Agentic Marketing: The Fine Line Between Agency and Accountability
In Part 1, we covered how decision systems work; in Part 2, how the Decision Policy Layer collapses; and in Part 3, the hidden cost of optimization: Exploration Debt.
Now it’s the right time to add a fourth dimension. The system no longer just optimizes. It makes decisions, allocates budgets, generates messages, and launches campaigns, and it does all this without waiting for human intervention. The agents have taken the stage!
When the Autopilot Is Switched On
In 1988, the Airbus A320 introduced fully automated flight systems to commercial aviation. Pilots were no longer flying the aircraft; they were supervising the system. Most of the accidents in the early years stemmed not from a malfunction in the flight system but from the blurred line between agency and accountability. Pilots couldn’t fully grasp what the autopilot would do, and when trying to override one of the system’s decisions, they could end up intervening incorrectly.
Today, marketing AI systems have reached a similar threshold. The difference: the aviation industry learned this transition gradually over decades. The marketing world doesn’t have that much time.
Agent-based AI systems (agentic AI) manage the campaign cycle autonomously from start to finish: audience selection, content generation, channel decisions, budget allocation, and performance optimization all happen in real time, without waiting for human approval.
This is an efficiency story. But it is also an accountability crisis.
In June 2025, Gartner shared the findings of a study it conducted. The results indicate that more than 40% of agentic AI projects will be canceled by the end of 2027. The reasons: escalating costs, unclear business value, and inadequate risk controls. In the same study, Gartner also defined a concept: “agent washing.” Among the thousands of vendors rebranding existing chatbots and automation tools as agentic AI, only about 130 offer genuine agency.
This picture sets the ground you need to understand before reading further: the technology has matured, but the governance framework has not yet.
Anatomy of the New Paradigm
We moved from manual decisions to rule-based automation, and from there to predictive AI. Now it’s the turn of fourth-generation agentic AI. The system no longer offers recommendations: it takes action. The difference isn’t in the word but in where accountability rests.
Agentic AI selects the audience, generates the content, allocates the budget, launches the campaign, and instantly readjusts based on the results. The entire cycle is autonomous. And the question that best sums up this transition remains unanswered: “When the autopilot is switched on, where does the pilot sit?”
Three Cases, One Tension
Salesforce Learning on Itself: The Wrong Flight Mode
When Salesforce launched Agentforce in late 2024, it also declared itself the first customer (Customer Zero), deploying the system at full scale across its own internal operations. But it hit a limit first, and only then reached success.
When the SDR (sales development representative) agent was first deployed, an unexpected problem emerged: the agent was giving technically correct answers, but the experience felt cold and mechanical. Human reps would open their responses to customer problems with “I’m really sorry to hear that.” The agent, by contrast, went straight to delivering the necessary information without any empathy. The autopilot was following the right course, but it was in the wrong flight mode.
This wasn’t a single-line-of-code problem. It was a critical observation that defined a structural limit of AI agents: the system could scale efficiency, but it couldn’t yet scale authenticity and context sensitivity. Once this gap was recognized, the system was reconfigured.
The results announced at the end of the first year were striking: of 1.5 million requests on the support platform, the vast majority were resolved automatically; the SDR agent worked on 43,000 leads and generated $1.7 million in new pipeline from dormant records; and the Slack integration saved employees more than 500,000 hours. But this success was only possible because the limit had been recognized.
Source: Salesforce, “What Salesforce Learned from Its First Year Using Agentforce,” September 2025 — salesforce.com/news/stories/first-year-agentforce-customer-zero
Air Canada’s Legal Limit: Who Was at the Controls?
February 2024. After his grandmother passed away, Jake Moffatt went to Air Canada’s website and asked the chatbot about bereavement fares. The bot told him that the discounted fare could be applied retroactively within 90 days of purchasing the ticket. Moffatt bought a full-price ticket. He later learned that no such discount policy existed.
In court, Air Canada’s lawyers put forward a surprising defense: they argued that the chatbot was a separate legal entity, independent of the company, and that the company therefore bore no responsibility for the information it provided. The British Columbia Civil Resolution Tribunal rejected this defense, calling it a “remarkable submission.” The ruling was clear: a company cannot be shielded from liability for information provided by its own agent.
The tribunal’s position speaks directly to the autopilot metaphor: “You handed over the controls. Accountability cannot be handed over.” Legally, this case was about a chatbot. But over the next five years, agentic AI systems will allocate budgets, generate price quotes, and send mass messages. Each time, the same question will apply: the system made the decision, but who is accountable?
Source: Moffatt v. Air Canada, 2024 BCCRT 149 — canlii.ca/t/k2spq
McKinsey’s Warning: The Alarm Is Sounding, but There’s No Protocol
Research published by McKinsey in October 2025 documented that 80% of organizations deploying agentic AI systems had encountered risky behaviors originating from the system. In most of the aircraft, the alarm had gone off, but the intervention protocol wasn’t clear.
McKinsey Partner Rich Isenberg described the situation this way: “Agency isn’t a feature; it’s a transfer of decision rights. The question is no longer ‘Is the model accurate?’ but ‘Who is accountable when the system acts?’”
The same research also identified the scenario in which risk management fails: leaving guardrails optional and allowing anyone to bypass them. This isn’t a technical failure; it’s a governance choice.
Sources: McKinsey, “Trust in the Age of Agents” — mckinsey.com, March 2026; McKinsey, “Deploying agentic AI with safety and security: A playbook for technology leaders” — mckinsey.com, October 2025
The conclusion is simple but profound: the more autonomous agentic AI becomes, the more deliberate its governance must be. Because as autonomy grows, errors grow too, along with their speed and scale.
The Agentic Marketing Governance Framework
Agentic AI systems can’t be governed with the traditional AI oversight model, because these systems no longer offer recommendations. They take action. This difference calls for a four-component governance framework:
1. Autonomy Map
Classify every agentic AI decision along two dimensions: magnitude of impact and reversibility. Low-impact, reversible decisions can be opened up to full autonomy. High-impact or irreversible decisions should require human approval. Design autonomy not as a binary choice but as a spectrum.
Of the decisions your agentic AI system can make today without human approval, if one goes wrong, within how many hours at the latest would you notice?
2. Decision Audit Trail
For every decision the agent makes, keep a traceable and explainable record: why it chose this audience, why it allocated this budget, which signal it interpreted and how. As the Air Canada case showed, when the system makes a decision, you are the one held accountable, and to answer for it, you need the decision trail.
When your agentic AI system made a campaign decision last month, could you explain it, with its rationale, in five minutes?
3. Circuit Breakers
Design thresholds that can override the agent’s autonomy:
- Stop the system when budget deviation exceeds a certain percentage.
- Raise an alert when the segment exclusion rate rises above normal.
- Require approval before making first contact with a new audience.
These mechanisms exist not to slow things down but to build trust.
Does your agentic AI system have a “circuit breaker” that has been triggered? If so, how many times has it kicked in over the last six months?
4. Accountability Protocol
“The system made the decision” is no longer accepted as a legal or ethical defense. Every agentic AI decision should have a human owner assigned in advance. This person should own not the system, but the outcomes the system produces.
If your agentic AI system makes a wrong decision tonight, who gets the first phone call in the morning?
“You can delegate the task. You cannot delegate the accountability.”
Agentic AI systems were designed to grant power, not to offload accountability. Salesforce learned from its own experience and updated its protocol. Air Canada didn’t learn; the court taught it. The McKinsey data tells us which of the two is the norm.
You Have a Message!
To Marketing Leaders: Which decisions is your agentic AI system making today without human approval? Have you systematically mapped the impact and reversibility of those decisions? If you don’t have that map, you don’t have an autonomy map either.
To Senior Executives: When your agentic AI system makes a wrong decision, whether it sends the wrong message to the wrong audience, wrongly excludes a segment, or misdirects the budget, who owns that decision? The answer to this question must be clear before the system goes live.
Next Installment: The AI-Native Marketing Organization
The fifth and final part brings together the five dimensions of this series: decision systems, the policy layer, Exploration Debt, agents, and the organizational architecture that will allow these four elements to work together.
An AI-Native Marketing Organization isn’t just a team that uses AI tools; it’s a structure that has designed its decision architecture, governance framework, and human–AI division of labor at the system level. This final part will serve as the strategic summary.
Two small notes: I rendered the tribunal’s wording as “remarkable submission,” since that’s the phrase used in the actual Moffatt ruling. And the line “You handed over the controls. Accountability cannot be handed over” is presented in the original as the tribunal’s formulation, but it reads as your paraphrase rather than a quote from the decision. You may want to frame it as “In effect, the tribunal’s message was…” so it isn’t mistaken for a direct quotation.
Comment(s)