TL;DR (Direct Answer)
AI agents are not coming for "jobs" in general. They are coming for specific tasks, starting at the bottom of the career ladder. Stanford economists found employment among 22-to-25-year-olds in the most AI-exposed occupations is now 19% lower than their peers — driven by weaker hiring, not layoffs. Meanwhile the WEF projects a net gain of ~78 million jobs by 2030. In aggregate, jobs are created; for individuals, careers break. The people who understand where the work is going land well.
I have worked with algorithms for more than twenty years. This is what I see, what the recent data says, and what I would do if I were twenty-three again.
1. The Lowest Rung of the Ladder Is Disappearing
Picture a 23-year-old who graduated this spring. Good grades, a clean CV, sixty applications sent. Most of them were never read by a human. The few answers that came back said the same thing: we are not hiring juniors right now.
This is not an anecdote. It is a measured trend. Stanford economists tracking US payroll data from ADP found that workers aged 22 to 25 in the most AI-exposed occupations now have employment 19% lower than their peers in less exposed jobs. A year earlier, the gap was 13%. The decline comes mainly from weaker hiring, not from layoffs. Companies are not firing the young. They are simply not hiring them (Computing, on the Stanford "Canaries in the Coal Mine" update, August 2026).
Lower employment for 22-25 year-olds in AI-exposed jobs
The other side of the ledger: the World Economic Forum's Future of Jobs Report 2025 projects 170 million new roles and 92 million displaced roles by 2030 — a net gain of about 78 million jobs, with 22% of the global workforce in churn.
Sources: Computing · eWeek summary of the WEF report.
Both statements are true at the same time. In aggregate, jobs are created. For individuals, careers break. The word I use is reshaper. Agents do not remove work. They move it, and they move it fast. The people who understand where it is going land well. The others find the floor missing.
That is why I am writing this. Not to scare you. To give you a map.
2. Agents Have Flaws. So Do We. They Are Different Flaws.
Let me start with the weaknesses of AI agents, because they are real.
They hallucinate. They produce confident, plausible and wrong answers. They leak. In April 2026, Capsule Security showed that prompt injection could pull sensitive corporate data out of Microsoft Copilot Studio and Salesforce Agentforce through public-facing forms. Both vendors patched the flaws. The OECD's incident monitor logged the case and noted that fully closing this class of vulnerability remains difficult (OECD.AI).
It is not isolated. AvePoint's State of AI 2026 survey of 750 enterprise leaders found that 88.4% of organizations had at least one AI-agent security incident in the previous 12 months. Half reported data leakage. And 21.1% could not account for the activity of unsanctioned agents (AI2Work summary, July 2026). That is a vendor survey, so read it as a signal, not a census.
Gartner predicted in June 2025 that over 40% of agentic AI projects would be cancelled by the end of 2027, citing rising costs and unclear business value (The Star). Badly built agents fail. That is normal.
Now look at us. We are slow. We work six or eight productive hours a day, on a good day. We get sick. We take holidays. We have bad weeks. We forget what we read last month. An agent has none of these limits. It answers in minutes. It does not sleep.
Of skilled human work a frontier model can complete in about half the time
METR, early 2026: a frontier model could complete, about half the time, tasks that take a skilled human around 14.5 hours. The confidence interval is wide (6 to 98 hours), the fitted doubling time was about 123 days, and METR warns its test suite is near saturation.
Source: Summary of the METR results. Read it as "the length of work an agent can carry alone is growing fast" — not "AI can do your whole job."
Here is the point most people miss. The result depends on the platform, not only on the model. An agent is a model plus tools, memory, permissions, checks and a feedback loop. The same model, wrapped badly, produces a mess. Wrapped well, it produces a colleague. When someone tells me "I tried AI agents and they don't work," my first question is: on which platform, with which guardrails, and who checks the output?
The honest summary is short. Agents fail in ways that are fast and confident. Humans fail in ways that are slow and tired. The winning setup is neither one alone. It is an agent that works, and a human who verifies.
3. Stop Paying for the Smartest Model
Most people I meet use the most powerful model for everything. They write an email with it. They reformat a table with it. Then they wonder why the bill is high.
The price gap is large. In an April 2026 price survey, DeepSeek V4 Flash cost $0.14 per million input tokens and $0.28 per million output tokens. Claude Opus 4.7 cost $5 and $25. A GPT-5.5 preview cost $5 and $30. That makes the frontier models 36 times more expensive on input and 89 to 107 times more on output (BestHub price survey, data as of 26 April 2026). Prices move every month and differ by source, so check them before you commit. A more recent comparison lists Claude Haiku 5.5 at $0.10 input and $0.50 output, and DeepSeek V4.1 Flash at $0.15 and $0.60 off-peak (Gradually.ai, November 2026).
My rule of thumb, from years of building these systems: about 95% of everyday tasks run perfectly well on a cheap, fast model. That is my estimate from practice, not a published statistic. The expensive model is a specialist. You call it when the task is hard, not when it is Tuesday.
Which model tier do you actually need?
| Task | Model tier |
|---|---|
| Summaries, extraction, classification, translation | Cheap and fast (Haiku, DeepSeek Flash) |
| Drafting, reformatting, routine replies, simple code | Cheap and fast |
| Long documents with strict structure | Mid tier |
| Architecture decisions, legal or financial review, novel research, long autonomous tasks | Frontier |
Two consequences follow. First, AI needs some expertise, or a good advisor. Routing the right task to the right model is a skill, and it saves real money. A small company that does this well can run a team of agents for the price of a few subscriptions. Second, a good agent does the routing for you. It sends the simple work to the cheap model and keeps the expensive one for the moments that need it. That is one of the first things we design into every agent we build at Kormaitech.
4. Two Trends: People Who Get Sharper, and People Who Get Duller
I see two groups forming, and the gap between them widens every month.
The first group uses AI to develop their intelligence. The second group uses it to delegate their thinking. Same tool. Opposite outcomes.
The research is early, so let me be careful. An MIT team followed 54 adults over four months. One group wrote essays with ChatGPT, one with a search engine, one with their own brain. Electrical brain activity and the essays themselves were measured. The ChatGPT group showed lower cognitive engagement, had more trouble quoting their own essays, and felt less ownership of them. The authors call this "cognitive debt." But only 18 people completed the final session, the authors themselves say the findings are preliminary, and outside reviewers argue some effects may come from task design (commentary on the study; preprint). It is a warning, not a verdict.
Use AI as three people: a professor who teaches and quizzes you, a developer whose work you read before you own it, and an executive assistant who carries the repetitive load while you keep the judgment.
— Nicolas Martin, working rule
The Professor
Ask it to teach you, then to quiz you. "Explain this concept, then ask me five questions and do not give the answers until I reply."
The Developer
Let it build, but read what it builds. If you cannot explain the result, you do not own it.
The Executive Assistant
Delegate the repetitive load: scheduling, drafts, research, follow-ups. Keep the judgment.
And one more habit that changed my work: use voice mode to think. Talk to the model for twenty minutes about your job, your doubts, your week. Ask it to challenge you. Ask what you are really trying to achieve, and what you would do if money were not the constraint. Spoken reflection goes deeper than typed notes. It is the cheapest coaching session you will ever have, and it helps you find purposeful aims instead of just productive ones.
5. AI Is Not Only Chatbots. It Is Already Everywhere.
When people say "AI," they picture a chat window. That is the visible 1%. The other 99% is invisible. It sits inside every IT service you use: search, maps, email filters, shopping, banking, and above all social networks.
Look at how accurate your feed has become. It knows what keeps you there. I will be direct: I have become increasingly addicted to scrolling, and I have to force myself to stop. I know it has an impact on my brain. I feel it in my attention span.
Courts are beginning to take this seriously. On 25 March 2026, a Los Angeles jury found Meta and YouTube liable in the first major social media addiction trial, over harm to a young woman who started on YouTube at age six and Instagram at nine. The jury found the companies negligent in the design of their platforms and in failing to warn of the danger to minors. The award was $6 million in total, split between compensatory and punitive damages. Both companies said they will appeal. The plaintiff's lawyer pointed to design features such as infinite scroll, autoplay and notifications (report of the verdict). To be precise: this case concerned minors and product design. It was not a ruling on adults, and it was not a ruling on algorithms as such.
But the principle matters to me. What is bad for kids is probably not good for us. Adults just notice the damage later.
Do not underestimate what an algorithm is. It is not a rule written by a clever engineer. It is a system tuned on billions of examples of human behavior, improved every day, optimized for a goal that is not yours. You cannot judge a suggestion when you do not know what is behind it. The same is true of an answer from a chatbot.
Ask three questions about every recommendation, human or machine: Who built this? What is it optimizing? What does it gain if I say yes?
6. Bubble or Not, AI Is Here to Stay
Let me take the bubble question seriously, because the warnings are serious.
On 28 June 2026, the Bank for International Settlements warned that debt-fuelled spending on AI data centers could raise the risk of a financial crisis if the boom reverses. It compared the situation to the dot-com crash and to the railway mania of the 1840s, and pointed to stress in private credit funds (Telegraph report, via Resultsense). I take this seriously. Some valuations will fall. Some investors will lose money.
Now look at what happened after the dot-com crash. Prices collapsed. The internet stayed. I believe the same will happen with AI, for four reasons.
Why AI survives the correction
- The money is enormous. Analysts count between $803 billion and $828 billion of capital spending in 2026 across Amazon, Alphabet, Meta, Microsoft and SpaceX. Amazon alone expects about $220 billion, and Microsoft about $170 billion (Hargreaves Lansdown, 30 October 2026).
- The physical assets are real. Data centers are among the best protected infrastructures in the world, with layered physical and digital security.
- Demand is not theoretical. The IEA reports that data-center electricity use grew 17% in 2025 and could double by 2030 (Pulse, April 2026).
- It spreads across hundreds of fields. Trillions of dollars are flowing into AI and everything built on it: healthcare, industry, logistics, finance, defense.
Robotics and drones are still small. IDC counted nearly 25,000 humanoid robots shipped in the first half of 2026, up 432% year on year, with Agibot in the lead at about 8,600 units. Counterpoint puts the growth at nearly 300% for the same period. The trackers disagree on the rate, but they agree on the direction (Humanoid.guide, 2 October 2026; Counterpoint Research). My expectation is that robotics and drones will grow rapidly from 2027. Today they are a rounding error. In a few years, they will not be.
My advice to young people
Be where the money is going. Not out of greed, but because that is where jobs, mentors and long-term wealth are built. If you are not close to these areas, wealth is hard to build over a lifetime:
- AI for business, including healthcare
- AI infrastructure: chips, data centers, networks
- Energy infrastructure: power, grids, storage. Every data center is an energy problem first.
- AI for industry: factories, logistics, robotics
- AI and space: SpaceX alone reported $48.1 billion of capital spending in a single quarter
You do not need to be a researcher. You need to be useful next to the machines, and to understand what they do.
7. Superintelligence Will Decide Our Activities, Not the Other Way Round
Now the part I find uncomfortable.
I believe superintelligence will decide what we do, not the reverse. When one system outperforms us at most cognitive tasks, it will allocate the work. Our role will be to carry out the parts it cannot, and to build and supervise the systems around it. This is my belief, not a fact. Smart people disagree.
If I am right, our economic value will diminish somewhat. I do not see how people can be useless and be paid for nothing. Money follows contribution, so we must stay contributors, and the most reliable way is to develop AI-related skills now. I explored the ownership side of this in Who Owns the Machines, Owns the World.
There is a second consequence that I like even less. To understand a superintelligence, we may need augmentation. Brain-computer interfaces are no longer science fiction. As of January 2026, 21 people were enrolled in Neuralink's clinical trials, controlling computers and devices with their thoughts (Benzinga). Today that serves people with paralysis. The ethical and neurological questions are serious, and I do not welcome the idea. But it is one of the trends in motion, and I would rather look at it clearly than pretend it is not there.
8. Patience, Resilience, and a Cool Head
The world looks unstable right now. Wars, economic stress, elections, political fights. It is tempting to blame AI for all of it.
History says otherwise. Every major transition, from steam to electricity to the internet, brought a hard decade. The pain did not come from the technology. It came from the slow adaptation of governments and societies to it: schools that taught the old skills, laws that arrived late, workers left without a path. The same will happen here. Difficult moments are coming, and they will come mostly from our lack of adaptation, not from AI itself.
So I ask for three things.
Patience
Give the new era a few years to settle before making irreversible bets on it.
Resilience
Keep learning. Skills decay faster than diplomas.
Detachment
Stay critical. Fear and hype are both emotions, and emotions are our main weakness as humans. Check the source. Wait a day. Ask who gains if you believe this headline.
What to Do This Week
Your 5-point action plan
- Pick one repetitive task in your work and give it to an agent. Verify the output every time for a month.
- Switch your default to a cheap model. Move up only when the task proves it needs it.
- Learn with AI, do not outsource to it. Ask it to quiz you once a day.
- Cap your feed time. Decide the limit before you open the app.
- Choose your direction. Pick one of the five growth areas above and learn its vocabulary.
Want to put this into practice? Kormaitech builds professional AI agents for small and medium companies. Spengtra teaches you and your team to build them yourselves. Same goal, two doors.
Sources
- Computing: AI is closing the door on entry-level jobs (Stanford "Canaries in the Coal Mine" update, Aug 2026)
- eWeek: 170 Million New Jobs, 92 Million Lost (WEF Future of Jobs Report 2025)
- OECD.AI Incidents Monitor: prompt injection in Microsoft and Salesforce AI agents, 15 April 2026
- AI2Work: AvePoint State of AI 2026, 88% of firms hit an AI agent security incident
- The Star: Gartner, over 40% of agentic AI projects scrapped by 2027 (June 2025)
- METR time-horizon results, summarized (April 2026)
- BestHub: 2026 AI model API prices (data as of 26 April 2026)
- Gradually.ai: Claude Haiku 5.5 vs DeepSeek V4.1 Flash (November 2026)
- PsyPost: ChatGPT and cognitive debt and arXiv preprint 2506.08872
- Standard Media: Jury finds Meta and YouTube liable in social media addiction trial (26 March 2026)
- Resultsense: BIS warns AI boom could trigger a financial crash (29 June 2026)
- Hargreaves Lansdown / Business Insider: hyperscaler capex, 30 October 2026
- Pulse: IEA, data-center power use up 17% in 2025
- Humanoid.guide: Agibot leads H1 2026 humanoid shipments (IDC data) and Counterpoint Research: humanoid shipments H1 2026
- Benzinga: Neuralink, 21 people enrolled (January 2026)
About the Author
Nicolas Martin, Founder of Kormaitech Pvt Ltd. Senior data scientist with more than twenty years of experience working with algorithms, 50+ data science and AI projects, and 500+ professionals trained.