From conversation to action: AI agents go beyond chatbots, IBM explains

I’d like to start with what we’re seeing in the market from an AI perspective and then transition into why agents will be fundamental to the future. At the core, LLMs today serve as the foundation for agents. In the market, almost all organisations have GenAI projects of some sort, but only a few are scaling them. Three issues act as a barrier in the journey from proof of concept (POC) to production. Number one is cost, because using models, especially large language models (LLM), has an associated cost. And what you do in a POC doesn’t necessarily translate into the same cost structure when operating in production.

The second area is complexity, which comes from multiple areas — customising data and taking an LLM that may know public data, but not your data to drive the right outcomes, and responsible AI components of trust, transparency, and robustness. There’s complexity because the AI system is not deterministic, so understanding how that AI works is required as it goes into essential systems.

The third is expertise, which comes from building these AI systems, and application to a particular problem. Domain expertise becomes a big part of it because AI will not do the transformation by itself, but act as an ingredient.

Everybody’s building agents and with different frameworks. You should be able to incorporate and build these agents to do autonomous work.

We are using our agent technologies at the intersection of AI agents and traditional automation, combining agentic AI and automation to drive process efficiencies and boost productivity in organisations. And so, we’ve introduced some areas where we’re driving a lot of efficiency.  

Some are automation and collaboration, including multi-agent collaboration. Agents are being used to enable conversational AI that goes beyond assistants that can do autonomous work on the backend.  

For agents to work, the foundation is a good model that knows how to reason and that you can iterate on with your data to have the right outcomes.  

We fundamentally believe much value will be captured in driving domain-specific agents, those pre-built for specific tasks. Within our portfolio, we have an agent in the HR space, which can be used to accelerate employee onboarding, candidate screening, and employee engagement.

We are also driving these domain-specific work in procurement and sales. Another thing is multi-agent orchestration to have one interaction paradigm across a number of agents.  

I could use an agent to communicate my vacation days or tell IBM I’m taking some time off. At the same time, I could use a sales agent. Through multi-agent orchestration, I can have one interaction paradigm, interact with several agents underneath, and have a single experience where the agent determines which sub-agent is essential to call.

GenAI uses an LLM trained on public data that may be customised. The foundation of what you’re using in an LLM or an SLM is the building block for agents.  

You can use an agent to plan a series of steps. For example, to build a piece of code, I may plan the steps to design it. Then, I may reflect. The reflection may use that same model to critique my code. With the LLM, you’re doing a Zero-shot, which is just writing the code. Now with the agent, you’re taking an iterative approach to planning how to write the code, execute, and potentially use that same model to critique that code, debug it, and rewrite it. This iterative nature gets you better accuracy than in Zero-shot learning.

The difference is you’re autonomously acting with agents. I’m not telling the agent every time to execute something. The agent is doing that iteration. So, the outcomes can be more accurate.   

By definition, AI agents can do autonomous actions. However, in many use cases where you may have had assistance before, you still can have an engaging experience.  

In a conversational agent experience, you are still augmenting the human, but using agentic technology. For employee engagement, you can still have, instead of an assistant, agentic technologies underneath that enable a conversational agent-style experience. You can also have these agents operate autonomously.

When we look at how Gen AI or agentic technologies are being used, it does mean teams need to understand how to utilise AI. It is about enabling and skilling our engineering teams to use the technology right. Internally in IBM, all our engineers, and even employees in different functions, use and are skilled in GenAI.

For example, sales agents. A seller’s role involves understanding and researching their organisation’s sales plays, then using research tools—potentially including web-based tools—to collect insights on target accounts. From there, they identify key contacts within those accounts and use that information for prospecting. Finally, they apply these insights to manage opportunities and advance deals through the sales pipeline.

When you’re using agentic technologies, it’s also a case of a conversational interface where we’re assisting the user, and it’s a much more engaging and productive experience for the seller where they’re probably going from opportunity identification to management in much less time with a lot more depth than earlier.

Every sector. There is not one industry agents will not apply to. Agentic technologies are being used in customer care and customer experience that are forward-facing, but also in back-office-facing ones. This includes sales, procurement, and HR. You can think about supply chain, manufacturing, and finance.  

I think it’s early, and that there’s a spectrum in agents. Of the four different use cases — executing tasks, automating things, collaboration, and orchestration, there are some areas where agentic technologies will be more easily used.  

If I’m using agentic technologies for tool calling, as an example, to do autonomous work, this will probably be adopted more quickly by enterprises than more complex use cases. It depends on the use cases. We’re already seeing some where agentic technologies can be applied easily. There’s a lot of interest in agentic technologies because of the promise of this intersection, of using agentic technologies to drive automation for productivity, efficiency, and innovation.