Jobs built around repeatable digital tasks face some of the strongest pressure from AI. Data entry, routine administration, bookkeeping support, customer service and parts of writing, translation and programming are exposed. Work involving physical care, repairs and responsibility for difficult decisions is generally harder to replace with current generative AI. The key distinction is between automating tasks and eliminating an entire job. 1 4
I’m a frontend developer, and I’ve been unemployed since June 2026. I suspect AI is part of what has made finding work harder, although I cannot separate its effect from hiring budgets, competition and the wider market. I still love AI and use it myself. Being excited about the technology and worried about your livelihood can exist at the same time.
This article looks at what the research supports, which tasks deserve attention and what the changes could mean for developers and other workers. It focuses on generative AI: tools that produce text, code, images and other digital output. Robotics can affect a different set of tasks.
Evidence reviewed: 12 November 2026. The studies below describe global exposure, employer expectations or US employment data; they do not establish why any particular person in Estonia is unemployed.
Which jobs are most at risk from AI?
The ILO’s 2025 global index estimates some generative AI exposure for roughly one in four workers worldwide. It expects job transformation to be more common than complete replacement. 1
The table combines research findings with practical examples of tasks. It is a comparison of exposure, not a ranking of replacement probabilities.
| Job group | Examples of exposed tasks | Work that still needs attention |
|---|---|---|
| Data-entry clerks and administrative assistants | Extracting information, completing forms and preparing standard documents | Unclear records, exceptions and responsibility for accuracy |
| Bookkeeping and accounting clerks | Invoice processing, routine entries and document matching | Reconciliation problems, unusual transactions and professional judgement |
| Customer-support agents | Common questions, request classification and standard responses | Escalations, disputed cases and decisions outside the script |
| Copywriters and translators | Generic product copy, rewrites and first-draft translation | Original reporting, terminology, audience and editorial accuracy |
| Graphic designers | Standard promotional assets and repeated visual variations | Art direction, brand consistency and evaluating whether the result fits the brief |
| Web developers and programmers | Routine code, standard interfaces and common components | System behaviour, integration, accessibility and production reliability |
| Financial analysts | Initial information gathering and routine reporting | Assumptions, unusual situations and responsibility for recommendations |
The ILO identifies exposure in clerical, financial, software and language work; Anthropic’s usage study highlights programming and support. OpenAI’s framework discusses bookkeeping, and WEF’s employer survey includes graphic design. The specific task examples above are my practical interpretation of these findings. 2 3 5 7 8
Data entry, administration and bookkeeping
These roles deserve particular attention because much of their output consists of structured information. However, the ILO also notes that automating clerical tasks can be harder in practice than theoretical exposure suggests. 2
Consider an invoice workflow. Reading a clean document is one step. Someone still needs to deal with a missing purchase order, inconsistent supplier details or a payment recorded against the wrong invoice. Automation could reduce manual entry while leaving exception handling with people. Whether that supports the existing team or reduces staffing is a separate business decision.
Customer support and routine service work
A support workflow may contain hundreds of similar questions alongside a smaller number of difficult cases. That makes it useful to evaluate individual request types instead of treating every customer interaction as equivalent.
Anthropic’s March 2026 study found no systematic unemployment increase among highly exposed workers, while reporting tentative evidence of slower hiring for younger workers. Its observed AI usage also covered much less work than theoretical capabilities suggested. This is early evidence from US labour data and Claude activity, not a universal prediction. 7
For a business, a working support system also needs current information, integrations and a way to pass unresolved cases to a person. I cover those practical considerations in my guide to AI chatbot costs for a website.
Writers, translators and graphic designers
The ILO’s media research describes high exposure for writing and translation, with some work shifting towards checking and supervising generated output. 3
The consequences can differ between a generic catalogue description and a reported article, or between a rough translation and a sensitive customer communication. In my view, the useful question is what the customer is paying for: producing an initial draft, understanding a subject or making a reliable editorial decision.
Graphic design also appears in the World Economic Forum’s employer survey among roles expected to decline. Those expectations cover several economic and technological forces through 2030; they are not measured job losses caused solely by AI. The finding should not be extended automatically to every design discipline. 5
AI exposure, automation and job losses are different
Four questions often get collapsed into one:
- Exposure: Could AI perform or assist with some of the work?
- Adoption: Has an organisation actually integrated it into a reliable workflow?
- Hiring: Does that organisation now recruit differently or open fewer positions?
- Employment: Have jobs been eliminated, changed or created?
A convincing demonstration answers only part of that sequence. A company also has to judge the cost of supervision, the consequences of errors and the demand for the finished work.
The OECD’s 2026 capability analysis places routine information processing closer to current AI capabilities than work requiring contextual judgement, interpersonal understanding and responsibility. It also emphasises that organisational choices affect the eventual outcome. 4
That is why an occupation can be highly exposed and still employ people. A task becoming cheaper could allow a business to do more of it. Alternatively, it could lead the business to maintain the same output with less labour. These are possible responses, not a single inevitable forecast.
Will AI replace programmers and frontend developers?
AI can take over parts of programming without replacing every programmer. My concern is that the pressure may appear first in the number of available jobs.
Software development is among the exposed fields identified by the ILO. At the same time, employers in the World Economic Forum’s survey expected software and application developers to be among the fastest-growing occupations through 2030. The two findings measure different things: changing tasks and expected employment demand. 2 5
A generated interface is one part of a working product
As a frontend developer, I would evaluate an AI-generated interface against the same requirements as any other implementation:
- Does it handle loading, empty, error and interrupted states?
- Does the permission model match what the API actually allows?
- Can someone use it with a keyboard and assistive technology?
- What happens when the data shape changes or a request fails?
- Is the result maintainable, observable and compatible with the existing product?
These questions are examples from software engineering practice, rather than claims that AI will never help answer them. They explain why generating code and taking responsibility for a product are different amounts of work.
For me, the practical response is to use AI while staying able to assess its output. Framework knowledge still matters when a plausible-looking component creates a state-management bug or hides a failed request. Clear requirements and good review become especially useful when producing a first implementation is quick.
My comparison of AI tools for WordPress development approaches tool choice through concrete development tasks.
The hiring problem can arrive before the job disappears
Stanford’s August 2026 analysis of US payroll data found an employment gap concentrated among 22–25-year-olds in highly exposed occupations, with reduced hiring playing a major role. Its authors explicitly describe the findings as non-causal; experienced workers did not show the same gap. 6
That cannot explain my own job search as an experienced developer in Estonia. It does make the hiring question worth discussing.
Imagine a team considering an additional frontend developer. If AI helps the existing team finish the planned work, management might delay that vacancy. Another company might use the same productivity gain to build a product it could not previously afford. This is an illustrative scenario, not a claim about what happened at a particular employer.
AI does not need to replace an entire profession to make jobs harder to find. Missing opportunities can matter before a job title disappears.
Which jobs are safer from AI?
Current generative AI has less direct reach into some kinds of work. Physical presence, real-world constraints and responsibility can make a role harder to automate, although they do not provide a permanent guarantee. 4 8
Skilled trades and work on site
Electricians, plumbers and roofers are examples of work with lower immediate exposure to language-based AI in OpenAI’s jobs-transition framework. Estimates and customer messages may change, while performing the physical work requires other capabilities. 8
Being able to describe a leaking pipe is different from reaching it, assessing the surrounding damage and completing a repair. Robotics and other forms of automation deserve separate consideration; a low chatbot exposure score does not cover every technology.
Hands-on care
Stanford’s employment dashboard places home-health aides among occupations with low AI exposure, contrasting their physical and relational work with more exposed digital occupations. 9
This distinction applies to tasks. It does not make every healthcare role equally protected or prevent documentation and administrative work from changing.
Work involving judgement and accountability
The OECD identifies contextual judgement, interpersonal understanding and responsibility as areas further from current AI capabilities. 4
My practical interpretation is to look at what happens when a situation becomes unusual. Who must understand conflicting needs, explain a decision and take responsibility for the outcome? Those parts of a role deserve attention when assessing how it may change. Having “manager” or “senior” in a title does not settle the question.
Will AI take my job? How to assess your work
Start with a normal working week. Separate the tasks you actually perform from the job title on your contract. Then ask:
- How much follows a stable pattern? Repeating the same process is different from resolving a new problem.
- How easy is it to check the result? A plausible answer can still require expensive verification.
- What information is missing? Real work may depend on context that never reaches a prompt.
- Who handles exceptions and consequences? Someone must own the outcome when the normal workflow fails.
- What would the business do with saved time? More output, different responsibilities and fewer hires are different possibilities.
This is a practical reflection exercise, not a scoring system that predicts unemployment. It helps identify which parts of your work you could improve and which capabilities you need to demonstrate.
Which skills are worth developing in the AI era?
My approach would be to improve both tool use and the ability to deliver a reliable result.
Use AI on a real workflow. Pick a bounded task, define what a good result looks like and check what the tool changes. Generating more output is useful only when the output helps finish the job.
Strengthen the knowledge needed to review it. For frontend work, that includes JavaScript, browser behaviour, accessibility, state, data flow and testing. A tool is easier to direct when you can recognise where its answer is incomplete.
Show evidence of decisions. A portfolio can explain the problem, constraints, trade-offs and verification behind a result. That gives a potential client or employer more to assess than a finished screenshot.
Learn the business context. Understanding why a workflow exists makes it easier to improve the whole process, including the awkward parts that a short specification leaves out.
If you are choosing tools, my ChatGPT and Claude model-selection guide provides a starting point. Learning to use AI can improve your options, but it cannot guarantee employment or compensate for every change in demand.
Frequently asked questions
Will AI take all our jobs?
The cited research does not support that conclusion. Exposure studies, usage data and employment studies answer different questions. The ILO expects changes within jobs to be more common than complete automation of occupations. 1
Will AI replace accountants?
Routine bookkeeping is exposed. An accountant’s work can also include handling exceptions, advising clients and taking responsibility for decisions. These tasks need separate assessment. 8
Is frontend development still worth learning?
If you want to build software, learn the foundations alongside AI tools. Be able to explain how an interface works, debug it and check whether it meets real requirements. No course or tool can promise a job, but those abilities let you evaluate and improve what you build.
Are junior developers more at risk?
Stanford’s US data shows the strongest employment divergence among younger workers in highly exposed occupations. Age is not identical to job seniority, and the findings do not prove that AI caused the change or that every junior role will disappear. 6
Which jobs will AI never replace?
No source here can justify “never”. Physical work, hands-on care and difficult human decisions have different constraints from routine digital tasks. Those differences can provide resilience today without guaranteeing permanent protection.
Does knowing how to use AI make your job safe?
It can help you work and adapt, but employment also depends on what organisations need and choose to fund. I would treat AI skills as part of professional competence, alongside judgement, communication and the ability to deliver.
I still love AI
I want AI to help me build better products and explore ideas that would otherwise take too long. I also want an honest conversation about people trying to earn a living while the expectations around their work change.
Being out of work since June has made this personal. My response is to keep learning, building and evaluating the tools I use, while staying clear about what I know and what I can only suspect.
Sources
- ILO: Global occupational exposure index, 20 May 2025.
- ILO: Occupational examples and practical limits, 29 September 2025.
- ILO: Media, writing and translation, 27 February 2025.
- OECD: AI capabilities and occupational exposure, 26 May 2026.
- World Economic Forum: Employer expectations for 2025–2030, January 2025.
- Stanford: Revised employment and hiring analysis, 12 August 2026.
- Anthropic: Observed exposure and early labour-market evidence, 5 March 2026.
- OpenAI: Jobs-transition framework, 25 April 2026.
- Stanford: Canaries employment dashboard, updated 21 October 2026.
Putting AI to work on your website?
Choosing a model is one part of a working feature. If your business needs a support assistant, content workflow or connection to an existing system, I can help define and build the interface, integrations and checks it needs.