Agentic AI is not one thing

"Agentic AI" has become very hot. It seems as if there's a new agentic announcement every other day - regardless of whether you're following ChatGPT (OpenAI), Co-Pilot (Microsoft), Gemini (Google), Claude (Anthropic), or North (Cohere). That's translated into a lot of executives demanding that their companies get agentic solutions.

"Agents" are AI systems that can perform tasks - typically as part of completing a bigger activity. You've experienced this when Gemini or ChatGPT goes and pulls some information off the internet to answer a question for you. In that case, the act of going out to the web was the "agentic" moment - it decided it needed more information, told another program to go search the web, and then used the information from that search to answer you.

Agentic AI is, indeed, truly an amazing step forward and deserves much of the hype it receives. However, there's quite a bit of confusion around what agentic AI is. Most crucially, it isn't just one thing. I think it's helpful to think of agentic AI in terms of how an organization might use it. In this way of thinking, there are three main categories:

- Operational efficiency: these agents automate tasks that a human might otherwise have done. Categorizing and filing documents, preparing and sending invoices, assigning specific work items to team members. When an agent does this, the primary benefit is that we've saved some human effort.

- Decision support: these agents help us make better designs, choices, and plans. Often their work begins with a request ("How many times did we have a client complain about our product?"), proceeds to the agent gathering and digesting resources (querying a database of client complaints), making sense of the resources gathered (grouping complaints by type), and then returning that information in a way that reflects the larger task ("Here are the categories and counts. Since you're preparing a report on our customer support, these are likely the most important categories...").

- Customer/user interaction: these agents also tend to have text-based interfaces and provide answers to users in need of help. In a sense, they're also providing operational efficiency gains - but they specialize in the customer or user-facing tasks.

Worth noting, there's even more agent diversity when you look inside these categories. For example, an agent that categorizes and files documents is very different from the one preparing and sending invoices. The upshot is that there are a lot of different kinds of agents out there - and even more that are on the horizon.

As a result, saying we're going to bring "agentic AI into a company" is kind of like ordering "food" at a restaurant. There are lots of kinds of "food" - and if we aren't more specific, there's a very good chance we'll be brought a bowl of chilli, when what we actually wanted was the strawberry-banana crêpe.

This reminds me somewhat of the early days of "cloud" in the 2010's. At that time, many companies thought that cloud was just one thing - and were sorely disappointed when they invested in "cloud" and paid for huge compute servers when what they really needed was massive data storage systems.

The good news is that every organization is fully capable of making the right agentic decision for their needs: purposeful consideration of what you want an agent to do combined with an assessment of the solutions available. Agents can be a huge benefit for your organization. Just pick the right one.