Last May I wrote an article here titled "The Agentic Curmudgeon" after my time on the show floor at Enterprise Connect witnessing the staggering amount of agentic washing that vendors had embraced. Lots of stories about how non-agentic products were now agentic through and through. Great stories with no reality behind them.
Today, I'm in the process of doing my "Conversational AI Wave," an analysis of leading vendors that provide self-service applications for customer service. As I dig deep into these vendors, I'm pleased to report that significant progress is being made with capabilities that pass the agentic sniff test.
To review, here are the six characteristics of an agentic application according to Forrester Research:<
- Reflection — The ability of an agent to assess its performance of a task and find ways to improve.
- Planning — The ability of an agent to understand the requirements and tasks that need to be performed to meet a higher-level goal.
- Memory — The ability to maintain state to know where in a process the agent is so it can take care of business in the right order.
- Tool use — The ability to leverage external resources, such as other agents, or to make API calls to back-end systems.
- Multiagent collaboration — The ability of a core agent to orchestrate the activities of multiple agents, allowing a system to break tasks down into component pieces and have different processes own and manage the execution of various tasks.
- Autonomy — This is the result of the above, an autonomous agent that can perform complex tasks with no human intervention.
Here are some things I'm seeing in these Wave demonstrations that are crossing the line into real agentic capabilities territory:
- Reflection — Agents that can look at bot interactions, score them, and recommend changes based on those scores. This type of functionality is being used to find where bots break down, identify coverage gaps in a knowledge base, and for application troubleshooting and regression testing.
- Planning — With modern conversational AI systems, an agent can be given a goal such as make an insurance claim for an accident. Instead of an endless program of explicit steps and many exception paths, you just give the agent general instructions for identifying a set of tools (other agents, back-end systems, scheduling apps, etc.), identifying information sources for answers, and rules about the order tasks need to be performed. This will be enough; the agentic system will take things from there. In many workflows the rules are too specific and limited to allow this sort of flexible approach to executing a process, but there are places where this will work and make a difference in customer service.
- Memory — A customer can walk away from a conversation with a chatbot on her computer, hop in the car, and call an 800 number from her cell phone and pick up the process right where she left off with all the needed context in place. You can even interrupt a voice bot by selecting something in a carousel in a chat conversation with the same bot. Think picking a vacation rental while the bot is describing it and having the bot stop talking and confirm the reservation.
- Tool use and multi-agent collaboration — Agents can call other agents to perform tasks such as getting information from a customer, finding and summarizing information from a knowledge base, or updating a service ticket in CRM. All this happens in the same conversation sharing the same memory. Model context protocol (MCP) is real; agents use it today to call agents or access back-end systems or other resources.
- Autonomy — Not even the craziest vendor is offering to let its agents run free without any supervision, but these bots are now capable of some surprising things.
Net-net: While conversational AI vendors aren't providing fully autonomous agentic bot workforces, they do provide systems that have agentic capabilities that make them more powerful, efficient, and easier to train than what has come before.
That doesn't mean that organizations are ready for prime-time deployment of these bots. Most organizations don't provide generative AI self-service options, much less anything agentic. But my frustration last May around the crazy agentic hype has turned into excitement to see how this rolls out into the world.
Max Ball is a principal analyst at Forrester Research.