Every week there is a new announcement about an “AI agent” that can do your job for you. But a new survey of 101 enterprise organisations from VentureBeat reveals a surprising truth: most of what companies call AI agents are actually just simple chatbots.
The survey asked enterprises to assess their own portfolios honestly. The results were striking. 71% of companies said that a quarter or fewer of their deployed “agents” were true multi-step, orchestrated workflows. Only 10% of companies had crossed the halfway mark. The rest were single-prompt chatbot wrappers dressed up as something more advanced.
Why the gap matters
This is not just a semantics problem. When businesses buy into the idea that AI agents are ready for prime time, they make decisions based on unrealistic expectations. They invest in platforms, retrain staff, and restructure workflows around technology that cannot yet deliver what was promised.
The survey found that enterprises judge AI agents primarily on task completion reliability (32%) and multi-step workflow management (28%). Yet most of their own deployments do not even do multi-step work. The ambition is real, but the execution is lagging.
Interestingly, the gap was narrower in larger enterprises. 62% of companies with over 2,500 employees said a quarter or fewer of their agents were genuinely orchestrated, compared to 77% of smaller ones. This suggests that the “chatbot trap” is especially pronounced for smaller businesses that may lack the resources to build genuine multi-step agents.
The cost control problem
Another worrying finding: more than a quarter of enterprises (27%) have no real-time, programmatic way to stop an AI agent before it runs up an unexpectedly large bill. They learn about runaway costs from the logs afterward. This matters because AI agents consume far more tokens than simple chatbots. An agent that loops through the same task repeatedly — or gets stuck in an infinite reasoning cycle — can burn through a monthly budget in hours.
The businesses building their own cost-control systems (23%) or using cross-model routing to arbitrage costs (19%) are the minority. Most enterprises rely on native caps and throttles built into their AI platform — a control only as good as the provider’s tooling.
On the platform side, the survey found that Anthropic’s Claude leads the enterprise agent orchestration market at 40%, more than double any rival. Microsoft holds 18% and OpenAI 13%. The main driver is “model gravity” — enterprises pick the platform that has the best underlying model. But 51% expect a hybrid control model by year end, keeping some control outside the provider to avoid lock-in.
What this means for Irish businesses
If you are considering deploying an AI agent in your business, the survey offers practical lessons. First, be specific about what you need. A simple chatbot that answers customer questions might be perfectly adequate — you do not need a multi-step agent for every use case. Most businesses would be better served by a well-designed chatbot than a half-built agent.
Second, set cost limits before you start. Most AI platforms let you set monthly spending caps, but few businesses use them. Decide in advance what a reasonable monthly AI spend looks like for each tool, and set hard limits.
Third, test before you trust. Run a pilot with real customers or real data before committing to a platform. Make sure the agent actually performs the multi-step tasks you need, not just the demo version.
The bottom line
AI agents are coming, and they will eventually deliver on the promise. But right now, most “agents” are still chatbots in expensive clothing. Treat vendor claims with healthy scepticism, test thoroughly, and keep your expectations grounded in what the technology can actually do today.