Agentic AI is probably the most talked-about development in enterprise AI right now. But there's still a lot of confusion about what AI agents are, how to deploy them effectively and what they can realistically achieve. This is why IBM recently ran a webinar to help clarify some of the main issues that people are still unclear on.

Called 'Fact or Fiction? Top Misconceptions About AI Agents', the webinar debunks the biggest myths about agentic AI and describes how agents are being successfully deployed by enterprise customers.

1. Businesses already work with a suite of siloed tools, and AI agents will add to it

This misconception only becomes true if you deploy AI agents incorrectly. Agents are not meant to operate in silos but work alongside humans as enablers of existing processes. The real value comes from using agents to redefine or rethink processes or workflows working alongside humans, not just layering them on top of existing processes to create separate silos.

You can create multi-agentic systems in which agents can speed up or enhance a manual process, cut unnecessary steps or consolidate tasks and quickly pass outputs over to humans to do the higher-level work. In this way, rather than creating more silos, they cut them and optimize processes by doing lots of things that previously suffered from sticky or slow manual handoffs.

IBM recently worked with a global life sciences company to redesign its regulatory approval process for introducing new products. Instead of relying on technical writers and medical engineers to prepare the comprehensive regulatory approval documents, which often stretch into thousands of pages, it now uses an agentic AI process. A team of AI agents, including data discovery agents, technical writer agents and evaluation agents, combine to produce the draft review by human experts. This has reduced the time to first draft from 6 weeks to just 8 minutes.

2.AI agents have too much unproven hype##

There is certainly some hype and a degree of agent washing going on. Every business problem certainly does not require an agent. You need to think of the right use case or specific part of a process where agents can really make a difference.

Two areas currently stand out as places where agents are having an impact. The transformation of operational business processes such as customer service, finance and procurement, and IT and software development lifecycles, where there is evidence of agents supporting human developers and solution architects to improve productivity.

As an example, a large home appliances company in Europe has used agents to enhance a customer service process already improved through digital transformation. It saw a significant uplift in productivity from turning a conventional chatbot that had been using natural language processing (NLP) into a fully agentic system. The AI agent can respond to a broader set of topics and intents than the chatbot, and not just answer questions, but actually solve problems for customers by getting work done.

It's very important to consider the level of accuracy you're seeking from agents. Research has shown that 56% of agentic pilots in enterprises fail. But sometimes that's because the use cases they are targeting require accuracy levels beyond what's possible right now.

In initial prototypes, you might get 70% accuracy from agentic AI. With progress towards fully functioning agents, you can expect to get up to 90%. But enterprises sometimes want an accuracy rate of 96% to 100% which isn't possible yet. So, look for use cases that can accept 70% to 90% accuracy. Otherwise, AI agents will not be a satisfactory solution for the task you are trying to automate or enhance.

AI agents are difficult to govern and increase errors

Governance and accuracy are critical and very achievable with the right approach. One IBM partner in the AI-powered advertising resource management sector is using watsonx to build various AI agents, including agents for automating media planning, media activation, financial planning and more. For them, governance is essential: if an agent makes a mistake in media buying, for example, it could cost millions of dollars.

It is important to think through what the agents should and should not do within a workflow and create appropriate guardrails and common-sense rules. Ensure transparency in the decision-making process. Was it the agent or the human that took an action or made a decision? Can you trace the data that the agent relied on for a particular suite of decisions?

If an agent is about to take action, include a human in the loop. And if an agent is handling a particularly important process, think about who within the organization should have access to that agent.

To reduce error rates, consider using multiple LLMs or SLMs to cross-check each other, helping to limit the impact of bias in training data.

Agents perform badly when they're not given the right context about how the organization makes decisions - including company policies, ways of working and data. To improve the accuracy and quality of agentic outputs, everything you would give employees to do their job, you should also give to your agents.

4. AI agents are expensive and hard to set up

Like anything, if you don't have the right guidance and knowledge, this can be true.

The agentic AI ecosystem is becoming very complex, with the CIO or CTO having to make numerous choices between different partners, platforms, software vendors, hyper scalers and more. This is both a challenge and an opportunity.

Ideally, you want to build an open architecture that allows you to run agents across any platform, system, LLM or cloud. Processes often sit across multiple platforms, so being able to easily orchestrate them across those platforms is really important.

Reduce complexity and cost by working with a partner who has thought through what works and what doesn't and has established robust and flexible frameworks with data connectors to common third-party platforms. An open architecture like this also avoids vendor lock-in.

To keep costs manageable, think hard about individual use cases and the level of accuracy they require. Many tasks don't need the latest and greatest expensive frontier models. Smaller language models can achieve the required accuracy while being significantly cheaper.

5. Agents are just better chatbots

There is a clear distinction between chatbots and AI agents. Chatbots can answer questions and provide insights, while AI agents can access tools, datasets and systems, enabling them to take direct action.

For example, a major Italian telco used agents to streamline media and advertising workflows, reducing them from 90 minutes to 6 minutes. Previously, teams across media, marketing and analytics were overwhelmed by slow, manual workflows requiring multiple handoffs and specialist intervention. Chatbots were unable to execute tasks, only answer questions, forcing analysts to manually run queries, prepare data and build reports.

The company addressed the shortcomings of the chatbot experience by deploying a coordinated set of AI agents capable of understanding requests, orchestrating processes, running analyses and delivering ready-to-use outputs, not simply replying with text. They included a project manager agent, a data analyst agent and a reporting specialist agent.

To calculate the real value of agentic AI, you should also consider how it gives you the opportunity to rethink and redesign processes to take advantage of autonomous planning, reasoning, tool use, and collaboration. Ideally, you are not just going to bolt an agent onto an existing workflow without considering how you can improve the workflow and make it more efficient and effective.

Far from being just upgraded chatbots, agentic AI provides an opportunity to rethink and streamline workflows and improve productivity. With proper governance and guardrails, appropriately deployed AI agents can work alongside human employees to free them of routine, repetitive tasks so they can focus on higher-value activities.

This blog was first published in the IBM Community.