How AI Is Changing Software Jobs: Where to Prepare
AI is changing software jobs by automating parts of coding and support while expanding the work around AI applications, evaluation, operations, and security. The evidence points to pressure on entry-level hiring, but nothing so far shows AI replacing software engineers as an occupation.
This guide explains the historical pattern, the evidence on Klarna and junior developers, and the limits of the employment data. It then maps the work around AI to five preparation domains, relevant certifications, and practice questions that test the technical decisions behind them.
How AI is changing software jobs
Start with the difference between a task and a job. A task is a unit of work; a job combines tasks. Generating code may be part of your job alongside understanding requirements, reviewing changes, connecting systems, and deciding whether a release is ready.
Automating a task changes how much labour the bundle requires. An employer might expand output, change the role, hire fewer people, or remove positions. The task approach helps you ask which part of the work is changing; it does not protect a job from disappearing.
That makes a better career question than asking whether your title is safe: which outcomes can you take responsibility for? If you can explain a requirement, test a proposed solution, and operate it after release, you have more ways to contribute than code generation alone. Treat that as preparation advice, not a prediction about what employers will reward.
The historical examples below show several possible outcomes. Natural-ice harvesting declined commercially, telephone operators declined inside a continuing communications business, and bank tellers initially benefited from branch expansion. None supplies a timetable for software.
When a job vanished while its industry boomed
Telephone operators and the long adoption gap
Almon Strowger, a Kansas City undertaker, suspected operators were diverting calls intended for him to a rival. He patented an automatic exchange in 1891, and the first commercial automatic exchange opened in La Porte, Indiana, in 1892 with about 75 subscribers. His suspicion belongs to the history of the invention; it is not evidence that the diversion actually happened.
The existence of the machine did not immediately end operator work. The Richmond Fed's history of telephone operators reports about 178,000 US operators in 1920, about 342,000 around mid-century, and about 40,000 in 1984. These are the operator counts the history reports, not a consistently Bell-only employment series.
Meanwhile, the Bell System employed about 1,009,000 people at its 1984 breakup, according to its historical booklet. A large communications business could coexist with a shrinking occupation. For software engineers, the relevant analogy is that demand for a service can remain strong while the work used to deliver it changes.
Who absorbed the loss?
Feigenbaum and Gross's research examines cities where AT&T automated exchanges between 1920 and 1940. Operator employment for women aged 16–25 fell by 50–80% in affected cities. Later cohorts' overall employment held up, though some moved into lower-paid work.
The outcome differed for established operators. Those over 25 at baseline were about seven percentage points less likely to be working a decade later than comparable operators in cities that had not switched. Some career changes led to lower-paying occupations.
The lesson is about distribution: an economy can adjust while particular workers lose income or careers. James Bessen also describes fewer switchboard operators and more receptionists as automated phone systems spread. That is growth in a neighbouring occupation, without proof that the displaced individuals all moved into it.
Other jobs that moved
Natural ice: the need survived the business model
Natural ice lost ground before household refrigerators became widespread. In Chicago, its share of the ice market fell from 82% in 1911 to 10% in 1922 as machine-made ice displaced it, according to the Chicago Public Library's history. The customer still wanted cooling; the method of supplying it changed.
The National Park Service's Fresh Pond account describes refrigerators undermining ice harvesting during the 1930s and a fire ending the company in 1935. Refrigerators did not cause the fire. The account supports commercial decline and a particular company's closure, not worldwide extinction of every related job.
Bank tellers: expansion can follow automation
In Bessen's account of ATMs, tellers per average urban branch fell from 20 to 13 between 1988 and 2004. Urban branches grew 43% as branches became cheaper to operate. His research on occupational automation reports teller employment, measured in full-time equivalents, growing about 2% annually from 2000 to 2013.
That historical expansion is not a permanent employment rule. The BLS teller outlook uses about 339,200 tellers as its 2025 baseline and forecasts 294,500 by 2035. The forecast points down, but those figures alone do not identify a single cause.
Typesetting and the limits of reassurance
Bessen describes desktop publishing alongside fewer typographers and more graphic designers. Research summarised by MIT on new work finds that about 60% of US employment in 2018 was in occupational specialties introduced since 1940. New kinds of work can become a large part of employment.
The same research finds stronger displacement effects since 1980, with new work shifting toward highly paid professions and lower-paid services. It provides no guarantee of equivalent opportunities for everyone displaced. For your own planning, follow the changing work while keeping the costs of transition visible.
What AI is doing now
What Klarna's customer-service claim actually meant
In its February 2024 announcement, Klarna said its AI assistant handled two-thirds of customer-service chats in its first month. It described that output as equivalent to the work of 700 full-time agents. This was a company-published output comparison, not a verified count of people dismissed.
In September 2025, Reuters reported that Klarna's chief executive acknowledged overemphasising cost-cutting and that human hiring had resumed. That doesn't mean the same workers returned, or that the company abandoned AI. For a software project, it suggests a practical review question: what outcome does the automation improve, beyond the volume of work processed?
Junior developers and a narrower entry route
Stanford's Canaries in the Coal Mine research, revised in August 2026 using ADP payroll data through June 2026, finds a relative employment gap of 19% for workers aged 22–25 in highly AI-exposed occupations compared with less-exposed peers. The gap mainly reflects reduced hiring. It is descriptive evidence, not proof that AI caused the difference, and the population extends beyond software developers.
The researchers discuss codified knowledge learned through education versus tacit knowledge developed through experience as a possible explanation. That mechanism remains a hypothesis. An implication for your preparation is to practise decisions whose consequences you can inspect: review a change, investigate a failure, and explain why a solution fits its requirements.
Indeed's July 2025 analysis found US technology postings for standard and junior titles combined down 34% against the pre-pandemic baseline, compared with 19% for senior and management titles. Across technology postings, the share asking for at least five years of experience rose from 37% to 42% between the second quarters of 2022 and 2025. These are posting statistics, not a count of junior developers losing employment.
Coding productivity depends on what was tested
A Copilot experiment published in 2023 found developers completed a particular HTTP-server task 55.8% faster with assistance. The experiment involved 95 participants and a single bounded task. It says little about the speedup across an entire engineering job.
METR's July 2025 experiment found experienced open-source developers took 19% longer with AI tools on familiar codebases. Afterwards, participants believed the tools had made them about 20% faster. Different tools, tasks, and participants produced different results; neither study gives you a universal rule.
Use that uncertainty to improve your own measurement. When you try assistance, include review and correction in the completion time, and judge whether the result meets the requirement. The feeling of moving faster deserves checking against finished work.
Freelance demand and support productivity are separate evidence
Demirci and colleagues found 21% fewer posts for automation-prone writing and coding freelance work relative to manual-intensive freelance posts within eight months of ChatGPT. Separately, Hui and colleagues' Upwork research found writing freelancers had 2% fewer jobs and 5.2% lower earnings after ChatGPT.
The 2023 working paper on AI-assisted support reported productivity gains of 14% on average and 34% for novices. That measures output from assisted workers. It doesn't show that employers hired fewer beginners as a result.
What the data does and doesn't show
Before using a statistic to plan your career, identify its denominator and what it measures. These comparisons keep different kinds of evidence separate.
- Exposure: overlap between AI capabilities and occupational tasks. / Displacement: work or employment actually removed.
- Productivity: output relative to the resources used. / Employment: people or positions employed, which may respond differently from output.
- Job postings: advertised demand on a measured platform. / Filled jobs: employment that postings alone can't show.
- Forecast: a conditional view of future employment. / Observation: a measurement of a stated population during a stated period.
The hiring cycle also matters. Indeed's July 2026 analysis put US software-development postings 27.5% below February 2020, despite a rebound of about 15% since February 2025. Much of the decline reflects the correction after the pandemic hiring boom, so attributing the whole change to AI would overstate what the evidence shows.
The ILO's 2025 assessment places one in four jobs worldwide in an occupation with some generative AI exposure and forecasts that most jobs will be transformed rather than made redundant. Exposure is a reason to examine your tasks. It is not a probability that your whole job will disappear.
Platform boundaries matter too. Anthropic classified 52% of Claude.ai conversations in its November 2025 data as augmentation and 45% as automation, while automated use dominated API traffic in the same economic report. Those are observations about its products, not the whole labour market.
Forecasts need equally careful labels. The World Economic Forum's employer-survey forecast projects 170 million jobs created and 92 million displaced by 2030 across technology, economic change, demographics, and the green transition. Calling them all AI jobs would change the meaning of the forecast.
New jobs around AI
LinkedIn placed AI engineer first on its 2026 US Jobs on the Rise list. This is a platform-based ranking. Application work around models includes retrieval, tool use, evaluation, and integration, which gives software engineers concrete capabilities to study.
Forward-deployed engineers work directly with customer teams to make systems useful in their business setting. Lightcast's posting analysis counted about 200 US postings in 2024, about 1,200 in 2025, and more than 5,200 during January–July 2026. The observation windows differ, and postings are not completed hires.
The surrounding work also extends beyond application development. LinkedIn's data-centre workforce report estimates more than 600,000 net new positions globally in 2025, including technicians. An Oxford Economics study funded by Scale AI estimated about 200,000 flexible US earning opportunities in data annotation in 2024, as described in Scale's published account. Flexible opportunities should not be read as an equivalent number of permanent careers.
For software developers themselves, the BLS outlook uses 1,717,800 US developers in 2025 and forecasts 1,892,600 in 2035, or 10% growth. Those figures cover developers alone, excluding the separate QA analyst and tester categories in the combined headline. A positive forecast can coexist with difficult entry-level hiring.
Five domains to prepare for
Build applications with models
Learn retrieval-augmented generation, or RAG: retrieve relevant information and provide it to a model as context for its answer. On AWS, Amazon Bedrock Knowledge Bases provides retrieval for applications using your information. Practise tracing an answer back to the material that supports it.
Then study agents that use tools to take steps in a workflow. Amazon Bedrock AgentCore provides a platform for deploying and operating agents securely. In a learning project, define what each tool may do and what should happen when information is missing.
Evaluate the result
Learn to judge retrieval relevance, answer correctness, and faithfulness to source material. Keep representative cases so you can compare changes rather than trusting an impressive demonstration. Include cases where the material does not support an answer, and decide how you want the application to respond.
In the Stack Overflow 2025 survey, 84% of respondents used or planned to use AI tools, while 46% distrusted their accuracy and only 3% highly trusted them. These are survey responses, not measured defect rates. They make checking output a useful skill to practise without establishing a count of evaluation jobs.
Operate the system
Study deployment, monitoring, cost control, and the model lifecycle. MLOps and LLMOps connect changes to the evidence needed to release and run them. Your project should make it possible to explain what changed and investigate a worse result.
Keep operational health and answer quality visible separately. A responsive application can still retrieve irrelevant material or produce an unsupported answer. Practise tracing the failure to the component that needs attention before changing the model.
Secure and govern it
Learn access control, sensitive-data handling, and the security of the containers and software supply chain underneath the application. Amazon Bedrock Guardrails helps detect and filter harmful or sensitive content. It reduces particular risks; it does not guarantee that an application's output is safe or correct.
For a learning project, document who may use the system, what information it may access, and who reviews failures. Treat content filtering and permission to perform an action as separate decisions. A harmless-looking answer can still come with an unauthorised tool call.
Understand a business problem
Choose a workflow whose successful outcome you can explain in ordinary language. Write down what the user needs, where mistakes matter, and when a person should take over. Use those requirements to decide which parts benefit from AI.
This is the customer-facing skill suggested by forward-deployed engineering, not a promise that a new title will protect your career. Build evidence of decisions you made and results you checked. That gives you something concrete to discuss alongside a certification.
Certifications and practice for each domain
Choose a track that matches the work you want to practise. The AI certification comparison helps you compare paths, and the agentic AI certification guide focuses on agent systems. Use the objectives as a study structure and connect them to a working project.
For application building, Azure AI Apps and Agents Developer Associate covers retrieval, evaluations, orchestration, and observability in its official AI-103 objectives. Use the AI-103 practice questions to check whether you can select an approach from requirements.
For AI concepts and evaluation, AWS Certified AI Practitioner includes foundation-model evaluation within its applications domain, alongside security and governance elsewhere in the outline. Work through the AIF-C01 practice questions, explaining why plausible alternatives fail the requirement.
For operations, Machine Learning Operations Engineer Associate covers MLOps, GenAIOps, quality assurance, and observability in the AI-300 study guide. Pair the AI-300 practice questions with the MLOps certification comparison when choosing an operations path. Google Professional Machine Learning Engineer covers serving, pipelines and monitoring on its official certification page; use the PMLE practice questions for production-ML decisions.
For supporting cloud infrastructure, AWS Solutions Architect Associate addresses secure, resilient, performant, and cost-conscious designs in its official outline. The SAA-C03 practice questions exercise those choices. AWS DevOps Engineer Professional adds delivery automation and deployment strategies described in its SDLC automation objectives; the DOP-C02 practice questions focus your release decisions.
For container security, Certified Kubernetes Security Specialist covers supply-chain and runtime security. The Linux Foundation's requirements include passing CKA before attempting CKS. Use the CKS practice questions to support that security study; the business-understanding domain still needs direct practice with requirements and outcomes.
Exam traps: choose the capability the requirement needs
Agent hosting versus specialised modernisation
For legacy COBOL and .NET modernisation, AWS Transform provides specialised discovery, planning, and transformation capabilities. AgentCore supplies a platform for agents you build; SageMaker AI serves model-development workflows; Kiro is an agentic development environment. The word agent does not make these capabilities interchangeable.
Try a practice question on legacy workload modernisation. Identify whether the requirement asks you to create an agent or use a service with existing modernisation expertise. Then explain what evidence would be needed to accept its output in a real migration.
Retrieval versus retraining
When source documents change, ask how the application gets the current information and shows its origin. RAG retrieves external material for the prompt, so updating indexed documents does not require retraining the model. A detailed instruction alone does not supply missing facts.
A practice question on changing internal policies tests that distinction. Separate access to current knowledge from changing the model's learned behaviour before choosing an answer.
Quality evaluation versus operational metrics
A system can respond quickly while returning poor answers. Retrieval and generation therefore need checks suited to their outputs, alongside operational monitoring. AWS describes those checks in its RAG evaluation documentation.
Use a practice question on evaluating retrieval and generated answers to distinguish a whole RAG evaluation from measuring only the model or response latency. Explain which part of the application each proposed measurement leaves untested.
Content safeguards versus permission boundaries
Content filtering helps address sensitive information and prohibited topics. Access control decides who may reach information or invoke a capability. Keeping those responsibilities clear helps you avoid choosing a control simply because it sounds security-related.
A practice question on chatbot safeguards and personal data examines content controls, while a practice question on cross-account role access examines access without distributing long-lived credentials. For an AI application, practise describing the content risk and the access risk independently.
Deployment success versus recovery from failure
Sending a small share of traffic to a change provides a chance to observe it. Recovery also needs a defined signal and a configured response. A staged rollout alone does not explain what happens when the new version fails.
A practice question on canary deployment and automatic rollback connects traffic shifting to alarms and rollback. Apply the reasoning to your preparation: state what you would measure after a release and what result would make you reverse it.
Frequently asked questions
Will AI replace software engineers?
Not on the evidence so far. It shows changing tasks, uneven productivity results, and pressure on some hiring routes. The BLS developer-growth projection is a forecast, so use it alongside the observed evidence rather than as a guarantee.
Are junior developer jobs disappearing?
Entry-level access appears under pressure, but the measurements cover different populations. Stanford's relative gap concerns young workers across exposed occupations; Indeed combines standard and junior technology titles. Neither establishes that every junior development role is disappearing.
Did Klarna replace 700 employees with AI?
Klarna claimed output equivalent to 700 full-time agents. That was not a verified layoff count, and its later return to human hiring does not prove that the same workers returned.
What should a software engineer learn now?
Practise building with models, evaluating results, operating systems, securing them, and understanding business requirements. Choose a project where you can explain the decisions and inspect the outcome, then use a relevant certification outline to organise further study.
Which AI jobs offer a reliable career path?
No evidence here makes a particular title reliable for everyone. AI engineering and customer-facing deployment work show growth in platform observations, while operations and security supply relevant skills. Evaluate the actual responsibilities and working conditions behind an advertised title.
Watch the full episode
The conversation makes the personal cost of each transition easier to hear, including the tension between an improving economy and a worker losing their route into it. Follow the historical comparisons and preparation discussion in AI Hit Junior Devs First. Where Should You Stand? on YouTube.