AI Agents, Beyond the Hype
Why I think agentic workflows will radically change operations
After my recent chat with Scott Brinker, I’ve been delving deep into the world of agentic AI.
I try to avoid excessive hype around these topics, but it really hit me that agents are how AI will deliver radical business transformation.
Generative AI has already reshaped how we work—personally, I’ve replaced 95% of my internet searches with LLM chats. Yet, if LLMs disappeared tomorrow, I suspect most businesses, including SaaS companies, would still continue operating as usual.
I believe we’ll only see a truly disruptive impact from AI—and a dramatic reshaping of the workforce—when the following conditions have been met:
We identify the right type of labor that AI should replace
We enable AI to access the necessary tools to fulfill those tasks
We have the frameworks / platforms / expertise to easily make these AI-driven processes a reality
These three pillars define the foundation for agentic AI to move beyond hype and into meaningful operational change.
In this article I share some early thoughts as well as resources I’ve been exploring. I don’t claim mastery of this topic, but hopefully you can come along on the journey or contribute resources you’ve found helpful.
What is agentic AI?
While there’s no universally accepted definition of “agentic AI,” most experts agree on two core elements:
A Large Language Model (LLM)
Automated or autonomous execution
For Scott Brinker, the defining characteristic of agentic AI is tool use—that is, the ability for the AI to take action in other connected systems, either via APIs or manipulation of human interfaces.
Cobus Greyling similarly distinguishes between LLMs, which generate media in response to user prompts, and Large Action Models (LAMs), which “make real-time decisions based on tools at [their] disposal, [including] API’s and other integrations.”
For the purposes of this post, I’ll follow the more expansive definition provided by Anthropic in their research paper Building effective agents.
They define two types of agentic AI—workflows and agents—based on how they handle decision-making and tool usage.
Some customers define agents as fully autonomous systems that operate independently over extended periods […]. Others use the term to describe more prescriptive implementations that follow predefined workflows. At Anthropic, we categorize all these variations as agentic systems, but draw an important architectural distinction between workflows and agents:
Workflows are systems where LLMs and tools are orchestrated through predefined code paths.
Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.
I think this broader definition is useful, because:
In practice, vendors are using the term “AI agents” very loosely. This level of precision offers a vocabulary to distinguish various applications.
Finer distinctions enable operators and system architects to chose the correct design pattern for their application.
Why many first-wave AI applications fall short
Many current implementations fall short of their potential because they focus on single-step tasks rather than full-process orchestration.
A few examples of what I mean below.
Prompt-based builders
Many platforms now offer chat-based interfaces for building assets as an alternative to traditional UIs. Instead of using clicks to build out your workflow/campaign/whatever inside the platform, you use a prompt.
Perhaps analytics show great results from these features, but I’d be surprised. I think this paradigm suffers from two fundamental flaws:
The outputs of these prompt interfaces are usually not precise or complete enough to deploy to production without the assistance of experts. So they don’t enable true independence.
Tasks often span multiple systems. (For example, an email campaign may require configuration in Marketo, Salesforce, etc.) A human expert is still needed to stitch everything together.
Unless the volume of assets is significant, a truly proficient user probably won’t save much time by chatting, as they have already mastered the existing UI.
Example:
Zapier introduced a chat-based interface for creating Zaps. It's cool, but the initial output is still relatively imprecise and requires additional configuration.
As a proficient platform user, it doesn't really save me time. I'd rather just retain control and build from scratch. These “build via chat” implementations often seem like solutions in search of a problem.
AI content
In Scott Brinker’s and Frans Riemersma’s recent report, Martech for 2025, they surveyed nearly 300 marketers on how they’re using AI today.
Five of the top ten applications—and three of the top five applications—are related to content.
Obviously marketers are finding AI useful for content creation. What’s the problem?
My concern here is that some marketers are using AI to optimize for the wrong things—focusing on content creation velocity vs. content quality.
I’m not suggesting AI has no place in content. I often use AI as a thought partner for refining ideas or coming up with first drafts for technical or descriptive content like podcast show notes. I used AI to copyedit this article.
The issue arises when people rely on AI to draft LinkedIn posts or articles that should convey authentic human thought. When it comes to content, the currency is authenticity and original insight—precisely the things that AI lacks and is definitionally incapable of producing.
On the other hand, process-driven tasks are ripe for agentic workflows that automate without sacrificing creativity.
So how do we identify valuable applications for agentic AI?
We first need to describe the type of labor AI is best suited to replace.
The work follows a structured and predictable process
It requires human-style evaluation and decision-making within component steps
The output has a low requirement for originality or uniquely human contributions (e.g., intuitive leaps, novel insights)
The output has a high requirement for consistency and conscientiousness
These criteria describe a middle-ground between purely deterministic workflows (that can be fully automated with existing iPaaS-type tools today) and tasks that are highly dependent on creativity, originality, and unique perceptions such as strategy development or thought-leadership.
These middle-tier tasks often demand significant skilled labor but could be largely automated by AI.
Practical Use Cases for RevOps
Considering my own field of revenue operations, I can see many types of work that fit this description. These are just a few.
(Note: most of these use cases are still speculative, NOT things I’ve rolled out in production. But they should be theoretically possible with tools we have today.)
Campaign Operations Agent
I estimate that somewhere between 25-50% of marketing operations resources are allocated to campaign building.
A campaign ops agent team could execute across multiple systems, interacting with users to adapt requirements for individual campaigns, and manipulating GUIs where system APIs fall short.
Lead Management Agent
Many teams have complex rules-based routing systems—but they will never be perfect, because the underlying data is never perfect, and require manual human review.
An agent could flag cases where human intuition currently compensates for rules-based system gaps, such as:
Leads misrouted due to mismatched address data
False negatives (e.g., due to email domain variation), where a lead that seems to be from a small company is actually related to a large account
etc.
Data Quality Agent
Rules-based data tools, even those with “fuzzy” matching logic, can still leave a lot of gaps. Many ops teams spend hours manually fighting bad data.
An AI agent could perform that next level of review—evaluating and merging suspected dupes, normalizing address data, updating parent-child account hierarchies, resolving conflicts around company size between different enrichment tools, and so on.
Reporting and Analytics Agent
Ops teams spend a huge amount of time producing reports that go unread and have no impact. An AI agent can perform the first level of analysis, calling out insights and trends that a human should pay closer attention to.
My friend Grant Grigorian is building a platform that does exactly this.
Ad Operations
Performance marketing agency NewForm created their own agent for ad operations.
This agent builds reports, detects issues/opportunities, and proactively makes changes in different platforms to address them.
They estimate they’ve saved 90 hours per week—that’s over 2 FTEs!
Content Operations
While I believe the heart of content should remain uniquely human, there is a huge amount of labor surrounding the production of content that is ripe for AI replacement.
For example, producing my podcast involves:
Editing the episode
Creating a title
Creating thumbnail graphics
Inserting the graphic and intro sound file into the podcast
Inserting ads at relevant points (if the episode is sponsored)
Creating show notes
Publishing the audio file to podcast networks
Publishing the video file to YouTube
and so on…
It can take 5-10 hours per episode easily.
I use in-platform AI tools in various ways at each step of that process, but there is no single tool to execute the entire workflow. But potentially an agent could, while keeping the human in the loop for important decisions.
Learning resources
These resources are a great starting point for building your understanding of agentic AI and how it can transform operational processes. In future posts, I’ll share more insights from my own implementation journey.
Overviews and Architecture
AI agents are the new iPaaS and the next frontier of intense competition in digital ops orchestration - Scott Brinker
Building effective agents - Anthropic
The Rise Of AI Agents And Agentic Reasoning - Andrew Ng, Snowflake BUILD 2024 Keynote
Agentic AI Examples
Building AI Agents for Marketing - NewForm
112: Stephen Stouffer: The dawn of AI Ops and the practical wonders of combining AI tools with iPaaS - Humans of Martech
Research a contact using AI workflow - Kyle Coleman
Courses
Introduction to AI Agents - DAIR.ai (I took this paid course, and it was $39 very well spent. You will walk away with first-hand experience building multi-agent systems in Flowise.)
Platforms and Frameworks
Flowise - Open source / low-code AI app and agent builder
Relevance AI - No-code and looks very easy to use
CrewAI - No-code UI and development framework for multi-agent automation
Tray.ai - iPaaS tool that is leaning into AI in a significant way
Make - another automation tool with AI capabilities