Let's start by zooming out for second.

Gartner is a 40 billion dollar agency that's been around for over 30 years, and they're calling AI part of a "combinational disruption" that's forcing us to reexamine our assumptions about strategy, tactics, and innovation.

In their words, these tools are the ones that make us more human.

Look at the industry trends and you'll see why. ChatGPT alone has a billion active users. It's the fifth most visited website in the world as of April 2026, sitting behind Google, YouTube, Facebook, and Instagram. Usage is up by 2.5 billion users per day, which is closely catching up to Google's 14 billion searches a day. And Google itself has implemented AI Overviews, which, depending on which study you read, appear in as many as half of all search responses.

The shift is especially pronounced in marketing. According to data from CMO Alliance, 77% of marketers are already using some form of AI agent in their day-to-day work, and another 22% are planning to. Everywhere you go, you hear about how agentic AI is changing the face of marketing.

This article is adapted from Avi's keynote at the AI for Marketer's Summit, San Francisco. Get your ticket for this year's summit to access even more cutting-edge insights on applied AI in marketing.

When I recently asked a room what their current experience was with AI agents in their marketing efforts, the answers were telling. Some were planning to start soon. Others said they were seeing some wins but hitting roadblocks. Very few said they were getting great results across the board.

That tracks with what I'm seeing in the wider market. Most organizations are building from the bottom up. Teams are dabbling with perplexity, Claude, custom GPTs, and whatever else they can get their hands on. The result is a disparate, fragmented approach with no top-down strategy, no real governance, and no thought given to scalability. That's why we're seeing wins, but also why we're hitting walls.

What agentic AI workflows actually are in marketing

Let me break this down simply. An agentic AI workflow is essentially an action designed to achieve a goal. There are three core components: AI agents, prompt engineering, and large language models.

AI agents are the doers. They autonomously carry out tasks. They make decisions on your behalf. They follow the orders they're given and execute with precision without needing constant human intervention.

People often conflate AI agents and agentic AI, and they aren't quite the same thing. AI agents are single-task specialists. Agentic AI describes self-managing teams of agents working together. One follows orders. The other runs the show.

A simple example helps. An AI agent might summarize a PDF for you. An agentic AI system has multiple agents connected together to handle complex, multi-step tasks, often with self-critiquing loops built in to keep optimizing as they go.

Prompt engineering is the second piece. Well-crafted prompts guide your AI agents to achieve the right outcome, but you always have to be on guard for hallucinations and deepfakes. Prompting is iterative work. As marketers, you want your AI agents to understand the nuances of a task so they generate responses that genuinely fit the audience.

The third piece is the LLM itself, which serves as the backbone for generating content. LLMs enable your AI agents to create tailored messages and understand diverse audiences.

The spectrum of agents

You'll hear a lot about the "spectrum of agents" or the "spectrum of vibe coding tools." When you zoom out, agents really do live on a spectrum.

On the simpler end, you have retrieval agents. In the middle, you have task-based agents. On the more advanced end, you have autonomous agents. The boundaries between them are fluid. You can start with a retrieval agent and optimize it into a task agent, then eventually evolve it into something more autonomous.

Best practices for retrieval agents

When you're building retrieval agents, there are three best practices that matter most: reliable data, appropriate audience targeting, and trustworthy delivery.

Think of your data as the foundation. Your audience-centric design is the strategy. Trust and credibility shape how the information is actually received.

As marketers, we usually have access to a lot of source material. Product marketing assets, competitive documentation, key messaging docs, campaign launch material, internal information. The question is how you overlay that material with your buyer personas.

In B2B, this matters more than anywhere else. You need to think carefully about the ratifier, the decision maker, the administrator, the champion, and the extended buying committee. Audience-centric design gives you a way to clearly delineate between those personas.

Best practices for task agents

Task agents are where the workflow loop becomes critical. The workflow loop is the heartbeat of a task agent. You think about your input, the model you're tapping into, the tools you're connecting to, the results you want to see, and the iterative loop between model and output.

A great example for marketers is integrated marketing campaign execution. Every team has to fill out a brief, develop key messaging, set UTM parameters, define ICPs, and run through countless other steps. In my 20 years in marketing, that whole process has been consistently cumbersome and time-consuming. Building a task agent to automate parts of that workflow saves enormous amounts of time.

Another use case is email communications and automation. Marketers do a lot of outreach through newsletters, often pulling content from multiple sources. This one varies a lot from team to team, but the foundation is data quality. Recently, I saw a webinar with n8n where someone built a pretty sophisticated dynamic newsletter in a few hours using n8n Cloud, trigger nodes, and Bright Data.

The third use case is social media and content automation. Here's a prompt I tested:

"I'm looking to maximize the reach of my thought leadership blog content on my website. I want to automate creating posts across LinkedIn. Provide me with detailed automation flows that use AI tools like Make.com, n8n, and Lindy. Show how I can track the blog content in a repository that includes the date, the AI post, and the status."

I ran this in Claude with the workflow loop already pre-loaded. Claude returned three options. Make.com, which is low-code, no-code. n8n, which requires JavaScript skills I no longer have. And Lindy, which is more prompt-driven.

I chose Make.com for the demo. I created an account, pre-configured the agent, and built out the scenarios. I connected an RSS feed from the HubSpot Marketing blog, mapped the XML fields to OpenAI through an API key, fed that into LinkedIn, and tracked everything in Google Sheets since I didn't have Airtable handy.

When I ran it, the workflow pulled that day's blogs, populated the Google Sheet, and posted directly to my LinkedIn profile with the prompts I'd designed, the right hashtags, and a link back to the original blog source. The whole thing took me about an hour to set up.

I built it as a one-time run rather than an ongoing loop, because the goal was to spark ideas. Now that you know how the pieces connect, you can scale this approach. Maybe you automate blog posts that currently live as islands on your website. Maybe you build something departmental. You can swap in Gemini, Perplexity, or another source as your reasoning layer. Make.com works well. Lindy.ai is another good option and is even more prompt-driven, so you can be up and running in minutes.

None of this required real coding. I just connected applications and built the workflow.

Stepping into autonomous AI agents

Autonomous agents are where things get more advanced. A lot of B2B marketers in the room are customers of 6sense or Demandbase or another ABM platform. So am I.

At my company, we use 6sense to drive outbound on what we call 6QAs (6sense qualified accounts), which others call MQAs. We see high volumes of these accounts delivered to our sales and BDR teams, but our conversion rates have been hovering between 5% and 8%. We want to get to the 12% to 15% range that aligns with our industry benchmarks.

What we've learned is that the gap isn't really an SLA issue between marketing and sales. The real challenges are redundancy in data, low velocity due to synchronization issues, and poor connect rates because being a 6QA doesn't automatically mean the account is ready for a conversation.

So we're building a timing-based optimization agent. It uses Copilot as the anchor, connects our Salesforce and 6sense instances, and applies rule-based logic to read digital body language, frequency patterns, when buyers are most active online, and seasonality factors. Then it feeds that intelligence to sales as a triggered push notification.

It's a complex autonomous agent and we're about halfway through building it out, working with our center of excellence team. It does require real coding to connect the applications and make sure the fields are feeding bidirectionally in the right way.

Steps to build your own AI agent

For most marketers, the path to building an agent looks something like this.

Pick a small problem. This is where most people go wrong. They try to build something general that can do everything. Stay narrow and well-defined so you can iterate and predict where things might break.

Choose your base LLM. Claude, Gemini, Perplexity, GPT-5. Each has its own strengths and weaknesses.

Define how the agent should act in the real world. Think beyond conversation. Think about the tools you can connect and the actual tasks you want to accomplish.

Build the workflow loop. This is the heartbeat. Without it, you just have an expensive chatbot. With it, you have something automated and repeatable.

Step one gives you clarity on success criteria. Step two gives you the reasoning engine. Step three gives you real-world capabilities. Step four ties everything together into something that works independently.

For more advanced builders, the work gets more technical. Connecting to a RAG, working with JSON, managing short-term context, using command-line interfaces like Claude Code, Cursor, or Windsurf. That's where you need real application development skills.

One thing I've learned from people who develop and write applications: you have to be agile. Experiment quickly, fail fast, make improvements, and keep your scope tight. Focus on one working agent that does its job well rather than trying to boil the ocean.

Common questions and where this is heading

A few questions keep coming up that are worth sharing.

Should I position agentic AI as a behind-the-scenes enabler or a visible differentiator?

This really comes down to how you're building. The technology helps you be more efficient and better at connecting the dots. From a customer experience perspective, it improves efficiency in your outreach. Most customers appreciate the benefits of intelligence and automation.

How do you reconcile corporate policies with the explosion of new AI tools?

Gartner has noted there are 3,000-plus companies building AI agents today. There's a lot of noise. Ultimately, it comes back to your business goals and objectives, then matching tooling to those goals. The same challenge has always existed in MarTech, where there are now over 9,000 vendors and the number keeps growing.

What are the non-obvious AI use cases beyond chatbots and campaign automation?

Connecting applications is one of the strongest. Another one I keep hearing about is BDR outbound automation, where you have large volumes of leads or accounts coming in from non-preferred segments. The customer acquisition cost is too high for human reps to chase small deals. That's where automation can fast-track engagement through conversational emails and similar outreach.

How do I approach governance and accountability when an autonomous agent goes outside SOP or brand values?

Ask me in two months. We'll have our agent fully deployed by then and will have a clearer picture of the outputs and effectiveness. This is a real challenge across the industry. Companies are building bottom-up, then leadership in the boardroom is asking about ROI on AI, and the teams can't always answer because there's no top-down approach to scaling.

If you take one thing away, take this. Start small, stay specific, and iterate. The marketers who win with agentic AI won't be the ones who build the most complex systems. They'll be the ones who pick the right problem, design a clean workflow loop, and keep humanizing the output every step of the way.