How to Prepare Your Sales Team for Agentic AI

Agentic AI is everywhere – or at least that’s what many would have you believe. But when you start to look more closely at many of the “agentic AI” use cases currently being promoted, it begs the question: How much of the hype for agentic AI is real (and ready for implementation today) and how much of it is still hypothetical value?  

Agentic AI vs. Generative AI and Multimodal AI 

To better understand what true agentic AI is, and which AI use cases are prior forms of AI under a different name, let’s quickly recap the AI innovation waves that took place over the past couple of years. 

First, there was generative AI. Generative AI refers to the ability of AI to ingest, analyze, and produce new structured or unstructured data, such as human conversations or computer code, using deep-learning models derived from a training data set or large language model (LLM). Think of it like the foundation on which multimodal and agentic AI use cases are built. In both of these newer forms of AI, generative AI is a critical component.  

Multimodal AI allows brands to point AI at use cases that may contain a mix of modalities such as images, voice, text, and video. For example, an auto insurer could use multimodal AI to assess automotive claim photo submissions from their customers, dealers, or insurance agents, which accelerates the end-to-end submission process and allows live agents to focus on other aspects of claim fulfillment.  

Now, agentic AI takes this foundation of capabilities and combines it with rapidly improving thinking and reasoning models to equip AI to handle an entire end-to-end workflow and verify its work along the way. In theory, this will allow AI agents to act on behalf of live agents or participate in tasks across an agent’s typical workflow, with less or even no oversight.  

Agentic AI Maturity Levels

Much like multimodal AI a year ago, agentic AI likely will not experience widespread adoption for approximately another year. According to a recent Gartner AI report, less than 1% of enterprise applications were using agentic AI in 2024.  

If we break AI use cases down into five distinct levels of agency, some weaker forms of agentic AI are already viable today – we just haven’t historically called them “agentic AI.”  

1. Do as I say. This first level of agency includes use cases where AI has been preprogrammed to respond in a specific, defined way. This has been around for a while and can be found in use cases like robotic process automation (RPA).  
2. Research and tell me. This second level is also widely available today. The multimodal AI auto claims process is one example of this level of agency in the wild.  
3. Recommend to me. This level of agency builds on “research and tell me” by taking multiple complex inputs, synthesizing them, and making a recommendation to a human in the loop (HITL) based on those inputs.  

When we’re talking about agentic AI – the kind that will really change the game for businesses – we’re really talking about the final two levels of AI agency. These last two levels unlock true AI-enabled decision-making that enables the cost-efficient vision many organizations have.  

1. Decide and tell me about it. The fourth level of agency involves minimal human oversight, beyond process and outcome auditing at the supervisor level.  
2. Do it and don’t bother me. The final level includes little to no human oversight, as AI agents take care of various processes and customer queries entirely on their own.  

These levels are largely not ready for implementation today, due to several different challenges at both the LLM and business process levels. But they’re getting much closer!

Challenges Facing Agentic AI 

Agentic AI adoption faces three big challenges. 

First and foremost is risk mitigation. As you increase the level of agency and decision-making delegated to an AI agent, you also accept the potential risks associated with those decisions. Depending on the industry, incorrect decisions could lead to lawsuits, public relations nightmares, and poor outcomes for the end customer.  

Secondly, these final two agentic AI levels require comprehensive training data that many organizations have not documented or curated for AI usage today. Unlike the AI that came before, agentic AI needs access to relevant knowledge bases, conversation data, business process context, as well as APIs to all the systems where the AI agent will need to “act” on its decisions. Many enterprises haven’t yet invested in the infrastructure and tooling, such as MCP servers, to enable this.

Finally, AI models are frozen in time. In other words, agentic AI today would make real-time decisions based on slightly outdated data. When we think about human decision-making, we’re constantly being bombarded by new information that refines our decisions. Data timeliness is a significant hurdle that needs to be cleared to help agentic AI make sound decisions.   

How Should Organizations Prepare for Agentic AI?

In addition to establishing a set of AI governance principles to organize and regulate a business’s AI adoption over time, there are a couple specific ways organizations can prepare for agentic AI.  

Start by experimenting with the first three levels of agentic AI now, while building the monitoring and support infrastructure that will enable risk-mitigated deployments in the future. These capabilities are ready today and can already deliver meaningful ROI. The learnings gained from experimenting with AI now will help focus and strengthen approaches over time. 

To be ready for levels four and five, organizations need to implement a business process management (BPM) approach to strategically analyze existing processes for optimization opportunities to spotlight which current processes and workflows might be ripe for AI to take over. By starting with a business process audit, organizations can prepare themselves for rapid agentic AI adoption when the time comes.  

Aaron Schroeder is director of AI Solutions at TTEC Digital, a global leader in customer experience orchestration.