The New AI Risk for Marketers: Flawlessly Bad Decisions, Beautifully Integrated
It used to be pretty easy to spot AI outputs because it was characterized by robotic copy, repetition, and hallucinations. While these failures were once easier to spot and correct, today’s improvements to frontier AI have introduced different and more subtle risks. And, I’d venture to say, more dangerous ones.
Today’s AI agents aren’t just used for creating content. They’re also used for managing operations autonomously. They can send emails, update CRM records, adjust campaign budgets, and more, all on their own, with little to no human management. A new failure mode of enterprise AI has emerged. Let’s call it beautifully integrated bad decisions. This risk isn’t related to weak output, but to the confident execution of bad calls.
The Evolution of AI: From Advising to Taking Action
AI can be independent and autonomous. It can also be flawed. Adoption has kept pace with advancements as organizations have been eager to realize the benefits related to efficiency and cost reduction. Salesforce’s 10th State of Marketing report claims that 75 percent of organizations are using some form of AI. Among high-performing teams, the number is even higher.
But that adoption comes with some bumps in the road. Only one in four marketing executives say they’re satisfied with the level of data unification—the process of combining fragmented information from disparate sources (CRMs, ERPs, websites) into a single, cohesive view to create a “single source of truth.” That’s a problem because unified data is foundational for realizing the benefits from AI technology.
Clean Execution Can Hide Wrong Decisions
It wasn’t hard for humans to spot flaws and make corrections to flawed AI output when this technology first emerged. Today, though, problems can be masked, giving marketers a false sense of security.
Flawed interpretations drive actions that can’t be taken back. A prospect’s hesitation could be misinterpreted as buying intention. Old data points could be mistaken for current preferences. Confidence scores could trigger actions that execute efficiently and look legitimate but are not.
The automation isn’t broken. The problem is that a bad decision now moves seamlessly, at breakneck speed, through all connected systems.
Context failure is often more damaging than content failure:
- The wrong customer receives the right message at the wrong moment.
- A misclassified account gets an aggressive sales sequence.
- A budget shift is triggered based on an unverified signal.
Cybersecurity agencies recently warned that “every individual component in an agentic AI system widens the attack surface.” That warning transfers directly to interconnected MarTech stacks. Without the proper context, agents don’t just underperform. They act with misplaced confidence upon faulty premises, executing decisively on signals that don’t mean what the system thinks they do.
Human in the Loop?
As organizations recognized some of the issues and risks with AI technology, they moved to insert people in the process, referred to as “human in the loop.” The approach was sound but not sound enough to contain the risks related to how processes are currently being implemented.
Senior executives across cybersecurity have openly debated whether traditional “human in the loop” governance is operationally viable as AI velocity increases. Their emerging consensus: It’s not.
That model is shifting to “human on the loop,” where humans supervise systems, set thresholds, and intervene on exceptions instead of approving every action. While the distinction sounds subtle, the operational implications can be significant.
To leverage these benefits, marketing leaders need to move governance upstream to a point before agents execute. They should:
- Audit context first. Know exactly which data each agent sees, what’s missing from its view, and where signals are too weak to act on. Agents operating on incomplete CRM or CDP contexts aren’t just less effective; they’re a liability.
- Deliberately constrain authority. Let agents act where mistakes can be corrected.
- Design for correction. Build approval thresholds, observability layers, and rollback capability into every system of action.
The next era of marketing AI will be defined by how responsibly marketing AI acts. The highest-performing organizations will be those that pair autonomy with judgment discipline, building not just for speed but for accountability at every point of execution.
Because in the agentic era, the most dangerous mistake is not the one that looks broken. It’s the bad decision executed so smoothly that it appears to be successful.
Stu Sjouwerman is cofounder and CEO of ReadingMinds.ai, an AI-moderated interview platform for conducting sentiment analysis. He also is the founder and executive chairman of KnowBe4, a cybersecurity platform that addresses human risk management. Sjouwerman is the author of Agent Powered Growth: Deploy AI Agents that Build your Marketing Pipeline 24/7.