CES 2025 has now come to a close, and one of the concepts everyone buzzed about was agentic AI. More than a few people took an opportunity to duck into a corner and frantically scroll through Google results to figure out what that is. If that was you, it’s okay – this is a safe space!
According to one study, 82% of companies are planning to integrate agentic AI in the next 1-3 years, and some large companies are already using them. So…
What is agentic AI, and how does it work?
RAGs or how to make sure the agents have the right input
Agentic AI is also unique because it’s iterative for your business data.
Think about what it’s currently like to use an AI chatbot or generative AI tool. You interact with it, and it responds to you based on what transpires during that single interaction. Maybe some of that data goes to improve the larger LLM that the chatbot is working off, but each interaction is individual. If you save your interactions, the LLM can learn from them in that single thread, but others using the LLM don’t benefit from that.
AI agents, on the other hand, are constantly being improved through a feedback loop, referred to as a “data flywheel”. After interactions, users can record nuanced information about how useful the agent was, and that data is fed back into the system to make the model more effective. The result is an AI that is constantly evolving to get better at the job it’s doing, and more effective at decision making.
So what could agentic AI look like in practice?
How can you get ready for agentic AI?
As we love to say at Pickaxe: your AI is only as good as the data it’s pulling from. You can have the coolest AI model in the world, but if it’s trying to RAG* from a database that’s a mess, it’s not going to get very far. Getting those data hygiene ducks in a row – developing a layer that is prepped and queryable for the AI agent to work with, ensuring that data is anonymized and in-house – is, as always, the unglamorous-but-crucial first step to making any AI run smoothly.
Additionally, because AI agents improve based on user feedback, your team needs to be ready to train the AI. That means everyone needs to know how to use the tool – this seems obvious, but prepping everyone on how to make an AI agent part of their workflow does take time and effort. Getting your team on the same page about how to give feedback, where to focus, and what to highlight for the agent can make it a lot more successful.
Start with safeguards. When we rolled out a media optimization agent for one client this winter, we didn’t give it access to make bids or traffic ads – instead, it made recommendations that the marketers could see, discuss, and then we stored and evaluates over time to see how much improvement they would have (or did) generate.
If everything pans out, agentic AI promises to be the return on investment in gen AI that companies have been making for years. But without the data fundamentals in place, any AI tool is about as useful as a digital paperweight.
*Let’s all agree this can be a verb now!
Let us help you lay a strong foundation
PS.
This technology is moving incredibly rapidly, so to help get you up to speed, we’ve pulled together some helpful links:
