Salesforce Unveils 5-Level AI Agent Framework, Dispelling Hype

Every time a press release about AI agents lands in my inbox, I can't help but feel a bit uneasy. It's not as bad as the dizziness I experience when someone tries to pitch me on "vibe coding," nor is it the cringe-worthy sensation I get from a PR rep using the word "convo" to ask for an interview. Yet, the hype around AI agents makes me skeptical. They're often overhyped, poorly defined, potentially risky, and their capabilities are quite limited. Some even suggest they could spell the end of life as we know it.

Despite my reservations, AI agents are all the rage. Microsoft recently made a series of announcements promoting their use of AI agents not just for businesses, but for every Windows user. Google followed suit this week, integrating AI agents into various applications, including coding tools. It's like letting the fox guard the henhouse, isn't it?

My biggest concern is that the promises made about AI agents often don't match their limitations, especially when it comes to how they interact across different ecosystems. It's a chaotic mix of AI promotion and innovation out there, but Salesforce brings a refreshing dose of sanity to the conversation.

Salesforce has introduced the Agentic Maturity Model, a framework that outlines key stages of AI agent adoption and capabilities. This model provides a common language to evaluate the flood of agent offerings from various vendors. Shibani Ahuja, SVP of Enterprise IT Strategy at Salesforce, emphasizes that while agents can be deployed quickly, scaling them effectively requires a thoughtful, phased approach. "Understanding the progression of AI agent capabilities is crucial for long-term success," she says, "and this framework offers a clear roadmap to help organizations achieve higher levels of AI maturity."

There's a significant gap between the public's perception of AI agents and what's actually possible. Vendors often label anything that follows a set of steps using AI as an "agent," which can lead to a lot of AI-washing. Even Apple's Siri, which has been around for a while, seems to be more hype than substance.

Salesforce's Five Levels of AI Agent Maturity

Salesforce breaks down the maturity of AI agents into five levels:

• Level 0: Fixed rules and repetitive tasks - These are basic automations with no AI learning involved. Think of customized email filters. While it might seem odd to include such basic tasks in a model about AI agents, it makes sense as a starting point. Imagine if these simple tasks could be enhanced with AI to become more intelligent. For instance, I've always wanted email rules that could sort press pitches into a folder but flag those directly relevant to my work. That's a job for a Level 2 agent.
• Level 1: Information retrieval agents - These agents pull in information and recommend actions based on it. A troubleshooting agent that searches for solutions or a shopping agent that compares prices are good examples. However, there's a catch. These agents are limited to the data within their ecosystem. For example, Microsoft Copilot won't be able to search your Google Docs or Notion database. So, Level 1 agents are effective within their host ecosystem.
• Level 2: Simple orchestration, single domain - This level addresses the ecosystem issue by focusing on a single data environment. Notion's AI, which derives knowledge from your Notion archive, is a great example. It can organize my articles based on specific criteria, but it can't integrate external data. This level is about low-complexity tasks within a single domain.
• Level 3: Complex orchestration, multiple domain - This is where AI agents start to fulfill their promise by orchestrating multiple workflows across different domains. It's challenging because it requires either cooperative APIs or screen reading/clicking methods. Social posting services illustrate the challenges; they can post to various platforms, but not all, and not to personal profiles on some networks. Level 3 can work if all domains cooperate, but there will always be limitations.
• Level 4: Multi-agent orchestration - This level involves any-to-any-agent operability across different systems with agent supervision. I once built a system like this with multiple AI agents working together to produce news articles. Each agent had a specific role, and they communicated seamlessly. However, achieving this level outside of enterprise settings seems unlikely due to the need for extensive control and integration.

The Agentic Maturity Model is a useful framework, but the name could be more engaging. Suggestions like the Agent Intelligence Scale or Agent Mastery Matrix might be more compelling. Regardless of the name, this system is a valuable tool for discussing AI agents and their capabilities.

What do you think about AI agents? Are they exciting, overhyped, or somewhere in between? Have you used any tools that fit into Salesforce's framework? What level do you believe most vendors are actually delivering, and how many are just AI-washing simple scripts? Do you think true multi-agent orchestration will be seen outside the enterprise anytime soon? Share your thoughts in the comments below.