AI moved forward, marketing did not

Marketers were among the first professionals to genuinely embrace generative AI. We opened ChatGPT, typed something, and got a result that was kind of amazing. We were among the first to have AI “wow” moments, and started weaving LLMs into our days. By most measures, marketing led the early adoption curve.

And then, somewhere along the way, we stopped evolving.

Eighteen months in, a surprising number of marketing teams are still doing essentially the same thing they did on day one. They open a chat window, type a request, edit the output, and move on. The workflow around AI hasn’t really changed. And only one step in the old process has been swapped out. We replaced the blank page with a draft. Everything else has, for the most part, stayed exactly as it was.

How we got stuck

It happened for understandable reasons. Inertia is one. It is just easier to keep doing things the same way (see my article about paving cow paths).

The early outputs burned trust. The first time you asked an AI to write something genuinely important, and it came back hallucinating facts, using your competitor’s name, or producing something so generic it hurt, you learned something. You learned to keep AI on a short leash. You used it for low-stakes drafts and kept real judgment firmly in human hands. That was rational at the time. The problem is that the lesson calcified into a habit.

Nobody owned AI adoption. In most marketing orgs I’ve spoken with, AI usage grew like kudzu: Everywhere, and with no structure. Individual contributors developed their own prompt tricks. Tools proliferated. Someone bought five subscriptions, someone else bought three different ones. There was no shared workflow, no center of gravity, no one asking the bigger question: what should this actually change about how we work? Without ownership, experimentation stayed individual and shallow.

The number of tools was genuinely overwhelming. At last count, there are over 1,000 AI tools marketed specifically to marketing teams. If you spent 30 minutes evaluating each one, that’s over 500 hours. Most marketers did what any reasonable person would do: they picked one or two familiar tools and used them for everything. Which mostly meant text generation. Which mostly meant the chatbot loop.

And so the pattern of prompt, response, and copy-paste got locked in. The ceiling of ambition remained low.

Your customers search everywhere. Make sure your brand shows up.

The SEO toolkit you know, plus the AI visibility data you need.

Start Free Trial

Get started with

But. The models that earned your skepticism evolved

The most difficult part of AI is the speed. The AI you tried 18 months ago and the AI available to you today are not the same technology. 

Eighteen months ago (Fall 2023). The GPT-4 generation excelled at drafting, summarizing, and generating. But if you asked it to reason through a multi-step problem, hold context across a complex task, use external tools, or check its own work, it fell apart. It was a brilliant single-task performer who couldn’t manage a project.

Twelve months ago (Spring 2024). GPT-4o and Claude 3 Opus brought longer context windows and better reasoning. Claude 3 Opus, in particular, could handle document-length analysis that would have broken earlier models. But tool use was still experimental and unreliable. Agentic workflows (sequences of AI actions executing without hand-holding) existed mostly in demos and developer sandboxes. The gap between generation and editing was still wide.

Six months ago (Fall 2025). This is where the real shift happened. Reasoning models such as OpenAI’s o1 and Claude 3.7 introduced AI that thought before it answered. They were working through problems step by step, catching their own errors, revising their approach. Anthropic’s Model Context Protocol (MCP), launched in late 2024, gave models a standardized way to connect to external tools such as databases, calendars, CMSes, and email platforms, turning a chat interface into something closer to a software agent. The outputs that once required five rounds of correction started landing right in two.

Now (March 2026). Claude Sonnet 4.5 can autonomously sustain complex, multi-step tasks for over thirty hours. GPT-5.2 has reduced hallucination rates to under seven percent. Researchers at METR, tracking AI performance across five model generations, found that the length of tasks AI can complete independently has doubled every seven months. The models that failed you in 2023 were replaced by systems that can plan a campaign, pull competitive data, draft variants, score them against your brand guidelines, and flag the top option for your review, all while you’re in your morning stand-up.

I had my own “wow” moment recently. I’d been using AI for content drafts for over a year, always with the same low ceiling. On a whim, I asked a current-generation model to take a published blog post, research three competitive angles I hadn’t covered, draft a follow-up piece with a different argument, identify the three best distribution channels for that piece based on our audience data, and write tailored intro copy for each channel, all in one session, without me touching the keyboard again until it was done.

It worked. Not perfectly. But close enough that my edit time was 20 minutes, not two hours. The ceiling had moved. And I didn’t realize how much it had moved until I pushed against it.

Recommended Articles

Salesforce launches Agentforce Operations to automate back-office work

Apr 29, 2026

10 lessons from real-world ESP migrations

Apr 28, 2026

Vibe coding is hollowing out the martech stack fast

Apr 28, 2026

What you could actually build right now

Let me give you a concrete example.

Every quarter, marketing teams produce a competitive landscape update. Someone scrapes three competitor websites, reads their latest blogs, checks their social cadence, and writes a summary. It takes a day. With a current-generation AI model connected via MCP to your web tools and CRM data, that can be triggered by a calendar event, executed overnight, and waiting in your inbox, complete with a changes-since-last-quarter comparison and a flagged things-to-watch section. Your job becomes reviewing and deciding, not gathering and summarizing.

The best part? You don’t need to know how to build it. You can simply put the context into the LLM, tell it what you are trying to do, and have it suggest the best approach. It doesn’t always get it right the first time. But we’ve come a long way since November 2022.

What these new approaches require is a willingness to redesign workflows and to push past the chatbot ceiling.

The bottom line

I know the reflex. I’ve felt it myself. Wait until it’s more reliable. Wait until there’s a best practice. Wait until someone else proves it out. It takes too long to build an automation.

But METR‘s benchmarks show capability doubling every seven months. That means the time to start experimenting is now. 

Try an experiment this week. Pick one workflow on your team that involves at least three handoffs and takes more than a day from trigger to delivery. Map it out. Then ask how a sequence of agents would handle this end-to-end, with one human decision point at the end? Then ask your favorite AI tool how to make it happen. 

You might surprise yourself.

The chatbot era was a fine start. We just don’t have to stay there.