How to Use Agentic AI Without Breaking Your Process

How to Use Agentic AI Without Breaking Your Process

There's a version of AI adoption that looks productive but isn't. You open a chat window, ask a question, get a polished paragraph, move on. It feels fast. But you're still doing the real work. You've just added a step.

That's not what agentic AI is.

A chatbot helps you think. An agent helps you do. For a marketing team trying to scale without losing quality, that gap changes everything.

We've been working through that shift in real time. It hasn't been clean. Here's what we built, what broke, and what we'd do differently from the start.

Agents Act. Chatbots Talk.

The word "agentic" gets thrown around enough that it's worth pinning down.

Jasmin Bennett puts it plainly: "It's an agent that is deploying work for you." It doesn't just respond. It acts. Connect it to Slack and it drafts and sends messages. Connect it to Monday.com and it pulls action items from your Zoom transcripts and creates tasks. "It really does allow you to have a second arm to your work," she says.

That framing matters because it defines what agents are not. They're not a replacement for your thinking. They're not a content generator you fire and forget. They absorb execution. The tasks that eat your time without needing your judgment, research, formatting, brief assembly, those are what agents can take.

Rachel Bascom frames it as a shift from simulation to action: "It's a shift from thinking to doing. That's where it can really benefit us as experts." The goal isn't to hand off strategy. It's to protect the time strategy requires by clearing the 80% that doesn't need you.

Why Chatbots Hit a Wall

We used ChatGPT early and often. It was useful for small, contained tasks. But it had a ceiling.

"It never really got that much better," Blake Nielson says. "With chatbots, you see one piece of the puzzle. With agentic, you see the process end to end."

The deeper problem was durability. Every chat session starts from scratch. You can't lock in a client's voice, a research sequence, or a structured output format in any lasting way. "With the agentic side, you can give very specific sets of instructions," Jasmin says. "Prompt-based conversations are limited in how far they can go."

The industry learned the same lesson around content. Early on, everyone assumed AI would flood the market with cheap copy and devalue the work. That fear faded when it became clear the output was flat. People want a human point of view. What didn't go away was the need to produce content at scale. Agents are the answer to that. Not by writing the content, but by building the process that makes writing it faster.

Where Building Got Hard: The Content Brief

One of our most valuable deliverables is the content brief. It pulls from multiple APIs, accounts for both traditional and AI search, and takes real time to do well. Automating it was an obvious target. Getting it right took longer.

Rachel went in skeptical. "Automated content briefs have been around for a while, and the quality has not been there. You can't pump out something with a machine, have no human touch it, and have it be good." The question was whether modern AI changed that, or just dressed up the same old problem.

Early answer: same problem, new packaging. The culprit was hallucination.

"We would build something, verify it later, and Claude would just say: I made that up," Rachel says. It looked clean on the surface. It wasn't. Fixing it meant finding every gap and wiring in real sources to replace what the model invented. "Connect this API. Build this skill. Package it back in."

That process was expensive beyond just time. "Burning through tokens," Rachel says. "Insane usage. For stuff Claude made up."

The Fix: Stop Building Tall. Build Modular.

The real breakthrough came when Jasmin started studying how other teams structured their Claude setups.

"I found Reddit threads that led me to GitHub," she says. "People were automating full marketing stacks. And a lot of what I found were pre-built skills you could upload directly into Claude."

What she saw was a structure completely different from what we'd been building. Our original approach was one large markdown file with eight or nine sets of instructions. Pull the research. Build the content frame. Format the output. All in one go. "That structure is just very bloated," Jasmin says. "You're telling it to do so many things at once that it becomes inefficient."

The fix is a skills pipeline. Each step in the process gets its own file. A persona skill. A SERP skill. A content framing skill. Each is self-contained and can run on its own or chain into a larger flow. "If I'm creating a persona skill, it could be used for a content brief or a customer journey map or a UX report," Jasmin explains. "It's a lot more flexible, and you stop confusing the AI."

The results were immediate. Rachel breaks down what changed: "Every time you ran a new brief with the old setup, it had to review the full skill, the full chat, all the context every time. Just not efficient." The second payoff was quality control. "With the big monolithic skill, it was hard to pinpoint where things broke. You'd fix one issue and something else would fall apart." With modular skills, you find the broken piece, fix just that, and put it back. Everything else holds.

Blake names three core wins: better troubleshooting, lower token use, and transferability. "Those skills move across disciplines. A persona skill built for a content brief can also serve ad copy, meta descriptions, or a customer journey map. That's the real value."

Don't Let It Run Without a Plan

The sharpest lesson from all of this is simple: don't let the AI act before you've agreed on a plan.

"Tell it not to act yet," Rachel says. "Always say: give me your plan first. What are you going to do?" Let it run without that and you get a B-minus output that tempts you to keep patching rather than questioning the base.

Jasmin goes further. When something breaks, she feeds the existing file back into Claude and asks it to find the problem. "I'd ask: what in this process is using the most tokens? And it would parse that down and find the specific areas slowing the system." Using the AI to diagnose its own inefficiencies turns out to be more reliable than guessing from the outside.

Blake puts it directly: "You have to hand-hold. You have to ask questions and problem-solve with it. Otherwise you get something very surface level." Faster platforms don't always produce better work. "I usually uncover more depth when I'm using Claude because it's more deliberate about each step. The end result is more usable."

Let the Model Write the Prompt

One of the most useful things Jasmin shared: when you need a strong prompt, have a model write it.

"I'll describe what I need. Say I need to audit an ads account. The way I'd write it myself would be short: look at campaign performance, look at keyword trends, tell me what you see. I put that into ChatGPT and ask it to build the prompt. What comes back is paragraphs. 'You are a senior ads specialist.' Specific steps. Explicit instructions."

What the generated prompt adds is the detail that someone who knows their own job tends to skip. "It tells it to give me the top five opportunities this week, the most immediate things I can act on. Those are things I wouldn't naturally think to ask." LLMs know how they process input. When you ask one to write a prompt for another, you get instructions built for the way the model actually works.

Rachel's parallel move: ask for three to five options ranked by your stated priorities. Cost, timeline, ease. "It usually tells me which one is best. I might agree, I might not. But at least I can see the tradeoffs."

What Gets Built on Top

The brief process is where the infrastructure gets its clearest test. But the applications go further.

Rachel built a Reddit-based audience analysis tool in Replit. She mapped out a Chrome extension in Gemini, built it in Claude, and hosted it on Replit. "That was a good example of planning the work before touching any tools. I told it everything I had access to, asked for existing software I could try first, and then asked for a build plan if I wanted to do it myself."

The brand sentiment tool followed the same path. What started as a manual process of querying each major LLM about a brand, logging results, and compiling insights by hand became an automated skill. When content marketers saw it run, they immediately asked if it could fold into the brief process. It did. "We can now crawl those LLMs with a single search," Blake says.

LLM answer analysis worked the same way. Manual incognito queries across ChatGPT and Perplexity became a skill that runs automatically and packages directly into the brief alongside everything else.

The Bigger Principle

All of this points past any single tool or tactic.

Agentic AI doesn't work well when you treat it like a faster chatbot. It works when you treat it like a system: modular, tested, planned, and built to extend. The payoff isn't a better output on one task. It's a foundation that compounds.

"Slow down," Jasmin says. "Ask questions. Do the brainstorming before you build. Because it will execute exactly what you tell it to. And it's very easy to send it in the completely wrong direction."

The teams getting the most out of agentic AI are the ones treating the plan as the work, and the execution as what the agent handles.

Resources: 

Connect with Blake Nielson on LinkedIn: https://www.linkedin.com/in/blakejnielson 

Connect with  Jasmin Bennett on LinkedIn: https://www.linkedin.com/in/jasmin-rock 

Connect with Rachel Bascom on LinkedIn: https://www.linkedin.com/in/rachelbascom 

Looking for an agency that'll be worth the investment? 97th Floor creates custom, audience-first campaigns that drive pipeline and conversions. Get started here: https://97thfloor.com/lets-talk/. 

About Blake Nielson: Blake is the Head of Accounts at 97th Floor, where he leads organic growth strategy and client performance for enterprise and high-growth brands. With deep expertise in SEO and AI Search. Blake specializes in helping teams adapt to the evolving search landscape and drive sustainable demand. Known for building holistic, full-funnel organic strategies, he helps clients turn search visibility into measurable, revenue-driving results.

About Jasmin Bennett: Jasmin has built her career around one idea: that the best marketing happens when strategy and collaboration go hand in hand. She works closely with clients across every channel always thinking backward from the goal to craft campaigns that actually move the needle.

About Rachel Bascom: Rachel is the Head of Content Marketing at 97th Floor, boasting over a decade of expertise in the realm of digital marketing and a fervent dedication to crafting audience-centric content strategies. In her tenure, Rachel has been a trailblazer in the development of the content marketing department, playing an integral role in the transformative journey that positioned 97th Floor as a comprehensive, award-winning, holistic marketing agency.