The hottest topic in Enterprise AI adoption these days is Agents. It's time to look at what this actually means, why Enterprise customers find them relevant, and how the big boys are implementing them today.

The key takeaway

For tech sales: As more companies adopt AI across their current applications or workflows, there will be an increased focus on performance, customization and value realization. AI Agents are well positioned as a "balanced" choice and will likely become the dominant form of Enterprise GenAI implementations that most users interact with.

For investors: How a tech company is approaching the topic of AI Agents is a good litmus test for their comprehension of what realistic Enterprise adoption would look like for most customers. Companies still stuck in the "GenAI features" phase or those that are trying to produce very "general" Copilot implementations are going to be behind on the adoption curve. From next year, the big topic to be on the lookout for will be actual benchmarking between companies offering similar types of agents natively within their products. This is where technical capabilities and access to top 10% ML talent will really pay off.

So what are these so-called “AI AGENTS”

Well, if we let the AI speak for itself:

AI agents are autonomous or semi-autonomous software entities that leverage artificial intelligence techniques to perform tasks, make decisions, or provide insights on behalf of users or organizations. In the enterprise environment, these agents are designed to enhance efficiency, reduce operational costs, and drive innovation by automating complex processes, analyzing large datasets, and interacting with users in a human-like manner.
Key Features of AI Agents:
• Autonomy: Ability to operate without continuous human guidance.
• Adaptability: Learning from data and experiences to improve performance over time.
• Interactivity: Engaging with users or other systems through natural language or predefined protocols.
• Goal-Oriented Behavior: Designed to achieve specific objectives aligned with business needs.

If we look at how ServiceNow positions their “Virtual Agent” i.e. customer support within the platform:

Source: ServiceNow Virtual Agent Data Sheet

Let’s not forget also the very meme-able AGENTFORCES amongst us:

Source: Salesforce Agentforce product page

So if we have to translate this now into tech sales speak:

Agents are machine learning-powered automated workflows.
We call them agents because if we also add LLMs to the mix, users can interact with them, and the workflows can also better handle tasks that require natural language comprehension.
While developers can build very custom agents already by making choices on multiple parts of their stack, in order to drive faster adoption, the software vendors are trying to provide pre-packaged modules that can be easily enabled. This removes the friction of having to build a new workflow from scratch every single time.
Agents are quickly becoming the "default" way for Enterprise adoption because their cost-to-value ratio is better in most scenarios. The majority of agents operate on small models, optimized for the specific workflow.
Agents are easier to pitch to customers since you can have faster testing and implementation periods that allow you to visualize the potential ROI of deploying them.
Agents are not a replacement for complex implementations for difficult use cases, but they do solve a growing demand for "easy productivity gains."

Do you have some numbers to back this story?

Funny that you mention that. Let’s look at the latest LangChain State of AI Agents report:

Source: LangChain State of AI Agents Report

Based on their survey among 1300 professionals (mix of tech and non-tech companies), half of the orgs already had deployed AI Agents in different parts of their company; almost 80% were currently working on doing that.

Predictably, the majority of use cases revolve around natural language. This, however, can be a bit misleading - code generation and data processing are "arguably" IT workflows that are technical in nature.

What becomes quickly very clear is that LangChain is fundamentally a developer-first org and its product marketing team has not thought deeply enough about the more important question: Which tasks currently deliver the highest ROI? Can you quantify the difference vs the non-agentic version of this workflow?

While there are no ROI figures, at least we get a direction of which types of activities would be considered the most valuable ones.

The most interesting aspect from a cost-to-value perspective is that the price of running the agents in production is not actually the biggest blocker to adoption - whether they do something useful is the bigger issue. Now we also need to calibrate these findings by putting them in context, considering who the data comes from - LangChain has a very technical audience focused on technical aspects of the operation of AI Agents.

If we take a look at VC-driven research, things look a bit different:

Source: Merlo Ventures 2024: The State of Generative AI in the Enterprise

While some of the core natural language tasks remain (meeting summarization and copywriting), there is a bigger focus on replacing workflows rather than just performing tasks.

More importantly, the primary focus when approving a pilot launch for a GenAI product from a leadership perspective is ROI and industry customization. This is a different argument than purely the cost-to-run; the priority is value realization.

Let’s see these agents in action

I thought you would never ask! Here is the most high-profile launch from last week (Microsoft Ignite event):

If 2024 was predominantly dominated by Copilot-style implementations, 2025 will be driven by agentic workflows as the primary interaction that B2B users will have. The implementation by Microsoft here is unusually intuitive for one of their products - users can interact with custom agents for specific tasks whenever they need them from a single overview.

More importantly, they are baking in quantifiable ROI as a core part of the performance analytics suite:

Source: “Introducing Copilot Analytics to measure AI impact on your business” article by Microsoft

While the building and deploying of agents is mostly on top of the existing AI tech stack, here are some additional companies that specialize in this type of product for further research:

This is not a perfect overview, but it gives you a general idea of potential companies to research around Agent Frameworks and how these could be added within products that you sell.

As we go into next year, it will be interesting to see from the hyperscalers a breakdown of revenue that specifically calls out inference from agents vs. "features" type of implementations. I would expect the figures to be in the high 90s.

The Tech Sales Newsletter #68: The state of AI Agents in late 2024

The key takeaway

So what are these so-called “AI AGENTS”

Do you have some numbers to back this story?

Let’s see these agents in action

The Tech Sales Newsletter #69: The WIZards among us

The Tech Sales Newsletter #67: On Palantir and high conviction bets