Agents vs. the AI Bubble

If 2023 and 2024 taught us anything, it was that Generative AI was on many people's minds. Capital was deployed as OpeEx by hiring data scientists or via CapEx by funding pilot projects to use the technology. In 2025, it is time for AI investments to translate to AI outcomes.

Many companies have invested significant resources to leverage AI but have little to show.

There are many reasons for the lack of outcomes from AI efforts, including but not limited to:

Unclear implementation goals

Cost of data cleansing and organization

Limited support or training

Cultural resistance

Leveraging intelligent automation to drive outcomes today differs from early, corpus-intensive machine learning. Services are more mature and have benefitted from iteration and refinement. However, implementing a large language model (LLM) that maintains data security and provides reliable, accurate information isn't easy.

On-Demand Delivery

Many organizations wanted to skip the 'crawl-walk-run' approach to utilizing technology and run a marathon—without training or shoes—in the desert. You get the idea.

Consultants at large firms have packaged and sold generative AI to offset the reduction in legacy service offerings. There is nothing new about growth in new areas balancing contraction in others, but Generative AI is an emerging technology still being understood. The areas of the business that shrunk? Those lines of business were commoditized and already in the process of being disrupted.

Legacy sales and delivery models were leveraged to promote new technology. Generative AI became an excuse for companies to take action and invest in areas they had been delaying. Implementing Generative AI became a new goal to achieve, regardless of business needs.

Perception vs. Reality

Generative AI became a convenient scapegoat—data migrations, governance models, and all sorts of ETL activities were added on. Generative AI became too big to fail, too expensive not to do since there were so many associated tasks and adjacent benefits. Many adjacent services were sold to support its success.

With so much investment came promises that could not be delivered. Anything is possible with time and unlimited resources, but 2024 was a 'wait and see' period. Patience has worn thin in 2025 and now we have a bubble about to burst.

Sunk Cost Fallacies aside, there is hope on the horizon. It has gotten a lot of talk in AI circles for the last 18 months (which means it has benefitted from R&D for far longer). AI Agents are here.

Say hello to your little friends

Agents are task-specific and far more straightforward than LLMs. They represent a direct relationship between business needs and the delivery of value.

AI agents and Large Language Models (LLMs) have overlapping capabilities but vary in how they are built and used. LLMs have been known to make mistakes, which impacts how we measure their effectiveness. An AI agent does more or less what you tell it to do.

An agent is a task-focused system designed to complete specific actions without oversight. It makes decisions on behalf of humans, informed by decision trees, algorithms, or neural networks. It can be programmed with failure protocols and fallback mechanisms to manage errors. Like humans, it can also make mistakes when provided with incomplete or conflicting data.

An AI Agent is highly accurate within a well-defined scope.

An LLM has a different job: to be generally effective and applicable. It focuses on language prediction, not task execution. If source data are ambiguous or training data is biased, this is reflected in the results. Due to their open-ended design, LLMs suffer from contextual misunderstandings.

An LLM relies on user input to provide feedback in error states. If a business wants specific, repeatable outcomes, an LLM isn't the best option compared to an agent. Most enterprise software is sold based on its ability to provide specific, repeatable outcomes in narrow contexts defined by users.

So, what technology should an organization use?

Implementation of Generative AI involves both LLMs and Agents

There are times when business value is achieved in a broader context. Combining LLMs with AI agents often leads to better outcomes than implementing one technology without the benefit of the other. Agents can leverage LLMs' language understanding capabilities while ensuring structured task execution. LLMs can be prompt-chained for specific uses. We are still working to understand the benefits of having both.?

The hybrid state of LLMs, agents, and old-fashioned automation will drive the outcomes businesses were looking for in 2023. Even if a company operates on a common platform, it's always a series of services, applications, and an ensemble of APIs that provide users with the agency they need to make an impact.

Let us know if this sounds complex to orchestrate and you need help—Gyroscope partners with organizations to optimize AI's ROI.