, consider three criteria: first, the expertise must be high-value, expensive enough that its democratisation represents a meaningful economic shift; second, the knowledge must be primarily pattern-based and articulable, which makes it tractable for large language models; third, the access barrier must have been primarily informational rather than physical, meaning the bottleneck was knowledge itself and not hands-on execution.
By these criteria, three domains stand out as sites of imminent, profound disruption.
Legal Reasoning
Law is perhaps the single most valuable form of expertise that has been systematically inaccessible to ordinary people and small organisations. In South Africa, as in the United States and the United Kingdom, the gap between those who can afford quality legal counsel and those who cannot is staggering, and consequential. Individuals navigate evictions, employment disputes, consumer fraud, and family law crises without representation not because they are incapable of understanding their rights, but because access to the reasoning required to assert those rights has been prohibitively expensive.
Legal expertise is, at its core, a pattern-matching exercise applied to language. Statutes, precedents, contract clauses, and regulatory frameworks are textual artifacts. The core of legal reasoning, identifying the applicable rule, mapping facts to elements, anticipating counterarguments, structuring a persuasive position, is precisely the kind of structured cognitive task that large language models perform with remarkable fidelity. AI systems are now able to draft contracts, identify legal risks in documents, construct basic litigation strategies, and explain legal concepts in accessible language at a quality level that would have required a junior-to-mid-level attorney a decade ago.
The economic effect is twofold. For large law firms, the immediate impact is augmentation: partners can supervise AI doing the associate-level work, compressing the leveraged staffing model that has defined Big Law for a century. For ordinary people and SMEs, the effect is access: for the first time, a food vendor in Alexandra or a sole trader in Pretoria can get a serviceable employment contract reviewed, a lease clause explained, or a dispute letter drafted without paying R2,000 per hour for the privilege. This is not the replacement of lawyers. It is the unbundling of legal expertise, separating the high-judgment, high-stakes work that requires a seasoned human from the routine tasks that once padded legal bills.
Medical Diagnosis and Clinical Reasoning
Medicine presents a more complex case, because the physical examination, the therapeutic relationship, and the procedural skills involved in healthcare cannot be commodified by language models alone. But clinical reasoning, the cognitive process of differential diagnosis, treatment selection, drug interaction analysis, and prognostic assessment, is highly pattern-based and has historically been concentrated in expensive specialists who are geographically unevenly distributed.
The diagnostic accuracy of leading AI systems now matches or exceeds average physician performance on standardised benchmarks across radiology, dermatology, pathology, and primary care triage. More consequentially for access, AI can deliver clinical reasoning to geographies and populations that have never had adequate specialist coverage. A nurse practitioner in a rural community with AI clinical decision support is not a doctor, but the quality of the reasoning available to that practitioner is transforming. The practical implication is that the bottleneck in many healthcare systems will shift from clinical knowledge to the regulated, physical, and relational aspects of care. That is a meaningful shift in what human expertise is for.
Financial Strategy and Investment Analysis
Institutional-quality financial analysis, portfolio construction, corporate valuation, tax optimisation, financial planning, has historically been available only to wealthy individuals and large organisations. The median South African household has never had access to a qualified financial planner, let alone the multi-disciplinary advisory team that manages the wealth of high-net-worth families. The bottleneck has been the cost of the human time required to perform the analysis.
AI does not replace the fiduciary judgment, the ethical obligations, or the relational trust involved in financial advice. But it can generate the underlying analysis, the scenario modelling, the tax-efficient structure options, the portfolio rebalancing logic, at marginal cost. The democratising potential is real. So is the risk: financial advice is a domain rife with conflicts of interest, and AI systems trained on available data can encode the biases and blind spots of the financial institutions that historically produced that data. The question of who trains the model, on whose behalf, and with what objective function is not academic. It is the difference between a tool that genuinely expands access and one that scales existing predatory practices.
III. The Death of Systems of Record
Alongside the commodification of expertise, AI is precipitating a quieter but equally fundamental transformation in the architecture of enterprise knowledge. The dominant paradigm of business software for the last thirty years has been the system of record: a structured, authoritative database that captures what a company knows, what it has done, and what it owns. CRMs, ERPs, HRIS platforms, and document management systems are all built on the same premise, that the primary challenge of organisational knowledge is storage and retrieval, and that the correct architecture for that challenge is a well-structured relational database that humans query and maintain.
That premise is being rendered obsolete. Not because storage and retrieval are less important, but because AI has fundamentally changed what you can do with unstructured information, and most of the knowledge that matters in organisations has never been structured in the first place.
The system of record assumed that the hard part of organizational knowledge was storage. AI reveals that the hard part was always synthesis. The data was never really the bottleneck. The reasoning was.
Consider what a CRM actually captures: a formalised record of customer interactions, structured according to fields that someone decided mattered when the system was configured, entered by salespeople who are measured on other things and who therefore minimise the time they spend on data entry. What it does not capture is the context, the subtext, the relationship history, the promises that were made in hallway conversations, the customer's actual unspoken objection, or the strategic account insight that lives in the head of the senior account manager who just resigned. The system of record captured the shadow of organisational knowledge, not the knowledge itself.
AI changes this in two ways. First, natural language interfaces and ambient capture mean that organisational knowledge no longer needs to be manually structured to be retrievable. A conversation, an email thread, a Zoom call, a Slack exchange, all of these can now be indexed, synthesised, and queried without requiring anyone to translate them into CRM fields. The friction cost of knowledge capture collapses. Second, the value of having a perfectly normalised relational database diminishes when an AI can generate equivalent insights from messier, richer, less structured sources. The CRM's value proposition was always that a structured database was the only way to make customer knowledge queryable at scale. That proposition is no longer true.
The implications for enterprise software are profound. The next generation of business knowledge systems will not be built primarily around storage and retrieval. They will be built around continuous synthesis, systems that don't just capture what happened, but continuously reason over the accumulated record of organisational activity to surface insights, identify risks, anticipate decisions, and answer questions that no one thought to ask. The Salesforce of 2030 does not look like the Salesforce of 2010 with a chat interface bolted on. It looks like an entirely different architectural philosophy: less database, more reasoning layer; less structured record, more living memory.
What Dies and What Survives
This architectural shift will not kill data governance or the need for authoritative records, regulatory and legal requirements will ensure that structured record-keeping persists. What dies is the primacy of the system of record as the interface through which organisational knowledge is accessed and used. The database remains in the back. The AI becomes the front. And the economic value that was previously embedded in building and maintaining the structured database migrates toward those who can build and maintain the reasoning layer on top of it.
For enterprises, the practical implication is that the return on investment for legacy ERP and CRM modernisation projects, already questionable given the cost and complexity of such implementations, has collapsed further. The question is no longer "how do we get our data into the right fields?" but "how do we build an AI layer that can reason effectively over all of our information, structured and unstructured, and surface the right insight at the right moment?" Those are fundamentally different engineering and organisational challenges.
For the workers who have made careers in data entry, data management, and the administration of enterprise systems, the transition is stark. These roles were the connective tissue of the system-of-record architecture. As ambient capture and AI synthesis reduce the need for human-mediated data structuring, these roles face direct substitution pressure — not from AI that is smarter than the humans doing the work, but from AI that makes the work itself unnecessary.
IV. Steering the Transition: Policy and Power
Brynjolfsson's central argument is that the economic outcomes of AI are not technologically determined. The same technology that can democratise access to legal reasoning can also be used to surveil workers more granularly, compress wages more efficiently, and eliminate the informational advantages that gave labor any leverage in employment negotiations. Which path is taken depends on choices, about regulation, about standards, about public investment, and about the distribution of ownership.
A few policy implications stand out as particularly consequential. The first is the question of access infrastructure. The democratising potential of AI expertise is only realisable if the people who would benefit from access can actually access it. In markets with low smartphone penetration, expensive data, inadequate digital literacy, and languages underrepresented in training data, the benefits of AI-mediated expertise will not flow to those who need it most. Public investment in access infrastructure, connectivity, digital literacy, local-language model development, is not optional if the equity promise of AI is to be anything more than rhetoric.
The second is the question of who owns the models. The productivity gains from AI accrue disproportionately to whoever controls the technology. In the current landscape, that is a small number of large technology companies with the capital to train frontier models. If the gains from AI-mediated expertise flow primarily to model owners rather than to the professionals whose outputs trained the models or the users whose data improves them, then AI will widen rather than close the inequality gap. Mechanisms to broaden ownership — whether through profit-sharing requirements, public model investment, or cooperative structures, will determine whether the AI productivity dividend is broadly distributed.
The third is the question of professional re-orientation. The professions most threatened by AI commodification are also the ones most capable of adapting if they redefine their value proposition. Lawyers who can reason at the strategic, ethically complex, and relational level, and who use AI to eliminate the lower-value work that currently consumes most of their billable hours, will find their expertise more valuable, not less. Medical professionals who focus on the physical, procedural, and human dimensions of care will become more valuable as AI handles the diagnostic reasoning. The transition requires investment in professional education that trains people for the residual human work rather than the work AI will do.
The professionals who thrive in the AI era will not be those who resist the commodification of their routine tasks. They will be those who move most aggressively up the value stack into the work that requires irreducibly human judgment.
V. Conclusion: The Expertise Bargain, Renegotiated
The economic compact at the heart of professional expertise, you pay a premium for access to knowledge that took years to acquire and that is scarce, is being renegotiated by AI. The scarcity is dissolving in some areas faster than others. The premium is compressing. The question of who benefits from the dissolution depends on everything except the technology itself.
For the global majority who have never had access to quality legal, medical, and financial expertise, this is potentially the most significant expansion of human capability since the internet made information universally accessible. The difference is consequential: information access did not close the expertise gap, because knowing where to find information is not the same as being able to reason over it. AI can reason over it. That changes what access means.
For workers in knowledge professions, the transition demands a clear-eyed assessment of which tasks within their work are defensible and which are not, and an active strategy for migrating toward the former. The systems of record they have maintained, the research they have conducted, the documents they have drafted, these are increasingly automatable. The judgment, the ethical reasoning, the empathy, the accountability, the irreducibly human relationship, these are not.
For policymakers, the challenge is to resist both the naive optimism that markets will automatically distribute AI's gains equitably and the fatalistic pessimism that technological disruption is inevitable and unsteerable. Autor and Brynjolfsson's combined message is consistent and clear: the distribution of AI's benefits is a policy choice, not a fact of nature. The tools available, investment, regulation, ownership structures, education, and access infrastructure, are adequate to the challenge if the political will to deploy them is present.
The great unbundling of expertise is already underway. The systems of record are already dying. The only open question is whether we rebuild the architecture of knowledge work in a way that concentrates its benefits or expands them. That question will not be answered by the models. It will be answered by us.
— END —
Sources and Further Reading
David Autor, "Work of the Past, Work of the Future" (AEA Papers & Proceedings, 2019).
Erik Brynjolfsson & Andrew McAfee, The Second Machine Age (W.W. Norton, 2014).
Erik Brynjolfsson, "The Turing Trap" (Daedalus, 2022).
David Autor, David Mindell & Elisabeth Reynolds, "The Work of the Future" (MIT Work of the Future, 2022).
