Marketing was the first department AI conquered, if you count by press releases. Every tool in the stack added a generate button, every vendor deck grew a copilot slide, and the martech landscape graphic, which passed 15,000 logos, kept growing anyway.
And yet when we sit down with marketing teams, the complaint is never "we can't produce enough content." It's the opposite. Generation is cheap now. What's expensive is everything around it: getting the campaign from brief to live across nine channels, knowing which audience should see it, reviewing what the tools produced, and proving any of it worked. The AI that earns a permanent place in the stack attacks that connective tissue, not the content itself.
The use cases that are actually working
Four patterns keep delivering across the marketing and MarTech companies we work with, and they share a shape: each one removes a handoff between people, not a click between screens.
- Campaign assembly, not copy generation. The valuable unit isn't a paragraph, it's a launched campaign: one brief becoming the email sequence, the landing page, the paid variants, and the social cuts, each within character limits, each in the brand voice, each traceable back to the brief. Teams that build this as a pipeline, with brand guidelines and banned claims enforced as validation steps rather than hopeful prompt instructions, ship campaigns in days instead of weeks. The human role moves from producing nine assets to approving one coherent package.
- Audiences you can query in plain language. Most segmentation is limited by who can write the query, which means the CDP's potential sits behind one overloaded analyst. A natural-language layer over customer data, "customers who bought twice in the last 90 days but haven't opened an email in 30," compiled to real queries against real schemas, puts segmentation in the hands of the people running campaigns. This is a well-bounded LLM problem: schema-aware generation with a validation step, not open-ended chat.
- Creative testing at machine scale. Instead of three hero variants argued over in a meeting, generate thirty within brand constraints, let the traffic decide, and feed the results back. The interesting engineering isn't the generation, it's the loop: performance data flowing back into what gets generated next. Teams running this stop debating taste and start accumulating evidence about what their audience responds to. Lumoo runs the generation half of that loop at scale, turning a fashion catalog into on-model imagery and campaign visuals hundreds of items at a time.
- Lead qualification that reads instead of counts. Traditional scoring counts events: pages visited, forms filled. A model can read the free-text answer, the company's website, and the email thread, and produce a routing decision with a written rationale. Sales teams trust it more than an opaque score because they can audit the reasoning, and marketing stops paying for SDR hours spent on leads a paragraph of context would have disqualified.
What doesn't work: the demo-feature graveyard
The failure cases are as consistent as the successes, and most of them were dashboards or chatbots.
The chat-with-your-analytics feature gets used for a week and abandoned, because marketers don't want a conversation with the data, they want the anomaly surfaced before they knew to ask. The generate button inside a tool that still requires six other tools to launch anything saves minutes in a process measured in weeks. And unreviewed AI content pushed straight to publication is how brands end up apologizing on social media; the cost of one hallucinated product claim exceeds a year of saved copywriting hours.
The pattern behind all three: they added capability without removing a handoff. The bottleneck in marketing operations was never typing speed.
If you're a MarTech vendor, the bar just moved
For companies building marketing software, the strategic picture is uncomfortable and clarifying at once. An AI feature, summarize this, draft that, is now table stakes; every competitor shipped one the same quarter you did. Worse, foundation models plus a spreadsheet can now replicate a shocking amount of thin SaaS functionality. If your product's core value is a workflow wrapper around a database, an agent can wrap that database too.
What defends a slot in the stack is depth the model can't route around: proprietary data (deliverability reputation, ad performance benchmarks, intent signals), genuine network effects, or ownership of an execution channel. And the product form that wins is shifting from tool to worker. A tool exposes screens a human operates; an agent completes the workflow and shows its work. "AI-powered email editor" is a feature. "Give it the brief Monday, review the launched campaign Wednesday" is a product with pricing power, because it's priced against headcount instead of against other software.
That shift is an engineering commitment more than a modeling one. An agent that operates across email, ads, CMS, and CRM needs reliable tool use, state management across a multi-step workflow, cost control, and above all evaluation: campaign-level correctness checks, brand-compliance validation, and regression suites that run on every prompt and model change. This is exactly the production discipline we've written about before, and marketing's tolerance for public mistakes is lower than most, because the mistakes ship to your entire audience at once.
Where to focus
For marketing teams: pick the workflow with the most handoffs, not the task with the most typing. Count the days from brief to live campaign and attack the gaps between people. Keep a human approval on anything customer-facing, and instrument the edits, they're your quality metric and your eval set.
For MarTech builders: assume the generate button is commoditized. Ask what your product knows that a foundation model can't, and what workflow it could complete rather than assist. The 15,000-logo landscape is going to consolidate around the tools that answer those two questions convincingly. The generate buttons are going to be remembered as a transitional artifact, like the mobile app every company built in 2011 because phones existed.
