There’s a new academic paper doing the rounds from Stanford, Harvard, and others that neatly explains something many marketers are going to experience firsthand. AI agents, simply put, are systems designed to plan tasks, use tools, and take action on your behalf with minimal human input, for example, taking a campaign brief, generating assets, updating the CMS, and logging results in your CRM without needing step‑by‑step human direction.

In theory, that sounds transformative. In practice, AI agents can look like magic in a demo… and then quietly disintegrate once you drop them into real marketing work.

That’s because agents require perfect prompts, clean data, narrow tasks, and no interruptions.
But what they actually get is messy briefs, half-finished assets, missing fields, shifting priorities, and poor human feedback. And suddenly the agent seems brittle, confused, or unreliable. In real marketing, that brittleness shows up as missed deadlines, off-brand assets, reporting errors, or compliance risks that quietly land on a desk to fix.

AI agents can be broken down into three core parts.

This matters because most real-world failures don’t happen inside one component, but in the seams between them, where planning assumptions, tool behaviour, and memory drift fall out of sync.

1. Planning: how the agent breaks a goal into steps
   (e.g. “create campaign → draft copy → check brand voice → generate variants”)
2. Tool use: how it actually does the work
   (search, APIs, browsers, code execution, CRM updates, CMS edits, etc.)
3. Memory: how it stores and retrieves context
   (previous campaigns, brand rules, user feedback, exceptions)

Most demo agents work beautifully because all three are perfectly aligned and static, a sort of quiet demo theatre driven by vendor incentives to show perfection, not resilience. As a prospect, you watch the magic happen. The work of many is performed by a single AI. You imagine a perfect world of 24/7 productivity, improved efficiency, and fewer errors.

But the real world isn’t perfect. It’s noisy, inconsistent, and full of edge cases that no demo ever shows. And unless an agent can adapt as conditions change, that initial sense of inevitability quickly turns into friction, rework, and loss of trust.

Cost pressure, headcount constraints, and rising expectations for speed and scale from the board are pushing CMOs and marketing ops leaders to increasingly consider agents. The danger is that many of these decisions are being made based on a single happy-path workflow.

So, how do you determine whether an agent will survive outside the demo?

If you’re evaluating or piloting an AI agent, here’s a pre‑purchase checklist you can use before buying:

1. Break the brief

• Give it an incomplete or contradictory brief
• Change the goal halfway through
• Ask it to handle two priorities at once

What to look out for: If it fails, it will struggle to scale beyond a tightly controlled pilot.

2. Interrupt the workflow

• Pause execution
• Add late feedback
• Remove a tool or data source temporarily

What to look out for: Does it recover, or does it restart from scratch?

3. Test memory, not intelligence

• Can it recall why a past decision was made?
• Does it adapt based on prior corrections?
• Can it handle exceptions as well as rules?

What to look out for: If the agent can’t explain why it made a past decision or adjust based on prior corrections, it may appear smart but won’t be dependable at scale. This is a failure mode that catches most agents out.

Agentic AI isn’t hype, but it is fragile

API call failures, CMS schema changes, and missing or renamed fields are part of the day-to-day of marketing operations. Yes, we get stuff wrong, miss things, and can’t find the information we need. Will your real-world marketing operations cause the agent to turn brittle?

So if you’re a marketer, CMO, or ops lead exploring agents right now, the question isn’t:

“Does this agent look impressive?”

It’s:

“What happens when marketing reality interferes?”

I believe the winners will be the marketing teams who treat agents as systems, not magic employees.

Let’s continue the conversation.

Connect with the author, David Sloly, on LinkedIn.

Subscribe to our weekly AI Marketing Newsletter and stay up to date with all the AI marketing news.