Introduction
With Five Cents, we’re exploring which jobs AI could really change by looking at the work inside them, rather than predicting which titles will disappear. We’ll cover administration, customer support, writing, software and analysis. Then we’ll separate demonstrated benefits from promises, and consider why results differ around the world.
The key points
The best starting point is the difference between a task and a job. A task might be drafting a routine reply, summarizing a file, or checking a standard form. A job combines those activities with others: coordinating with colleagues, handling exceptions, understanding customers, and taking responsibility.
Generative AI is particularly relevant when work is digital, repeatable, and information-heavy, especially when its output can be checked against a clear standard. A global exposure estimate published in 2025 suggests about one in four workers is in an occupation with some potential exposure. That does not mean a quarter of jobs will disappear. It describes tasks that technology could affect, not actual redundancies.
There are also two different ways work can change. Automation means a system completes a task with little human involvement. Assistance means it helps someone who still guides, checks, or revises the result. The same tool can support either approach, depending on how an employer uses it. And producing an answer is not the same as completing useful work. If someone must spend substantial time correcting that answer, the apparent saving may shrink.
For workers, the practical question is which activities become faster, which still need judgment, and whether the job is reorganized around that difference. Clerical and administrative work is especially exposed because many tasks already follow digital processes. AI can help prepare standard correspondence, organize records, or process routine documents.
But an administrator also notices missing information, resolves conflicting requests, and coordinates people. Bookkeeping illustrates the distinction. Categorizing a straightforward transaction is different from investigating an unexplained discrepancy or accepting responsibility for the accounts.
Customer support offers measured evidence of assistance
Customer support offers measured evidence of assistance. A workplace study involving more than five thousand agents found an average fourteen percent increase in issues resolved per hour with an AI assistant. Less experienced workers benefited more. The most experienced saw little or no gain. That demonstrates an improvement in one setting, not a universal result or proof that support teams are becoming unnecessary.
Writing and software development show a similar split. AI can produce a first draft, rewrite an explanation, suggest code, or generate documentation. Those are meaningful parts of a working day. But someone still has to decide what the text should accomplish, verify its claims, and judge whether it suits the audience. In software, plausible-looking code must fit the wider system, pass tests, and meet security requirements.
Analytical and financial work also contains exposed tasks, such as summarizing reports and organizing information. Yet interpreting incomplete evidence, validating inputs, and making accountable decisions are different from generating a polished summary. The work may shift toward reviewing and integrating output. That can change skill requirements without removing the whole role.
How confidently can we describe these effects? A demonstration shows that a model can produce something under particular conditions. A controlled experiment or workplace study can measure speed or quality on defined tasks. Neither alone establishes what happens to staffing, earnings, or employment across an economy. Reviews of observed evidence find real but uneven productivity gains, with large-scale displacement still limited in the evidence examined. That is not a guarantee about the future.
Results also differ across countries and workplaces. Reliable digital infrastructure, access to tools, and the actual tasks within a role all matter. An accountant or support worker may do quite different work in different settings. Some lower-income workplaces may face competitive pressure before workers gain access to the tools that could help them.
What it means
The biggest mistake is treating exposure as a forecast of unemployment. Another is assuming that time saved automatically means fewer employees. An employer might reduce staffing, increase output, or improve service. To assess your own work, identify recurring tasks. Then ask whether AI can perform them reliably, and how much checking remains. Include confidentiality, the cost of errors, and responsibility for the final result.
Learning to evaluate AI output may be more useful than simply learning to generate it. And keep this discussion separate from robotics: evidence about digital text and code does not establish what machines can do in physical workplaces. AI is most likely to change particular tasks. Its benefits vary, and workplace choices help determine what happens to jobs.
To explore the practical next steps, you can generate two more Five Cents podcasts: Building Skills for AI-Assisted Work and How Employers Measure AI Productivity. And with that, you're up to speed in a few minutes.