Every time I get a press release about AI agents, I get a slightly queasy feeling. It is not quite as bad as that dizzy feeling I get every time someone insists on pitching me about vibe coding, nor is it the nails on a chalkboard feeling I get every time a PR rep sends me something with the word “convo” in it when asking for an interview or discussion with one of their clients.
Even so, AI agents are overhyped, under-defined, possibly dangerous, and extremely limited; they could possibly spell the end of life as we know it.
Also: AI agents aren’t just assistants: How they’re changing the future of work today
And yet everyone is all about agents. Microsoft did a series of announcements last week that promoted its extensive use of AI agents, not just for the enterprise, but for every Windows user. Google this week did a series of announcements that included AI agents in a wide range of applications, including writing your code. Because that’s not like letting the fox guard the henhouse — not at all.
But my real concern about agents is that they seem to be over-promised because there are so many limitations in the interaction of agents between ecosystems.
Into this crazy bouillabaisse of AI promotion and innovation, Salesforce enters with a fairly impressive dose of sanity.
Salesforce is introducing its Agentic Maturity Model, a framework that defines key stages of AI agent adoption and capabilities. This can help give us a common vocabulary when evaluating agent offerings from the various vendors who are flooding the market.
“While agents can be deployed quickly, scaling them effectively across the business requires a thoughtful, phased approach,” says Shibani Ahuja, SVP of Enterprise IT Strategy at Salesforce. “Understanding the progression of Al agent capabilities is crucial for long-term success, and this framework provides a clear roadmap to help organizations move toward higher levels of AI maturity.”
Also: The top 20 AI tools of 2025 – and the No. 1 thing to remember when you use them
See, there’s a big gap between the public’s picture of an AI agent and what’s possible. To vendors, AI agents are pretty much anything that can follow a bunch of steps using AI capabilities. This allows vendors to AI wash almost any offering, even if the true capabilities are fairly uninspired — or as in Apple’s case with Siri, largely vaporware.
But Salesforce gives us five levels:
- Level 0: Fixed rules and repetitive tasks
- Level 1: Information retrieval agents
- Level 2: Simple orchestration, single domain
- Level 3: Complex orchestration, multiple domain
- Level 4: Multi-agent orchestration
Essentially, we’re going from basic scripts all the way up to teams of agents working in concert to accomplish complex tasks across a variety of infrastructures.
This is very helpful because then we can look at an offering and determine that, yeah, it is “agentic,” but it is really not much more than a script — Level 0. Or, wow, you’re talking about an entire supply chain that’s automated, intelligent, and highly adaptive across vendors — Level 4.
Using the Agentic Maturity Model, let’s look into each of the five levels in a bit more depth.
Level 0: Fixed rules and repetitive tasks
Salesforce describes this as “automation of repetitive tasks using predefined rules, with no reasoning or learning capabilities.” A great example of this is your customized email filters. There is no real AI involved whatsoever, but those rules do help get the job done.
On one hand, I was thinking that it did not make sense for Salesforce to bundle in basic repetitive tasks in a model describing AI agents. But after thinking about it for a while, it did make sense. That is because you have to start somewhere.
Many of the repetitive tasks we script computers to help us with, such as email rules, programmers’ makefile scripts, and services that post updates on social media, are automated and save time.
Also: AI will change the trades too – and field service technicians can’t wait
But what if we could add intelligence to them? I’ve wanted intelligent email rules since generative AI was a thing. I don’t just want my email rules to sort my mail based on email address.
I want my email rules to determine press pitches and put them in their own folder, but identify those that are directly in my area of focus and flag just those for my attention. That’s an AI agent job, but it’s not at Level 0. Based on the framework, that challenge is probably a Level 2 agent.
But first, let’s move on to Level 1.
Level 1: Information retrieval agents
Salesforce defines this as agents that go out and pull in information and, as a result of that information, recommend actions.
They use the example of a troubleshooting agent, where you describe a problem, the agent does some searching, and then recommends a fix. Another example might be a shopping agent that can compare offerings and prices and make recommendations.
Also: Why scaling agentic AI is a marathon, not a sprint
But here’s where we run into agentic walls, a factor that Salesforce’s model does not cover. Let’s say you want to start a company to make buying recommendations and send people to Amazon so you can score affiliate revenue if they buy. Some vendors have tried this. But Amazon controls its data, and while you could certainly scrape web pages, Amazon would always have an advantage because they have all the raw data in their systems.
So, for an information retrieval agent to be able to do its thing, it needs to be able to get at that information. As we see vendors like Apple, Microsoft, and Google offer tools that they say can do informational retrieval actions, keep in mind that they are counting on you being solely in their ecosystem. Microsoft Copilot, for example, is sure not going to be able to search your Google Docs library and then your Notion database.
So, let’s extend the Level 1 definition to be agents that go out and pull in information from sources within their host ecosystem.
And now, let’s look at Level 2.
Level 2: Simple orchestration, single domain
Level 2 directly addresses the ecosystem issue by specifying that agentic activity take place in a siloed data environment. What this means is that all the data used is stored and available from one environment.
Notion’s AI is a perfect example of this. Notion’s AI derives its knowledge primarily from the selection of notes and documents you maintain in your own Notion archive. I, for example, write all my ZDNET articles in Notion and then transfer them to ZDNET’s content management system. The Notion AI could provide knowledge and activities based on my articles — but not based on any other ZDNET author, because those articles are not in that siloed data environment.
Also: I’m an AI tools expert, and these are the only two I pay for
As to orchestration, the idea is that simple orchestration means low-complexity tasks. We could, using the same data example, ask the AI to create a list of articles where I’ve discussed Salesforce, and then organize them based on Salesforce offerings discussed. But we could not get the AI to perform complex API connectivity and relate those articles to external deep research, for example.
Remember, the core of this level is the limitation of siloed data in one data store. This is what Salesforce calls single domain.
Do you want to work across data siloes? Then you are moving up to Level 3. That’s next.
Level 3: Complex orchestration, multiple domain
Now we start to get to what the whole agentic AI concept promises. Salesforce describes this level as “autonomously orchestrate multiple workflows with harmonized data across multiple domains.” In other words, your application will not break if you need to get data from different ecosystems or sources and integrate them using other systems.
Let’s be clear: this is very hard. There are really two choices architecturally for making this work. The first is using a series of APIs to communicate between the systems via microservices. In theory, this would work, but it would require all the information ecosystem vendors in the app to be willing to cooperate. There will be holdouts, and therefore there will be holes in the implementation of this system.
Here’s an example. There are a ton of social posting services that will take your posts, schedule them, and post them to Facebook, X, Instagram, Bluesky, etc. But there are two challenges. First, the more fringe socials are not always represented. Second, Facebook will allow these tools to post to Facebook pages, but not personal Facebook profiles. So if you wanted your AI to post to your Facebook profile, it is not available with Facebook’s permission.
Also: Anthropic is expanding Claude AI to the enterprise with domain-specific AI agents
That leads us to the second architectural choice, which is screen reading and screen clicking. In other words, the AI uses the browser the way you would and clicks, drags, moves, and types the same way you would. This blasts past all the limits of APIs because if a human can access the web page, so can the AI.
But it is highly unreliable. Trust me. I have built these solutions. Pages change constantly. I built a Twitter/X autoposter. I found that the structure of the X page changed almost weekly, which necessitated that my code re-learn the page every few days. It was a pain.
What does this all mean? Basically, Level 3 can work if all the domains and workflows are part of a cooperative ecosystem. That means there will always be outliers and parts of the solution that will not be able to be implemented.
And that leads us to Level 4…
Level 4: Multi-agent orchestration
Salesforce defines this as “Any-to-any-agent operability across disparate stacks with agent supervision.”
About 15 years ago, I built a subset of this. My AI Editor system consisted of a series of virtual servers, each hosting a single AI agent. I had a bunch of agents scanning news feeds, each looking for its dedicated area of interest. Those agents fed news pointers to an agent that wrote fresh news articles about the news items the first flock of agents identified. I had another agent that did nothing but identify images appropriate to each article from an enormous library of available images. And still another agent was the managing editor, assembling the article components, doing a final appropriateness check, and publishing the articles to my custom-built content management system.
Also: As AI agents multiply, IT becomes the new HR department
You can see how each agent took on part of the job and had to communicate with other agents. In my case, the agents and data sets had already been normalized — or I had agents who did the normalizing before sending on the data to the next AI team member. In Salesforce’s Level 4 definition, those data sources do not need to be normalized or even interoperable, and the agents may or may not be from the same orchestration class.
I do not think we will really see Level 4 agentic AI except in enterprise-level implementations where IT teams can control the scope of the overall project. But when it works, like with my AI Editor, it can have an enormous force-multiplying effect.
Can we do better?
I actually quite like the five levels and Salesforce’s definition for each of them. I think they fairly represent the stages of AI agentude and what sorts of tasks they can perform. But the name of the model, Agentic Maturity Model? Well, that could be better.
I recommend calling it the Agent Intelligence Scale, the Smart Agent Scale, Agent IQ Levels, Agent Power Index, Agent Mastery Matrix, AI Agent Stages, Agent Growth Scale, Automation Intelligence Scale, or Agent Intelligence Scale. Any of them would be stickier and more compelling than Agentic Maturity Model.
That said, I think this system works, and I will be referencing it as I talk about agents in the future.
Also: Crawl, then walk, before you run with AI agents, experts recommend
What about you? Do you find the idea of AI agents exciting, overhyped, or somewhere in between? Have you tried any tools that fit into Salesforce’s framework? What level do you think most vendors are actually delivering right now, and how many are just AI-washing simple scripts? Do you think we will see true multi-agent orchestration outside the enterprise anytime soon? Let us know in the comments below.
Get the morning’s top stories in your inbox each day with our Tech Today newsletter.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.