Are AI agents just a big bubble?

Why are people calling AI agents a bubble now?

Well, AI agents look bubbly today because the market is selling one big dream while the evidence shows several very different realities.

First, we can understand why some people are suspicious.

Gartner said in June 2025 that more than 40% of agentic AI projects could be canceled by the end of 2027, mostly because costs rise, business value stays fuzzy, or risk controls are too weak. That is a serious warning because it does not come from AI doomers; it comes from enterprise technology buyers seeing what happens after the demo.

The second warning is vocabulary inflation. Gartner also pointed out that many vendors are basically “agent-washing” old chatbots, RPA tools, and assistants. That matters because if every workflow automation tool suddenly becomes an “AI agent,” the market size starts lying to us.

The third warning is money. PitchBook reported that AI venture funding in Q1 2026 was already larger than all AI funding in 2025. Morgan Stanley estimates almost $2.9 trillion in global data-center construction costs through 2028. That is not normal software-market optimism. That is industrial-scale capital being placed before the end demand is fully proven.

So yes, people are right to smell bubble dynamics.

If you want more recent data on this point, please see our latest agentic AI market report.

Are companies actually using AI agents, or just talking about them?

Companies are using AI agents now, but mostly in a shallow way.

McKinsey’s latest State of AI survey is useful because it separates broad AI use from agent scaling. It found that 88% of organizations use AI in at least one business function, but only 23% are scaling agentic AI somewhere in the company. More telling: in any single business function, no more than 10% of companies say they are scaling agents.

Capgemini tells a similar story from another angle. In its survey of 1,500 senior executives, only 2% said they had deployed AI agents at full scale, 12% at partial scale, 23% were piloting, and 61% were still exploring or preparing. That is not a dead market but one where the funnel is wide at the top and very thin at the bottom.

LangChain’s 2026 agent-engineering survey looks more optimistic, with 57% of respondents saying they have agents in production. But that survey is naturally closer to builders and technical teams. Read together, the picture is clearer: builders are pushing agents into production, while the wider enterprise is still figuring out how much of that production is actually business-critical.

So the answer is yes, AI agents are being used. However, today, most agents are still sitting inside bounded workflows, not replacing full departments.

Are AI agents making real money already?

Yes actually, but the money is concentrated in a few sharp use cases.

The cleanest revenue signals are coming from coding, customer service, and professional workflows. Cursor said in June 2025 that it had passed $500 million in ARR and was used by more than half the Fortune 500. Salesforce reported in its Q4 FY2026 results that Agentforce ARR reached $800 million, up 169% year over year, with more than 29,000 Agentforce deals since launch. Sierra said it entered its third year with more than $150 million in ARR in enterprise customer service.

Those are not tiny numbers. But they also tell us something important: the strongest agent businesses today are not selling vague autonomy. They are selling very specific work compression. Cursor compresses software development. Sierra compresses customer support. Salesforce is trying to plug agents into existing enterprise workflows where the buyer, data, and budget already exist.

Are AI agent startups overvalued?

Many AI agent startups are currently overvalued, but the valuation problem depends on the type of startup.

A coding agent with huge usage, recurring revenue, and daily developer dependency is not the same thing as a generic “AI employee” wrapper. Cursor’s reported ARR, Salesforce’s Agentforce revenue, and Sierra’s customer-service traction show that some agent companies are selling into urgent budgets. Those businesses deserve to be taken seriously.

The froth shows up when valuations price every agent company like it will own a core workflow. That is a much harder claim. Business Insider recently described how OpenAI, Anthropic, Cursor, Canva, and others are all moving into each other’s territory. That overlap is dangerous for startups because the surface product is easy to copy, especially when the model provider, workflow platform, and incumbent SaaS vendor all want the same customer.

There is another risk: platform dependency. A startup building on frontier models can grow very fast, but the model company can also become a competitor overnight. It means revenue quality matters more than the headline ARR. We need to ask: is the product deeply embedded in the workflow, or is it a nice interface on top of someone else’s model?

So yes, parts of the startup market look overvalued. The strong companies have workflow lock-in. The fragile ones have a demo, a big story, and very little protection.

If you want more recent data on this point, please see our latest agentic AI market report.

Will most AI agent projects fail inside companies?

A lot of them will fail, and that is probably the base case.

Gartner’s cancellation forecast is one hard signal. Capgemini’s deployment funnel is another: if only 2% of organizations are at full scale while 61% are still exploring, then many projects are still far from proving value. Forrester’s 2026 work says the same thing in plain enterprise language: companies are excited, but many remain stuck in pilots because ROI is unclear, governance is weak, and teams do not know whether to buy a SaaS agent, build custom, or rely on a services partner.

The IBM signal from June 2026 is even more concrete. In a survey of 2,000 CIOs and CTOs, only 11% said they were completely prepared for large-scale AI agent deployment. Two-thirds said they were accountable for AI systems they did not fully control. That is exactly how failed enterprise projects happen: the business wants autonomy, but IT still carries the risk.

So the failure rate will probably be high. But that should not surprise us. Most early CRM, cloud, RPA, and analytics projects also failed when companies treated the tool as the strategy. AI agents are going through the same cleanup, only faster.

Is the technology reliable enough for real work?

For some work, yes. For unsupervised work, not yet.

The capability curve is genuinely improving. METR found that the length of tasks frontier AI agents can complete with 50% reliability has been doubling roughly every seven months. That is a strong signal because it measures something closer to real work than a normal benchmark: how long a human would have needed to do the task.

But deployment is still much messier than capability. A May 2026 industry study across twelve companies found that most were still using agents as assistants or compensators. Only one had reached multi-agent orchestration. The reason was simple: teams could show more advanced agent behavior in experiments, but they still needed humans to check the output before putting it into daily operations.

A 2026 MIT-linked AI Agent Index also found that many agent developers disclose little about safety, evaluations, and societal impact. That is a problem because enterprise buyers do not just need an agent to work. They need to know when it fails, why it fails, what it touched, and how to reverse it.

So the technology is good enough for supervised acceleration. It is not yet good enough for the fully autonomous story being sold in many pitch decks.

Are AI agents replacing workers already?

Not broadly. They are replacing pieces of work before they replace jobs.

The data does not support a mass-replacement story today. McKinsey found that only 39% of companies attribute any EBIT impact to AI, and most of those say the impact is still below 5% of EBIT. That is too small to support the idea that agents are already reshaping the labor market at scale.

The stronger signal is inside specific tasks. Developers report productivity gains from AI tools. Customer-service agents are being supported or partially automated. Internal ops teams are using agents for routing, summarization, research, QA, and follow-ups. But these are fragments of jobs, not whole job categories disappearing overnight.

This is where the conversation often gets lazy. “Agents will replace workers” sounds dramatic, but the measurable story today is work unbundling. A job is made of tasks. Agents first attack the digital, repetitive, easy-to-check tasks. The harder human layer remains: judgment, escalation, relationship management, accountability, and deciding what good actually means.

So AI agents are not replacing workers at scale right now. Instead, they are changing what workers spend time on, and that can still be a very big deal.

Is enterprise demand real, or are vendors forcing it?

Enterprise demand is real, but buyers are more cautious than the headlines suggest.

LangChain’s 2026 survey says 57% of respondents have agents in production, and quality is now the top production barrier. That detail matters. If cost were the main objection, we would call this a budget issue. If safety were the only objection, we would call it a risk issue. But when quality is the top barrier, it means buyers want the product and are testing it; they just do not fully trust the output.

IBM’s June 2026 survey adds another layer. AI agent use is expected to grow 38% by 2027, but 77% of CIOs and CTOs say current governance frameworks are inadequate. In other words, demand is running ahead of control.

Salesforce gives us a buyer-side commercial signal too. Agentforce closed more than 29,000 deals since launch, and production accounts grew nearly 50% quarter over quarter. But Salesforce’s own stock pressure and slower wider growth show that agent demand does not automatically fix every SaaS growth problem.

So demand is clearly more conditional than vendors want to admit. Buyers want agents when the workflow is clear, the output can be checked, and the business case is measurable.

If you want more recent data on this point, please see our latest agentic AI market report.

Is the AI agent infrastructure buildout rational?

Partly, but this is where the bubble risk gets largest.

The infrastructure numbers are huge. IDC said AI infrastructure spending hit $89.9 billion in Q4 2025 and is heading toward $1 trillion by 2029. Morgan Stanley estimates nearly $2.9 trillion in global data-center construction costs through 2028. Menlo Ventures noted that foundation-model companies had announced close to $1 trillion in AI infrastructure commitments.

Those numbers can be rational if agent usage explodes. Agents are compute-hungry because they do not just answer once. They plan, call tools, retrieve documents, check steps, retry, and sometimes run multiple model calls for one business outcome. One useful agent workflow can consume far more inference than a simple chatbot query.

But here is the uncomfortable part: enterprise scaling is still narrow, as we saw earlier. If agent deployment stays mostly in pilots, coding tools, and support automation, some infrastructure assumptions will look too aggressive. The risk is that usage grows, but not fast enough, not profitably enough, or not autonomously enough to justify the capex curve.

So the infrastructure bet is the most bubble-like part of the agent story. The software shift can be real while the physical buildout still overshoots.

What is most overhyped in AI agents?

The most overhyped thing is the “autonomous employee” narrative.

Today’s best evidence points to supervised systems, not independent digital workers. Capgemini found that trust in fully autonomous agents fell from 43% to 27% in one year. IBM found that only 11% of tech leaders feel fully prepared for large-scale deployment. The May 2026 industry study found most companies still below true orchestration.

That is a very different story from the pitch-deck version. The pitch-deck version says agents will take a goal, figure everything out, and execute across the company. The deployed version usually says: give the agent a narrow job, connect only the tools it needs, keep logs, cap its permissions, measure the output, and put a human nearby.

The winners will probably look boring from the outside: agents for claims, invoices, code review, IT tickets, customer replies, compliance checks, sales research, and internal knowledge work.

Autonomy is overhyped. Bounded execution is the real product.

If you want more recent data on this point, please see our latest agentic AI market report.

What is most underhyped in AI agents?

The most underhyped part of AI agents is the control layer around them.

Most people are still watching model intelligence. We think the more important market is everything that makes agents safe enough to use: evaluations, observability, permissions, identity, audit trails, cost controls, fallback rules, and human review. LangChain says observability is now table stakes. IBM says organizations that build control into AI systems see stronger performance outcomes. Salesforce even reported production work on scalable inference for compound AI systems, with lower latency and cost savings versus earlier static deployments.

That is where the non-obvious value sits. If agents become common, companies will not just buy “agents.” They will buy the operating system for managing agents. Who can this agent talk to? What can it change? What did it do? Why did it make that call? How much did it cost? When should it stop?

This is why the best agent businesses may not all look like cute digital coworkers. Some will rather look like infrastructure, governance, and workflow-control companies. Less sexy, probably more durable.

So, are AI agents just a big bubble?

AI agents are a real software shift wrapped in a bubbly market structure.

The real part is easy to see now. Companies are using agents. Some agent-native products are making serious revenue. Technical capability is improving quickly. Buyers are asking for agents in real workflows, not just playing with demos.

The bubbly part is also clear. Most companies are not scaling deeply yet. Many projects will fail. The autonomous-worker story is ahead of the deployment reality. Startup valuations assume workflow ownership before it is proven. Infrastructure spending assumes a very large future demand curve that current enterprise adoption does not yet fully support.

So the clean answer is this: AI agents are overhyped as independent workers, underhyped as workflow compression, and very real as a new software-control layer.

If we had to make the call today, the highest-probability outcome is not a total agent crash but rather a label correction. “AI agent” stops being magic language. Weak wrappers disappear. Generic demos get repriced. Enterprises cut projects that cannot show ROI. But the useful agentic workflows stay, especially where the task is frequent, digital, measurable, and easy to review.

The thing most likely to “burst” is not the idea of agents. It is the belief that autonomy alone creates value.

If you want more recent data on this point, please see our latest agentic AI market report.

Check	Answer	Comment
Are companies actually using agents?	Yes	Usage is real, but deep scaling is still limited.
Are agents making money?	Yes	The best revenue signals come from coding, customer service, and enterprise workflow tools.
Will many projects fail?	Yes	Poor governance, unclear ROI, and weak integration will kill many pilots.
Is the technology reliable enough?	Partly	It works best when tasks are bounded and humans can review outputs.
Are agents replacing workers?	Not broadly	They are taking task fragments before full roles.
Is enterprise demand real?	Yes	Buyers want agents, but only when quality and control are credible.
Are startups overvalued?	Many are	Workflow-native companies look stronger than generic wrappers.
Is infrastructure spending rational?	Mixed	Compute demand is real, but the capex curve needs much broader agent usage.
What is overhyped?	Full autonomy	The “digital employee” story is ahead of real deployment.
What is underhyped?	Control systems	Evals, permissions, observability, and workflow redesign may capture durable value.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.