Despite the prevalence and demonstrated superiority of ML and AI tools, many companies are still making decisions — even relatively trivial decisions, like determining messaging frequency — manually.

What’s worse, each manual decision affects more than a single individual; even simple decisions typically require time and attention from multiple teams, which leads to significant bottlenecks and time invested in low-value activities.

A quick example: Setting up a simple marketing campaign

Let’s say you want to send a simple message to a segment of high-value, active users. To launch a single message, most companies require:

  • An engineering team to pipeline, clean, and stage the data from your various data sources to a single location for combination/comparison
  • A data science/analytics team to create a user segment from a combination of mobile and web last-seen data and POS data to understand which users fall into the definition of “active, high-value users.”
  • An operations team to help pipeline the resulting segment (along with CMS and other data) into your engagement tools.
  • A creative team to draft copy and creatives for each campaign.
  • A CRM and/or product team to load the new campaign into the engagement platform, orchestrate the process (minding user frequency limits, balancing and prioritizing this messaging with existing messaging, triggered messaging, and requirements from different business units, etc.), and analyze and report on the results.

Bottlenecks at every step 

With the setup above, you have to wait for your engineering team to turn around the data processing. You have to wait for your data science team to create metrics, and you have to rely on many humans to take the information and use it both appropriately and promptly. Frankly, this doesn't happen often and is not the human's fault. Every user in your app has a different experience and different preferences, and there is no way for a human to be able to handle this flood of information. 

There is no analytics team large enough to analyze each individual user, and there is no CRM team large enough to act on individual insights, even if you produced those insights. 

The modern communication stack is limiting

What these teams end up doing — and what the modern communication stack is designed and limited to do — is to set a bunch of static rules: user segments, message triggers, and a lot of “if-then” logic. 

We know this isn’t the ideal way to effectively engage individual customers. Methods like segmentation are just tools that allow us to treat the true problem of effectively engaging with hundreds of thousands of unique individuals as something less complex that human teams can manage.

Agentic AI is currently revolutionizing this process.

What is an AI Agent?

An agent is a particular type of AI that can operate autonomously within boundaries. If you have used a Large Language Model (LLM), like Chat GPT, then you have used an agent. LLM’s take boundaries: if you tell GPT to talk to you in English, it will return English, or if you tell it to talk to you in code, it returns code. There is a lot of room for creativity within those boundaries, but it does take your guidance. 

Agents operate on behavior

More than LLMs, Agents don't only have to operate on language. They can operate on behavior as well. Most notably, they can operate on your event stream data — all of it. 

They can take all of this data, process it for an individual user, and then choose the best next steps to engage that user. They do this through four basic capabilities, some of which they have in common with LLMs, some of which are actually unique to a behavioral agent:


If you’re an eCommerce app, you ultimately want users to buy. If you’re a streaming app, you want users to listen or watch. If you're a gaming app, you want them to play. However, every time you reach out to a user or they come to your app, they're not going to perform the desired action — but they will often do something.

 A surrogate model is a particular type of machine learning model that will take a goal event and identify precursors and predictors. So the agent is able to assess intention. It's able to say, “This user isn't doing this thing. They aren't going to the finish line yet, but they're moving in the right direction.” This is a core human capability. We use it every time we talk to one another. Surrogate models allow a behavioral agent to mimic this capability.


Embeddings are an information map. If you take the full event stream from your app, agents can condense it into a representation that they can navigate easily to find similarities and differences between users. So, if a user is not converting but is doing other things, the agent that's assigned to that user can then navigate the embeddings to find other users who have similar behavioral profiles but are converting and ask those agents what worked for them and then apply those lessons to the user it's assigned to.


Edges are connections in a graph. If you have two pairs of shoes of the same brand, they are connected by an edge in the graph. If you have two songs connected by a genre, then they have an edge in the graph. If two messages have the same value proposition, there's an edge in that content graph. The edges allow agents to identify what users have experienced so far and what they haven't experienced and see what the next best step is. What's the next thing you could do to engage that user? 


All of this rolls up into weights - a way for a virtual agent to maintain working memory. If you send a message about how Nike shoes are popular right now, and the user responds positively to that, the agent has a representation of Nike, and it will increase the weighting of that representation. Further, it will have a representation of the weight of shoes, and will increase that as well. It will also have a representation of the notion of popularity, and it's going to upweight that, too. So, it allows the agent to learn what works and what doesn't and to maintain that memory over time. 

The Agentic Advantage

Just as it's possible to train an AI to complete a sentence, it is possible to train an AI to read the room, connect the dots, and apply lessons learned at a scale that human teams can’t possibly achieve.

Unlike human teams, AI Agents can provide truly individualized attention to each user. They can monitor each user’s preferences for aspects like timing, offerings, tone of voice and more, and serve each user accoerding to their needs. They can strategize with other user agents to find new and interesting ways to help get their assigned customer back on track. 

In short, they make marketing personal in a way that’s it’s previously been impossible to achieve.

If you want to learn more about Agentic AI, with examples for CRM and Product teams, please check out the video, below.