The release of ChatGPT on November 20, 2022 launched the generative AI revolution and with it wave upon wave of more than the usual hype accompanying a game changing new technology. The hype made it seem like large language models (LLMs) were capable of anything.
Three years later, however, the bloom is off the rose, and the waves of hype are crashing on the beaches of reality. The industry is currently engaged in figuring out what LLMs are good for, and understanding what they are not capable of.
The Agentic AI Hype Cycle
AI agents leverage LLMs to execute tasks such as dynamic pricing or employee onboarding, and can reason and act independently, preserving the necessary context and memory.
As with the overall generative AI hype cycle, when the agentic AI hype cycle began around mid 2023, Bill Gates, Mark Benioff, and others predicted that AI agents would be capable of anything and would become the ubiquitous new method for human-computer interaction, dramatically changing the future of work.
“In the next few years, they [agents] will utterly change how we live our lives, online and off,” Gates said. In his vision of agentic AI, people would not use individual applications any longer, but would instead instruct agents to do it all automatically for them.
But the prediction assumed that human language could successfully be translated into computer language, which is not the case in reality.
Unlike computer language, processing human language is probabilistic and imprecise. It’s well understood that LLMs often return inaccurate, and sometimes completely fabricated responses, also known ashallucinations..
Salesforce recently revisited their original predictions about AI agents, based on overconfidence about the technology.
A Voice of Reason
Boomi CEO Steve Lucas, however, clearly understood the challenge of using generative AI from the beginning, and describes it in his best selling book, Digital Impact. He discusses the challenges created by this lack of understanding in the context of the recent MIT study documenting the high failure rate of generative AI projects.
“Agentic transformation is about rethinking workflows, not just digitizing them. It’s about blending deterministic processes (like payroll) with AI agents that can reason and handle probabilistic tasks like expense approvals to increase agility and efficiency,” Lucas said.
In other words, those who claimed that autonomous agents would be soon taking over the world were, to put it in AI terms, “hallucinating.” Instead of entirely replacing existing deterministic tasks, AI agents are best used to complement them with the kind of probabilistic tasks they are good at.
How to Think about Deploying AI Agents
Because AI agents are non-deterministic, and pre-generative AI IT systems are deterministic, it’s important to find the right place in a deterministic workflow to introduce an AI agent, assign a non-deterministic task to it, and deploy accordingly.
In this sense, deploying an AI agent is a matter of finding the right role and task for it to play within a business workflow.
Many organizations are using AI agents for generating application code, and this appears to be the most widely adopted form of agentic AI and proving to be a significant productivity gain.
But this activity supports application development and is not a task directly related to operational business workflows.
You could have a series of agents, on the other hand, execute the steps of a travel reservation, as long as you are flexible about flight and hotel choices it makes.
An AI agent could also help evaluate substitutes for a product purchase when multiple options would be equally acceptable to the customer.
And you could have an AI agent do the analysis on setting up a convenient meeting time for a team or set of external clients.
In other words, AI agents are good for tasks that require a lot of research and analysis and do not necessarily have a single, correct answer.
Once you find the right role for an AI agent in a workflow, the next step is to figure out how to categorize and govern agents consistently across the organization. You get the most value for your investment in AI agents when identifying the right use case and deploying with enterprise guardrails to prevent costly security breaches.
Setting Up the AI Agent Ecosystem
Once you identify the tasks in your organization that AI agents are good for and set organizational policies, you can create the ecosystem for your agentic AI solutions.
You can input your requirements and policies into a dedicated solution such as Boomi Agentstudio, for example, to create and support organizational development and governance practices.
With Boomi Agentstudio you can create, deploy, and govern agents along with the APIs they work with, such as MCP servers for accessing tools and data.
Agentstudio works within the Boomi Enterprise Platform, which includes fundamental capabilities for agents, such as the Data Management and API Management capabilities, integrations, and workflow configuration.
When combined, probabilistic agentic requirements and policies become an agentic blueprint — enabling consistent development, deployment, and governance of agentic AI with tools such as Agentstudio.
The Intellyx Take
Generative AI agents were born and raised in the over-hyped environment of generative AI. But they are maturing in the enterprise IT landscape.
Early promotion from Salesforce, among others, predicted a world dominated by autonomous AI agents that not only figure out how to execute a given task, such as placing an order or solving a customer issue, but also automatically execute the task without human supervision.
But issues with erroneous content, hallucinations, and incorrect results are characteristic of LLM technology, which operates on a statistical match, not an exact match. AI agents, based on LLMs, inherit these issues.
The best approach is to identify where in the IT landscape agentic AI can add the most value – where its inherently probabilistic nature can become an advantage rather than a liability – and then establish strong governance, policies, and development patterns for use across the organization.
And then pick the right solution, such as Boomi Agentstudio, to help you build, govern, and orchestrate the AI agents.
In the next blog we’ll discuss the importance of establishing a foundation of trusted, high quality, and reliable data sources.
Copyright © Intellyx BV. Boomi is an Intellyx customer. Intellyx retains final editorial control of this article. No AI was used to write this article. Image source: Google Gemini.