The biggest AI risk in digital advertising is probably not the one most people are talking about.
It’s not creative generation.
It’s not job replacement.
It’s not whether AI can optimize campaigns faster than humans.
The real risk is what happens when autonomous systems begin making commercial decisions without clear ownership, governance or accountability. That concern has been on my mind for a long time, partly because of a class I took years ago at Stanford with Ronald Howard, the father of decision analysis.
Howard had an unusual way of teaching risk. In one exercise, instead of answering multiple-choice questions normally, students had to assign probabilities to each possible answer. If you were completely certain and correct, you scored full points.
But if you were completely certain and wrong, the penalty was effectively catastrophic. You’d get a “negative infinity” score on that question.
In other words, one bad decision could mathematically destroy your score for the entire class.
The point wasn’t to grade mechanics. It was about teaching the implications of asymmetric risk. Small mistakes are survivable, but overconfidence in the wrong system can be fatal.
That framework increasingly feels relevant to where digital advertising is heading with AI.
Because while AI will almost certainly reduce many small operational problems in our industry, it may simultaneously increase the probability of systemic failures that are far larger and harder to contain.
In the Ron Howard class, one bad decision could kill your entire quarter. In an AI future for digital advertising, one agent without the right controls could bankrupt a company.
AI will reduce operational friction and potentially increase existential risk
Done properly, AI should eliminate a significant amount of day-to-day operational inefficiency. Fewer trafficking mistakes, reporting mismatches, broken workflows and campaign setup errors.
Those improvements are real and valuable.
But many companies are treating operational upside as if it automatically means lower overall business risk. That assumption is dangerous.
What AI changes is not simply the volume of decisions being made. It changes the scale, speed and autonomy of those decisions.
As we move into an AI-driven operating environment, systems may coordinate without explicit instruction, optimize without understanding context and act without a clearly accountable owner. And when those systems operate at machine speed, the consequences of failure can scale far faster than our current governance structures were designed to handle.
Other industries are already seeing early warning signs
Some of the clearest lessons are coming from outside advertising.
Air Canada faced legal consequences after a chatbot provided inaccurate bereavement fare information that conflicted with the airline’s official policy. The airline argued that the correct information existed elsewhere on the website, but the court determined that customers should be reasonably able to trust a chatbot’s response. In effect, AI-generated output created contractual liability.
Zillow’s highly publicized iBuying initiative offers another example. The company deployed algorithmic home-buying models at scale, but incorrect assumptions inside the model amplified losses quickly enough that the division ultimately had to be shut down.
Then there was Microsoft’s Tay, the chatbot that learned from user interactions and rapidly devolved into generating offensive content within hours of launch.
Different industries, different systems – but the same underlying lesson. When autonomous systems fail, they can fail at scale.
Advertising is particularly vulnerable
Digital advertising already struggles with governance and operational clarity in normal conditions. Many companies still lack mature frameworks around operational accountability and decision ownership.
AI raises the stakes dramatically.
One reason this matters is because our industry increasingly operates through interconnected systems making rapid decisions across pricing, targeting, optimization and creative deployment.
Now imagine those systems becoming increasingly autonomous.
A few examples quickly become concerning:
- Illegal or offensive creative at scale
- Privacy and regulatory violations
- Autonomous contracting risk
- Targeting protected audiences
The pricing collusion scenario nobody wants to discuss
But one area that particularly concerns me is pricing behavior.
Historically, allegations of pricing collusion typically involved clear human coordination in the form of meetings, communications, agreements or observable intent. AI complicates that framework.
What happens, for example, when independent optimization systems learn – without explicit human coordination – that aggressive price competition is economically irrational?
You could plausibly end up with algorithmic behavior that appears coordinated even if no human explicitly instructed it to happen. There are at least two versions of this risk.
Implicit coordination: Independent models learn to avoid competing aggressively on price because maintaining pricing discipline maximizes long-term revenue outcomes. No direct communication occurs, but the market behavior begins to resemble coordinated pricing.
**Explicit agent coordination: **Far more concerning is the possibility that autonomous agents directly exchange signals or optimize collaboratively without sufficient oversight.
At that point, regulators may not care whether humans explicitly instructed the behavior. They may simply conclude that companies allowed anticompetitive systems to operate unchecked, which creates enormous legal and regulatory exposure.
Importantly, many current governance structures are not designed for this new world.
The industry needs a control layer
Most discussions about AI in advertising focus on capability, but very few focus on control architecture.
That is the gap.
If autonomous systems are increasingly making economically meaningful decisions, organizations need formal control layers sitting above those systems. In practice, I believe this requires three core components. 1. Configuration authority
Every AI-driven process needs a clearly defined owner.
Not “the algorithm.” Not “the platform.” Not “the model.” But, rather, a person or accountable function. Someone must own bidding logic, optimization rules, targeting parameters, pricing boundaries and system oversight.
Organizations cannot outsource accountability to software.
2. Acceptable risk thresholds
Companies should not deploy AI faster than they can contain it. That means defining failure boundaries before deployment.
What triggers intervention? What thresholds require human review? What conditions force rollback or shutdown?
Too many companies currently approach AI implementation reactively: deploy first, investigate later. That is backwards. The thresholds need to be defined before the AI is rolled out.
The organizations that succeed will define operational guardrails before systems scale into production environments.
3. Reversibility and auditability
If systems fail, organizations need to understand what happened, why it happened and how to stop it quickly. That requires audit logs, process visibility, rollback procedures and operational traceability. In traditional enterprise software environments, rollback mechanisms are relatively mature. With AI systems – especially adaptive or continuously learning systems – rollback may be significantly harder.
Which makes planning for reversibility even more important.
Human operators become more Important, not less
One of the biggest misconceptions in the current AI conversation is that oversight itself can be automated away. I think the opposite may happen in the medium term.
Experienced operators who understand failure modes may become more valuable, not less valuable. Because somebody still needs to recognize when a system is confidently wrong.
But right now, many organizations are attempting to fund AI initiatives by reducing the very oversight layers needed to supervise them safely. That is a dangerous trade.
You cannot cut your way to effective AI governance through a depleted talent pool.
The people who understand pricing logic, operational edge cases, campaign workflows and market behavior are precisely the people organizations may need most as AI systems become more autonomous.
The right approach is not to ‘move slow’
None of this means companies should avoid AI – the opportunity is far too significant for that. But there is a major difference between experimentation and production deployment.
I increasingly think organizations need two separate operational tracks.
Frontier track: This is where experimentation happens: bounded pilots, human oversight and controlled operational exposure.
The objective is learning.
Fast-follow production track: Only after systems demonstrate reliability inside of controlled environments should they move into scaled production workflows. This is where mature governance, auditability and operational controls become critical.
The mistake would be deploying frontier experimentation directly into core production systems without sufficient control architecture.
The companies that win will maintain control at machine speed
The winners in the AI era will not necessarily be the companies that deploy AI the fastest. They will be the companies that can deploy AI at machine speed while still maintaining human governance, operational visibility and accountable control structures.
That is a much harder problem. But it’s probably the more important one.
Because the real challenge of AI in digital advertising is not simply building smarter systems. It’s ensuring that those systems remain governable after they become powerful enough to matter.
“Data-Driven Thinking” is written by members of the media community and contains fresh ideas on the digital revolution in media.
Follow James Deaker and AdExchanger on LinkedIn.