Consumer vs. enterprise AI agents

Are you wondering about the AI agents that people keep posting about on social media? Are you curious how only a few clicks are needed to create them? And are you wondering how these AI agents can improve productivity in your company?

No worries. As with many things AI these days, there is more smoke than fire.

Why are consumer AI agents relatively easy?

Many of the AI agents you read about on social media are what I call consumer AI agents. People are using AI agents in personal lives and as self-employed or small business owners. These AI agents include auto-replying to new comments on Instagram, summarizing emails from Gmail and sending them to WhatsApp, or automatically generating content for social media.

There are many vendors who popped up this year to offer no-code platforms to develop AI agents. They claim that no technical background is necessary. Anybody without any data or programming experience supposedly can use them.

Sometimes it works, usually when there is no additional information or data required. But whenever data is needed, not having any data and programming experience may be a detriment. Often, AI agents use Large Language Models (LLMs) to extract, digest, and summarize information. Like with any AI chatbot, the results may be hallucinations.

I tried some of them.

I created an AI agent to first extract a list of companies from a website based on their annual revenues. In the second step, the agent was supposed to find people on LinkedIn who work at these companies in specific roles. To verify this for accuracy, I run this process manually in parallel. The output from the first step was incorrect after comparing it with real data. In the second step, the AI agent wasn’t able to find people in these specific roles, although I was able to do that manually.

The failures of this AI agent were due to LLM hallucination and inconsistent data structure. During the first step, the LLM ignored instructions and “explored” other alternatives. And in the second step, it struggled with the data structure on LinkedIn.

AI agents may work well in situations with streamlined data flows, low complexity with only a few interdependencies, and requiring tools and code repositories that are fairly common and vetted by experts. Apparently my task was already too complex for this simple AI agent that I created.

Despite all the errors and hallucinations, consumers keep using AI agents because the consequences are negligible. Nobody loses money when the comment reply on Instagram sounds cringe. Nobody gets hurt when the email is incorrectly summarized. And nobody gets fired for posting AI slop on LinkedIn.

The consequences in enterprise are consequential.

How are enterprise AI agents different from consumer AI agents?

The key difference between consumer and enterprise AI agents is not only where they are used, but also the organization, process, and data complexity they need to address.

In business, AI agents don’t run independently of the business process. They are embedded in existing processes, they connect with company’s software systems, and they use enterprise data to execute tasks.

Without repeating much of my blog article from a month ago, AI agents are created to address a specific business need. They analyze data and triggers to take actions, sometimes independently of a human. Hence, they require a clear understanding of the business processes they are supposed to automate. Enterprise AI agents range from complex AI systems to intelligent process automation with a pinch of AI thrown in.

And here is where data and programming experience comes in.

People experienced with AI know that AI needs data. They know that data is never perfect and oftentimes needs to be cleaned and prepared before it can be used. And they know that LLMs tend to hallucinate and their outputs often need to be verified for accuracy.

Experts also know that errors multiply. An AI agent involves several process steps. The more steps, the higher the overall error rate. Being generous and assuming an average accuracy of 90% at each step, the total average accuracy is 73% with only three steps. The automation will be accurate only 73% of the times it runs. And that only with three steps. Just imagine what the accuracy would be for a 10-step process left to its own devices … 35%!

Companies such as ServiceNow and Salesforce are successful with AI agents. But for them it’s simply AI and they use the terms “AI agents” or “Agentic AI” for marketing purposes. They employ experts – data scientists, data engineers, ML engineers, software engineers, etc. – since AI agents are nothing more than AI embedded in process automation. You need expertise for that. Drag-and-drop solutions from no-code platforms won’t suffice in enterprise.

Why is technical expertise needed for successful AI adoption?

How are AI agents relevant for AI adoption in business, you may ask. Good question since after all, most AI agents are just rebranded AI solutions or intelligent process automations.

Keep in mind that successful AI adoption happens if the AI solutions are embraced by business users because they see value in using them. If these AI solutions don’t work, provide false recommendations, or screw up business processes, they will never be accepted.

Despite the marketing claims of AI providers, AI agents are not easy in enterprise. They need technical expertise to function properly within the enterprise system. They need the same level of technical rigor and attention that have been applied to all AI solutions. Without that, there won’t be any AI adoption.