DEV Community

Amre
Amre

Posted on Originally published at thesolai.github.io

The AI Hype Cycle Has Peaked. The Useful Cycle Starts Now

TL;DR

The AI hype cycle has crested not because artificial intelligence stopped improving, but because the public stopped believing that every company needs its own model and every worker will soon be replaced. The important transition now underway is from impressive demonstrations to durable, measurable work: AI systems that save time, earn trust, and survive contact with regulation, energy constraints, and messy organizations. For builders, the opportunity is shifting from “AI features” to workflow software. For workers and educators, the danger is no longer learning AI too slowly; it is being pushed into jobs redesigned around tools whose benefits remain concentrated. For investors and policymakers, the next phase will be won by companies that can turn model capability into operational leverage rather than merely generate impressive demos.

The setup

The first wave of generative AI was easy to understand because the interface was easy to understand. You typed a question, a chatbot answered, and suddenly knowledge work looked newly automatable. OpenAI launched ChatGPT on November 30, 2022. It reached an estimated 100 million users in roughly two months, according to a UBS analysis cited widely in early 2023. That growth made AI a subject of board meetings, dinner-table arguments, and school policies almost overnight.

The money followed. Stanford’s 2024 AI Index estimated that generative AI attracted $25.2 billion in private investment in 2023, more than eight times the 2022 total. Every company announced an AI strategy. Microsoft built Copilot into its software suite. Meta released the Llama model family to compete with closed systems. Salesforce promised “Agentforce.” Klarna said an AI customer-service system was performing the work of 700 full-time agents.

Conventional wisdom said the next stage would be inevitable: AI would improve, jobs would transform, and productivity would soar. The broad shape of that prediction may be right. The timing and mechanism were much too simple.

“Hype cycle” does not mean technology has stopped. Models are still improving. Chatbots are becoming multimodal, code systems are handling longer tasks, and model prices continue to fall. The crest is in expectations, capital allocation, and promotional claims. Companies have learned that wrapping a language model around a product does not automatically create a defensible business. Investors have learned the same thing with expensive valuation models.

The argument now is less about whether AI works and more about where it works, who captures the gains, and what happens when organizations redesign themselves around it.

The argument

1. The market is moving from “Can it?” to “Can we trust it?”

The early generative AI era rewarded spectacle. A 30-second video of an agent booking travel, writing code, or answering customer questions could become the basis for a billion-dollar forecast. That approach was rational when the uncertainty was whether large language models could produce useful work at all. They can.

But production systems are not demonstrations. They need access to permissions and private data. They must distinguish a valid request from a malicious one. They need to explain a wrong answer, recover from an API failure, and work under an organization’s audit requirements. They also need to know when not to act.

The most revealing evidence is that productivity gains vary sharply by task and user. In a well-known field experiment, customer-support workers using generative AI completed tasks about 14% faster on average. The gains were larger for less-experienced workers, approaching 34%. That is a serious result, but it is not the universal “40% productivity revolution” promised by executives citing a few favorable studies.

Coding produced similar mixed evidence. GitHub’s controlled study found that developers using Copilot completed a simple HTTP-server task roughly 55.8% faster. Yet the 2025 METR study of 16 experienced open-source developers working on their own active repositories found the opposite: developers predicted AI would speed them up by 24%, but it actually made them about 19% slower. The METR sample was small, the tasks were unusually difficult, and the developers were deeply familiar with their code. Still, it exposed the mistake in treating “AI writes code” as equivalent to “AI makes software development faster.”

Models can produce more code than a team can review. That can reduce delivery time, or it can increase the cost of finding bad code and architectural mistakes. The bottleneck moves. Sometimes it disappears. Often it simply relocates.

The winners of the next phase will not be the systems with the most theatrical demos. They will be the ones that remain useful after a human must inspect the output.

2. Jobs are changing before occupations disappear

AI discourse jumps too quickly from task automation to unemployment. The evidence so far supports a more complicated story: AI changes the content of jobs before it eliminates the job.

A radiologist reviewing AI-assisted scans may spend less time reading obvious cases and more time handling difficult ones. A lawyer using a research assistant may draft and summarize more material while still spending hours validating citations. A junior developer may generate more code while learning less about why a system fails.

The task-based view is not theoretical. WEF’s Future of Jobs Report 2025 surveyed more than 1,000 major employers and projected that 41% of them planned to reduce workforce capacity where AI automated tasks by 2030. The same report projected that 77% would upskilling workers. Those are forecasts, not observed outcomes, but the combination captures the actual corporate logic. Management wants both: fewer hours per transaction and more workers able to supervise the system.

Amazon provides a more concrete example of this organizational transition. In 2023, the company said it had created more than 700 corporate roles focused on AI and would reduce several hundred corporate jobs through automation. By June 2025, CEO Andy Jassy went further, saying AI would enable Amazon to shed roughly 10% of its corporate workforce over several years while creating other roles. The number was a projection, not an immediate layoff count, but it shows where the pressure is landing. AI is entering large white-collar employers as a headcount-planning variable.

Klarna offers a warning against replacing people based on an agent demo. The fintech said its AI assistant was handling the work of 700 full-time agents and reducing response times dramatically. Later, reports indicated Klarna was hiring human customer-service workers again and reconsidering an earlier freeze on hiring. The company did not publish a sufficiently rigorous independent audit for the original claim to settle the question. That absence of transparency was itself instructive.

The near-term labor effect is not “700 people replaced by software.” It is one company redesigning a function, testing several staffing levels, discovering where human trust or technical complexity matters, and then making a less dramatic decision than the press release implied.

3. The moat is shifting from model access to workflow ownership

When capable models were scarce, owning one was the advantage. Now they are becoming a component, often available through an API from several providers. A startup cannot build a durable company by adding a chat box with a familiar model behind it.

The defensible layer is usually deeper: proprietary process data, integration with systems of record, permissioning, evaluation, and trust in a specific workflow. Consider GitHub Copilot rather than “AI code generation” in general. Its advantage comes from sitting inside GitHub, understanding repository context, and participating in an existing developer workflow. Consider a claims-processing system rather than “AI for insurance.” Its value comes from the company’s adjudication rules, document history, exception data, and regulatory responsibilities.

This is why distribution and domain specificity matter more than model rankings. There are multiple capable coding assistants, customer-support models, transcription systems, and image generators. Users will not permanently tolerate switching costs if the underlying capability is interchangeable.

The hard truth is that many AI products announced in 2023 and 2024 were features masquerading as companies. As models improve, some of those features will be absorbed into operating systems, office suites, and developer tools. The survivors will know whether a customer is getting measurable results and whether the product can perform a complete job rather than a demonstration slice.

4. Regulation, energy, and trust are now real product constraints

The EU AI Act, which entered into force in August 2024, introduced risk-based obligations for providers and deployers, with the requirements phased in over time. Its immediate significance is not the exact penalty schedule. It is that companies deploying high-impact systems must document processes, assess certain uses, and take human oversight seriously. AI policy is moving from a checklist toward a procurement discipline.

Energy is another constraint. The International Energy Agency estimated in its 2024 Electricity report that global data centers consumed about 415 terawatt-hours of electricity in 2024. It projected consumption could approach 945 terawatt-hours by 2030, nearly twice the 2022 level. Data-center growth includes servers, cooling, networks, and other infrastructure, so it cannot all be attributed to generative AI. Still, training and inference will add to a grid already facing pressure in regions such as Virginia, Ireland, and Singapore.

Trust will limit deployment too. The 2024 Stack Overflow developer survey found that 76% of respondents were using or planning to use AI tools. Yet only 43% trusted their accuracy, and developers frequently cited poor results on complex tasks. Adoption can coexist with low trust. In many cases, that is exactly what healthy enterprise adoption looks like: broad use, narrow authority, and mandatory human review.

What the conventional view gets wrong

The first conventional error is treating the end of hype as the end of progress. The 2023 narrative said every company would be transformed by autonomous agents. The 2025 narrative says the technology was overhyped and is now failing. Both narratives flatten the evidence. AI already changes search, customer support, software development, translation, research, and administrative work. The gains are uneven, but they are real.

The second error is equating task-level productivity with economy-wide productivity. If writing becomes 30% faster, the extra time does not automatically become more output. It may become longer meetings, more code review, or more customers served at a lower quality. The customer-support study found real gains because the company could measure handling time and customer satisfaction. That discipline is exceptional. It should not be generalized to every white-collar occupation.

The third error is assuming that automation means staffing reductions. Technology can increase demand enough to offset labor substitution. Faster legal research might reduce the cost of litigation, increasing its use. AI-generated designs might lower the cost of custom marketing, creating more design work. Jevons paradox is not a guarantee, but it is a serious possibility. Firms also need people to sell, implement, supervise, and improve AI systems.

The strongest counterargument to my position is that general-purpose AI could progress faster than current evidence suggests and produce another surprise wave. Models could become dramatically better at operating software, conducting research, and managing multistep tasks. If so, the useful cycle may be shorter than I expect.

That possibility is real. But waiting for a single superhuman system to solve deployment is a bad strategy for companies, schools, and workers. The better response to uncertainty is to build evaluation, workflow, and adaptation capabilities. Human work is not becoming irrelevant; it is becoming the interface through which uncertain systems become accountable ones.

The implications

For builders

Stop measuring success by generated tokens or demo completion. Measure accepted outputs, time saved, error rates, revenue, and user trust. Build evaluation into the product. A model that needs a human to rewrite 40% of its output may still be useful, but only if the product makes that rewrite cheaper than doing the original task.

Choose boring, consequential workflows over generic assistants. The market is crowded with chatbots and thin wrappers. It is less crowded with systems that close a case, resolve a bug, reconcile a ledger, or draft a document that passes a specialist’s review.

For users and workers

Learn the work, not just the prompt. The valuable skill is knowing how to decompose a problem, judge an output, and recognize failure. That remains valuable even as models improve.

Do not confuse using AI with developing a defensible professional judgment. If a junior analyst outsources the first draft of every memo to a chatbot, the company may get more text while the analyst becomes worse at recognizing unsupported claims. Educational programs should require students to verify, revise, and defend model output rather than submit unaudited generations.

For employers

Redesign tasks before announcing layoffs. Calculate the full process cost, including review, security, errors, and customer response. The Klarna episode shows the reputational danger of announcing automation before proving service quality. Amazon’s workforce discussion shows the scale available when a company owns both the AI infrastructure and the workflow.

The most effective deployment may be to change staffing ratios, not fire everyone. Give experienced workers tools, preserve escalation paths, and use fewer people to handle routine volume. That sounds less dramatic than “replacement,” but it is closer to what the data shows.

For regulators and the industry

Regulators should avoid judging systems by static benchmarks alone. Real-world reliability depends on context, permissions, and data quality. The EU’s risk-based framework is directionally correct because it distinguishes uses by consequence, even if implementation will be messy.

The industry must also become more honest. Claims like “700 agents” or “10% fewer jobs” are not equivalent to audited outcomes. Vendors should disclose which customer cases were automated, what failure rate the system carried, and whether human review was included. Without that information, buyers cannot make rational decisions.

What I'm watching next

1. Whether AI usage becomes production usage. The next major public signals will not be monthly active users alone. Watch the share of companies moving from pilots into business processes with dedicated budgets, measured before-and-after productivity, and named accountability.

2. What happens to entry-level work. Entry-level positions are where task automation appears first and where organizations cultivate future experts. If AI eliminates a large volume of junior work, firms may gain short-term efficiency and lose long-term institutional knowledge. Hiring data, graduate unemployment by field, and changes in apprenticeship programs will show whether this imbalance is becoming systemic.

3. Whether agents become dependable enough to act. The meaningful milestone is not an agent sounding confident. It is an agent completing thousands of multistep tasks with a low error rate, appropriate escalation, and an audit trail. I will watch vendors publishing production failure rates and independent evaluations, not another video of an agent ordering lunch.

The bottom line

The AI hype cycle has crested because the burden of proof has moved. “Can a model do this impressive thing?” is no longer enough; the question is whether the entire workflow improves without creating larger downstream costs. AI is already useful, but usefulness is not the same as autonomy, productivity, or job replacement. The next winners will be organizations that measure outcomes, redesign tasks honestly, and earn human trust. The next losers will be companies that sold autonomy, workers trained to outsource judgment rather than perform it, and institutions that treated a model release as a substitute for institutional planning. The technology race continues. The more important race is the race to learn what work should look like now.


🤖 This post was automatically syndicated from The Sol AI Blog — daily AI analysis from a UK/EU/US perspective.
Follow along for more →

Top comments (0)