AI Did Not Cause the Crisis. It Made It Visible.

The following seven points summarize the backbone of the article. Behind each one is the detailed discussion and the cited source.

AI did not cause the crisis. It made it visible. A market overview of the current state of the Hungarian IT profession, of the kind that does not tally complaints but looks for patterns.

Introduction: scattered complaints or a market pattern?

If you read through a few Hungarian IT conversations lately (Reddit threads, forum posts, professional groups), the same motifs come back again and again. Entering as a junior is hard. There is pressure because of AI. The senior title means less and less. LinkedIn and recruiter outreach are noisy and not credible. Wages stagnate or grow below inflation. Interview processes have many rounds and are unpredictable. Transparency is weak. And from many people there flows a hard-to-pin-down decline in morale: the joy of the work and professional status have both been damaged.

One commenter summed up the mood in a single sentence: it is as if the whole profession had been slapped in the face. At first, these may look like scattered, individual complaints. Read together, however, they add up to something else: a market pattern in which several independent forces meet in the same time window.

METHODOLOGICAL NOTE

This article does not treat these conversations as representative research but as a qualitative market thermometer. The threads show mood, recurring narratives, fears and perceptions, not statistics. The aim is to read them together with international data and research to give a market overview. Throughout the text, I distinguish what is supported by data, what we see from market conversations, and what is merely an analyst's hypothesis.

The main thesis is simple: the Hungarian IT profession's sense of crisis cannot be explained by a single factor. AI is only one accelerator, if a very visible one. At the same time, there is the economic slowdown, the cooling of global tech hiring, the productivity expectations created by AI, the automation of junior tasks, the inflation of seniority titles, the vulnerability of the service center model, wage stagnation, weak recruitment transparency, and the long-unresolved uncertainty of measuring knowledge work. This article is neither anti-AI nor techno-optimist. AI really does speed up certain kinds of work; some developers really do become more productive with it; entry-level and routine tasks, however, become more vulnerable; and in bad organizations, AI does not solve existing problems but magnifies them.

The current mood can only be understood in light of the previous decade. The 2010s and the post-Covid period were an overheated era of tech hiring: a wave of digitalization, service centers opening one after another, outsourcing, remote work, and a labor market in which a developer's visibility almost guaranteed the attention of demand. This period ended with the interest rate environment after 2022. Capital became more expensive, investors started to expect returns, and the tech sector learned the language of cost discipline instead of growth.

The numbers support this. According to Challenger, Gray & Christmas, there were more than 806 thousand announced layoffs in the United States in 2025, the highest figure since 2020, and the tech sector was hit hardest, with more than 89 thousand layoffs on its own. This is not a single bad quarter but a recalibration that has been going on for several years. It is important, however, to read the numbers in context: only part of these layoffs can be tied directly to AI. A larger share is the result of the interest rate environment, the correction of excessive pandemic hiring and general cost discipline.

Hungary played a particular role in this system: it was a cost-optimized but high-quality delivery location. A workforce that was cheaper than in Western markets but well enough trained made it an ideal target for global companies. This model works well exactly as long as the difference between the Western "expensive core" and the Eastern European "cheaper delivery" is large enough. If the gap narrows (a strong forint, slowing real wages, or a technology that lets fewer people do the same work), the model becomes vulnerable.

The Hungarian macro picture adds nuance. According to the Hungarian Central Statistical Office (KSH), consumer prices rose by an average of 4.4% in 2025, while by early 2026 inflation had already sunk close to 2%, in an election year. Industrial production, however, was 3.2% below the previous year: output grew in only two of the 13 manufacturing subsectors, most visibly in the manufacture of computer, electronic and optical products (+14.5%). The picture is therefore not uniformly dark: there are technological islands with strong momentum. But the broader labor market mood is gloomy.

DATA

Hays Hungary Salary Guide 2026: more than 70% of Hungarian labor market participants are pessimistic about the next 2-5 years, and only one in ten is optimistic. Wages grew noticeably more slowly in 2025: the most common raise was 6-10%, but many received only 2.5-5%, and the share of raises above 10% fell drastically.

It is worth asking the question the market itself keeps raising: is this a crisis, or merely the normalization of an earlier, unnaturally overheated period? The honest answer: partly both. Compared with the pace of 2021, this is undoubtedly a cooldown. Measured against a longer historical average, though, it is more of a return to earth. The problem is that the cooldown and the structural realignment caused by AI arrived at the same time, and the two together are much more unsettling than either would be separately.

We misunderstand AI's role most when we either blame it for everything or expect it to save us from everything. The more accurate reading: AI does not cause the sense of crisis by itself, but it accelerates every earlier pressure. Where the junior market was already tense, it is now more visibly tense. Where developer work could not be measured before, it is now measured badly even faster.

The background is real and large in scale. According to Stanford's 2025 AI Index, corporate AI investment reached 252 billion dollars globally, and organizational adoption jumped to 78% in a single year (from 55% before). Companies are therefore not only curious about AI but also expect productivity gains from it, and this expectation by itself rewrites hiring and pay decisions.

The employee experience, however, is twofold. In Stack Overflow's 2025 developer survey, 84% of respondents use or plan to use AI tools (76% a year earlier), so adoption is beyond dispute. Yet trust collapsed at the same time: 46% of developers do not trust the accuracy of AI output (a year earlier only 31% were this skeptical), and only 3% say they "highly trust" it. The most experienced developers are the most cautious. The most common frustration (for 45% of respondents) is the "almost right, but not quite" solution, which makes debugging more time-consuming; 66% of developers spend more time fixing almost-right AI code. Seven in ten would still rather ask a human for help when they do not trust the AI's answer.

CONTRADICTORY EVIDENCE

The optimistic side, PwC 2025 Global AI Jobs Barometer: in the industries most exposed to AI (finance, software publishing), productivity growth has nearly quadrupled since 2022, and AI skills carry an average wage premium of 56% (25% a year earlier).

The skeptical side, METR randomized controlled trial (2025): 16 experienced open source developers, 246 tasks, on mature codebases. The developers expected a 24% speedup from AI and afterwards still estimated 20%. In reality they were 19% slower with it. Reasons: low reliability (fewer than 44% of suggestions were accepted), the burden of validation, the cost of context switching.

Neither figure is false, and they do not necessarily contradict each other. PwC is an industry-level analysis based on about one billion job postings; METR is a narrow, deep engineering experiment on mature codebases. The authors themselves stress METR's limitations: a small sample, developers who knew their own projects very well, and tools from early 2025, since when coding agents have improved a great deal. Still, the two results warn of the same thing: AI does not speed things up evenly. On new, well-scoped tasks, boilerplate and prototypes, the gain is large; in mature, complex systems with high validation demands, the benefit can disappear or even turn negative.

This brings an important distinction to the center, one that has also crystallized in market conversations: the difference between the AI-literate engineer and the vibe coder. The former knows what they want, uses AI as a tool, reviews, validates and takes responsibility for their code. The latter accepts what AI generates without understanding it. In the Stack Overflow survey, 72% of developers do not call their work vibe coding, so most of the profession is already instinctively defending this line. Management, however, often cannot see the difference between useful AI assistance and uncontrolled slop production. And here is the key: in bad organizations AI does not solve existing problems but amplifies them, while in good organizations it is real engineering leverage.

If there is one point where market mood and hard data agree most, it is the junior market. Entry-level positions have become scarce, and AI is automating or speeding up precisely the tasks on which juniors used to learn: the old "droid work," simple tickets, CRUD, writing tests, documentation.

According to SignalFire's analysis, the share of new graduates at big tech companies fell to just 7% of new hires in 2024: a 25% drop in a single year, and a decline of more than 50% from the pre-pandemic level of 2019. At startups, the share of new graduates plunged from 30% (2019) to below 6%. A Harvard study based on 62 million LinkedIn profiles and 200 million job postings identifies "seniority-biased" change: at companies adopting generative AI, junior hiring falls sharply while the level of senior hiring stays stable, and this is driven mainly by slower hiring, not by firings. A Stanford working paper documents a 13% relative decline in employment for 22-25-year-olds in AI-exposed occupations.

NUMBERS IN CONTEXT: HANDLE WITH CARE!

According to US BLS data, employment in the "programmer" occupation fell by 27.5% between 2023 and 2025, while in the "software developer" category it fell by only 0.3%. The two numbers do not contradict each other: they refer to different occupational categories. This is a good reminder not to build a whole narrative on a single dramatic number; category boundaries and definitions matter a great deal.

This leads to a structural paradox: the market wants seniors while dismantling the training ground from which seniors emerge. If nobody hires juniors, where will the seniors of five years from now come from? In the short term, though, fewer but more experienced people plus an AI assistant is more tempting for companies. This is not an exclusively Hungarian problem (Indian IT service providers have cut entry-level roles by 20-25%), but it may hurt especially in Hungary, precisely because of the cost-optimized delivery legacy.

Market conversations also shed light on a deeper layer that cannot be seen from employment data alone: learning and delivery are coming apart. The juniors and mid-level developers who are inside do not necessarily develop healthily either, because the workplace does not say "understand it, build professional intuition," but "get it done, with AI, fast." In the old model, the junior learned the system through boring but instructive tasks. If we automate precisely these, what remains is fast delivery, and deep understanding falls away. This is not simply that "juniors can't get jobs"; it is a crisis in the profession's pipeline of new talent.

Expectations are also being structurally rearranged. According to PwC, the skills sought in AI-exposed occupations are changing 66% faster than before, and the demand for formal degrees is eroding (for AI-augmented jobs, the share requiring a degree fell from 66% to 59%). According to the WEF Future of Jobs 2025, 39% of key skills will change by 2030. The old career ladder of junior, mid-level, senior is not disappearing but compressing: more autonomy, communication and judgment are expected at the entry level. Put plainly: the junior role is now partly asked for what used to be asked of a mid-level developer.

If the entry level slides upward, the top of the hierarchy does not remain untouched either. A recurring realization in market conversations is that in many places the "senior" title is no longer a professional guarantee but a pay band and a bargaining label. A company calls someone senior because it places them in a certain salary category, not necessarily because the person is responsible for systems at production level.

This has two directions, and both are worth taking seriously. On the one hand, real title inflation is under way: the old senior is in many places now a staff, principal or architect, because "senior" has been devalued. On the other hand, there are young seniors who really are partly on a faster learning path: with AI, better learning materials and denser feedback, knowledge that used to take longer can be gathered in three or four years. The two phenomena mix, and that is exactly why it is hard to read anything from a title.

HUNGARIAN PAY BANDS 2026 (MONTHLY GROSS, INDICATIVE BENCHMARK)

Junior developer: approx. HUF 700,000-1,000,000 · Senior developer: HUF 1,750,000-2,100,000

Senior Data Scientist / Data Engineer: approx. HUF 2,000,000 · Senior ML Engineer: approx. HUF 2,100,000

Lead / Architect: approx. HUF 2,300,000-2,500,000

Source: Hays Salary Guide 2026 and the Fizetesek.com 2026 benchmark. For comparison: according to KSH, the gross median full-time earnings in December 2025 were HUF 607,700; ICT is among the sectors with outstanding earnings.

The big jumps between pay bands themselves encourage title inflation: with a more than twofold difference between the junior and senior bands, there is plenty of room for companies to maneuver with the title instead of the pay, or the other way around. This does not solve the underlying problem, however: the number of years is a bad metric. Real seniority is not the same as time served. Ownership, a feel for tradeoffs, production responsibility, stakeholder management, mentoring, deep knowledge of systems: these are the actual signals. A senior who does not write code but prevents a bad system from being built is often worth more than one who commits a lot. The trouble is not the inflation itself, but that the market has no good language for expressing real seniority, and AI only amplifies this gap, because the value of mere code writing falls while the value of decisions rises.

The labor market has become not only narrower but also noisier. One of the angriest strands of market conversations is precisely about the breakdown of the job search infrastructure. Postings tagged as junior that turn out halfway through to be looking for a senior. Outreach without a salary band. Keyword-based recruiter outreach that serves reach, not the candidate. The LinkedIn filter, even set to junior, brings up director and architect positions requiring 10+ years.

This places a hidden cost on the candidate: with few real entry-level positions, they first have to dig out which posting is not lying to them. The search cost rises and the candidate's sense of bargaining power deteriorates, especially because many approaches carry the background threat that if you do not take it, the cheaper offshore or AI alternative will. In market conversations this tone is often stereotyped and angry, but the underlying economic feeling is clear: Hungarian IT people feel less and less that they are in a strong negotiating position.

Here is the paradox: the profession looks down on the visibility game, while the labor market is increasingly turning into a visibility game. Many consider LinkedIn cringe, corporate self-promotion, and they are not entirely wrong. But given the state of the market, some minimal presence is unavoidable. A healthier compromise also surfaces in market conversations, and it echoes the broader trend toward transparency: not daily empty posting, but your own professional trail with real evidential weight (a technical case study, a blog or Substack article), which you then share briefly. You do not need to be an influencer, but you do need to leave a trail.

MARKET SIGNAL + DATA

According to Hays Hungary, pay transparency in Hungary is still in its infancy: 61% of employees say there has been no meaningful progress. This will soon change, however: following the EU pay transparency directive, mandatory pay transparency comes into force in Hungary in the summer of 2026, which may rewrite expectations about salary bands, career paths and job postings. Candidates are thus getting more structural bargaining power right now, while the market has otherwise tilted against them.

There is an uncomfortable truth worth stating: we could not measure developer work well even before AI. Hours worked, story points, ticket counts, lines of code and meeting attendance are all problematic metrics, because the value of knowledge work rarely shows directly in activity. AI exploded into this already uncertain measurement environment.

Now come the new, temptingly quantifiable but misleading metrics: token usage, the number of AI-generated PRs, tickets run in parallel with agents, the amount of generated code, "how much did you close." These easily create productivity theater: the metric rises while the real value does not necessarily follow. Quality work, moreover, is often invisible: good design, less code, the avoided bug, the precise specification, the good test, the well-considered review, the question asked at the right moment are all hard to quantify.

WHAT HAS TO BE KEPT APART

Activity (how much you do) → output (how much gets made) → outcome (what it achieves) → business value → reliability. AI makes activity and output cheap and visible, which is exactly the least useful thing to measure.

The lesson of the METR experiment haunts us here too: the developers misjudged their own productivity, believing they were faster while they were slowing down. If an organization did not understand the difference between activity and value creation before, AI only makes the mismeasurement faster. More thoughtful frameworks exist in the profession (the DORA and SPACE metrics encourage looking at several dimensions together: lead time, reliability, satisfaction, cognitive load), but they spread slowly, because they are less comfortable than a simple ticket count.

Market conversations also bring a surreal side phenomenon to the surface: the very concept of "working time" is falling apart. Many talk about between zero and four or five hours of actual focus time a day, and the debate quickly turns to what we even call work: the meeting, planning over coffee, waiting for the agent, knowledge picked up earlier? This is not laziness but a symptom of the nature of knowledge work. And this is exactly what makes it so dangerous when someone tries to measure value in token consumption or the amount of generated code.

One of the most concrete sources of declining morale is pay. A recurring picture in market conversations: raises below inflation, games between bonus and base salary, and the feeling that the hub creates value while the employee gets a 1-3% raise. Reactions are divided (many say "that's the market, go elsewhere"), but the frustration is real.

Hungarian macro data explain why this is coming to a head right now. According to KSH, net real earnings grew by an average of 9% in 2024, an exceptionally fast pace since the change of regime. In 2025, however, the tables turned: one in four Hungarian workers (28% of employees) found that their pay was worth less in the shops than a year earlier. According to KSH's commentary, this was mainly due to a decrease in benefits beyond base pay, such as allowances and bonuses. Put plainly: companies did not necessarily cut the pay scale, but the bonus and the extras, which is exactly the experience that comes through in the conversations. The share of those in the 0-5% raise band rose from 16% to 28%.

MARKET SIGNAL

The 2025 setback after the exceptional real wage growth of 2024 is psychologically sharper than continuous stagnation would have been: people got used to improvement and were then suddenly confronted with its end. This contrast by itself damages morale, even if inflation has meanwhile eased.

And this is where the pay question meets AI. According to PwC, AI skills carry a 56% wage premium, but that is a global average, not a Hungarian guarantee, and it rewards those who bring AI skills to the market, not everyone. The rawer danger is this: AI raises the expected output, but it is not certain that pay rises. If the employee feels only a faster pace, more context switching and a stagnating salary, then AI is not a liberating tool but a morale-destroying pressure.

The key question, then, can neither be avoided nor answered with confidence: who wins AI's productivity gains, the employee, the company, the customer or the shareholder? This is a question of distribution, not of technology. PwC's data suggest that some employees with AI skills really do realize a wage premium. But McKinsey's data also show that most companies do not yet turn AI into enterprise-level profit (more on this in Chapter 8), so the "winnings" to be shared are in many places still more promise than reality. Anyone who claims to know the answer for certain is overestimating the current evidence.

The labor market impact of AI has an underestimated dimension: pricing. The current public mood assumes that AI is essentially free and unlimited. In the short term this is partly true (free access will probably remain), but professional, intensive AI use is increasingly becoming usage-based, quota-limited or enterprise-controlled.

The most telling signal is GitHub Copilot's pricing model. Billing for premium requests started in June 2025, and from 1 June 2026 all Copilot plans switch to token-based, usage-based billing: "premium request units" are replaced by GitHub AI Credits, consumption is calculated from input and output tokens, and requests above the limit cost extra (0.04 dollars per request in the earlier model). The key lies in the model multipliers: advanced reasoning models can consume the allowance at 5× or 20×, and some top models in fast mode at up to 30×. The earlier "fallback," when you were automatically switched to a cheaper model once the quota ran out, is going away; instead, the allowance and the administrator's budget set the limit.

THE DOUBLE MOVEMENT

The price of inference is plummeting: according to the Stanford AI Index, running a GPT-3.5-level system became 280 times cheaper between late 2022 and late 2024 (from about 20 dollars to 0.07 dollars per million tokens). Training top models, on the other hand, is getting more expensive: training the newest systems sometimes costs more than 100 million dollars, which only the tech giants can afford. Commodity inference is thus cheap and getting cheaper, while heavy, long-session, large-context, agentic workloads become expensive exactly where the real engineering leverage is.

The labor market consequence of this is important and somewhat surprising: if serious AI leverage costs money, then it does not replace everyone; rather, it increases the efficiency of good professionals. In a world of cheap, unlimited AI, it is easy to imagine that "everyone is replaceable." In a world of metered, costly AI, what matters is who can use the expensive resource with the greatest business value: who can do good prompting, good architecture, good evals and good cost control. This raises the value of engineering judgment rather than lowering it.

The strategic conclusion is the same at the individual and company level: LLM-agnostic systems, model routing (the model that fits the task, not always the most expensive one), evals, cost control and governance. Whoever understands this not only increases their own productivity, but builds exactly the layer that AI does not commoditize.

Global trends affect every market, but in Hungary there are a few factors that amplify their effect or cast it in a different light. The most important is the service center and nearshore legacy mentioned earlier. Much of the work came here because Hungarian labor was cheaper than in London, Zurich or Germany, but of good enough quality. This cost-arbitrage model is structurally vulnerable: it works exactly as long as the cost difference is large.

On top of this comes the strong forint. Both macro data and market sentiment show that the strengthening of the forint makes Hungarian labor relatively more expensive in euro terms for foreign companies, just when global cost discipline is stricter anyway. (Market conversations also mention a concrete 20-25% increase in cost; this is best treated as a qualitative market reading, not a precise metric.) If Hungarian delivery is no longer noticeably cheaper, it is not enough to be a "good coder"; something extra that is harder to replace is needed: domain knowledge, trust, ownership, compliance, local and European context, end-to-end responsibility.

On the demand side there is another Hungarian tension. A recurring realization in market conversations is that the problem is not that there is nothing to digitalize; there is plenty of technical need. The problem is that demand that can pay and can decide is scarce. SMEs and institutions wait, haggle, want cheaper development by pointing to AI, or try to tinker internally. It must be added that there is no strong, public Hungarian data series on this, so I put it forward as an analyst's hypothesis, not a proven fact. But the picture is consistent with what we know about the economic slowdown and caution in investment.

ANALYST'S HYPOTHESIS

Because of the Hungarian market's small size, global tech and AI trends spill over faster and more sharply than in larger, more diversified economies. A narrow market has less cushioning: if a few large hubs or a few dominant clients change their hiring or pricing logic, the whole profession feels it immediately. This means faster pain, but potentially faster adaptation too.

Finally, there is the slowness of state and large corporate systems, the remote/hybrid realignment and the uncertainty of the contractor market. Hungarian contractors and body shops are particularly exposed: in-house senior-plus-AI teams can take back work that was previously outsourced. In sum: Hungary is not "finished," but it has to step up from the plain delivery model toward domain, production and decision value, and the small market gives less time to do so.

The picture is harsh, but not hopeless, and certainly not apocalyptic. The realignment does not mean that the profession is finished; it means that what counts as value is shifting. Mere code writing is losing value, while specification, architecture, product judgment, domain knowledge and production responsibility are gaining. The winning profiles follow this axis.

THE COMMON THREAD

The winner is not the classic "coder," but the one who can learn, uses AI in a controlled way, understands business problems, builds systems, validates, communicates, and leaves a visible professional trail behind. This is the layer that AI does not commoditize but amplifies.

The "slap in the face" felt by the Hungarian IT profession is thus a real experience, but it would be misleading to put it down to AI alone. We are rather in the middle of a labor market regime change, in which the old social contract ("learn to code, start as a junior, the market will absorb you") is being rewritten. The new contract is harsher and less predictable, but it is not closed: it opens up to those who move to the points of the value chain where decisions, responsibility and domain knowledge matter. In the following parts, I will look at these points one by one.