The AI Agent Mandate No One Wrote
The demand gen agents performed, but no one decided where they should stop.
Each month, this series follows a fictional composite leader through a real professional challenge. The situations are composites drawn from patterns I observe across B2B marketing teams in AI transformation. I invented the names and companies. The failure modes, however, are real.
Nadia approved the AI expansion in February, during a quarterly business review where her demand gen results were the one bright spot in a flat pipeline quarter. Eight months earlier, she built an agentic AI system to handle prospecting sequences, content personalization, and pipeline qualification. The results held, and the team gained capacity. Her CEO asked where they could apply it to other use cases, and Nadia thought about customer renewal.
Her reasoning was defensible. The same personalization logic that interpreted buyer signals in demand gen, she reasoned, could do the same in retention. She moved forward, setting firm guardrails on volume and frequency. But she didn’t define what the agents should never be allowed to decide.
In week three of the new deployment, an agent generated a renewal sequence for three enterprise accounts in active contract renegotiations. The system generated the sequence on schedule. But the tone read like a script, completely misaligned with a delicate, multi-million dollar renewal. No one on Nadia’s team flagged those accounts as outside the agent’s mandate, because no such mandate existed. Two account executives received escalations from their contacts. One account went quiet.
Nadia was in a pipeline review when her VP of Sales texted her about the incident.
THE SITUATION
Eight months of results made the next decision feel inevitable
Nadia’s demand gen system didn’t fail once during the eight months it operated. That track record shaped the February decision more than the capability analysis. Consistent performance at one layer of the customer journey created a specific kind of confidence: that the agent understood the work as well as executed the task.
Nadia tightly governed the first deployment, with named KPIs, a defined scope, and human review at key points. The second deployment inherited the first one’s results, not its architecture.
The agents ran as designed. But the design was the problem.
WHY THIS MATTERS NOW
Agentic AI is scaling faster than the criteria to constrain it
Right now, B2B marketing leaders are operating under an intense executive directive to scale automation. According to BCG’s AI Radar 2026, roughly 90% of CEOs expect measurable ROI from AI agents this year, and companies dedicated nearly a third of their AI budgets to agentic deployments. CMOs aren’t expanding agent access in a vacuum; they are responding to massive top-down pressure. The conviction, however, is running ahead of the evidence.
McKinsey found that 62% of organizations are experimenting with AI agents, but only 39% report any impact on enterprise-level EBIT. A March 2026 BCG analysis found that 60% of companies have seen minimal or no business value from AI despite significant efforts, and nearly two-thirds report uncontrollable scaling expenses. Gartner projected in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Gartner’s analyst named the root cause directly: most projects are “mostly driven by hype” and “often misapplied.” The mandate gap is where the failure lives.
The technology performed; the agent generated the renewal sequence exactly on schedule. The failure was operational. Because executive pressure demands fast deployment, leaders are scaling AI without building the governance architecture that defines where it should stop. The agents ran as designed. Nadia never built the constraint layer.
THE GAP
What she governed, and what she forgot to define
Her initial demand gen deployment succeeded because it had three structural pillars that the customer retention expansion did not: a tightly defined scope, a clear human owner for every automated action, and a mandatory review checkpoint before any text reached a live account.
Nadia didn’t actively strip these guardrails away for the second rollout. She assumed the governance structure had transferred.
But continuity is a dangerous assumption when changing customer motions.
Unlike cold prospecting, enterprise accounts live in a high-stakes relationship layer. Here, slight changes in timing, tone, and context carry massive revenue consequences. A conversion-optimized AI script simply isn’t equipped to read those nuances.
The three accounts currently in contract renegotiations went completely unflagged because the system lacked any concept of relationship states. Tracking a sensitive, ongoing negotiation requires human context and judgment, but Nadia had never designated anyone to feed that critical insight to the AI.
She wasn’t alone in this governance vacuum.
According to a 2026 industry survey by SmarterX, only 13% of organizations have all four foundational governance elements in place for AI deployment; nearly a third have none.
Nadia’s customer retention expansion fell into the latter camp. The incident report documented what the agents sent. Nadia never defined what should have been off-limits.
WHERE WE LEAVE NADIA
The incident report closed. The governing question stayed open.
Her VP of Sales spent three weeks in direct conversations to repair the damage and bring back the account that had gone quiet, conversations that a proper mandate would have prevented. They resolved the escalations, but the structural gap remained. Nadia acknowledged the mistake in a leadership meeting and committed to reviewing her AI expansion criteria.
The underlying problem was systemic: no one asked her what those criteria were before she approved the deployment. She hadn’t asked herself.
The question Nadia struggled to answer in that meeting is the one this series picks up: which decisions should an AI agent never make? She can document what the agents sent. The mandate that would have prevented it was never written.
This series, in four parts:
Part 1 — The Leadership Brief: The mandate that was never written (this post)
Part 2 — The Framework: The Agent Scope Map (paid)
Part 3 — Real-World Examples: Where strategic restraint held — and where it didn’t
Part 4 — The Debrief: What Nadia decided, and the question she’s still carrying
Sources
BCG. “As AI Investments Surge, CEOs Take the Lead on Decision Making and Upskilling Themselves.” January 15, 2026. https://www.prnewswire.com/news-releases/as-ai-investments-surge-ceos-take-the-lead-on-decision-making-and-upskilling-themselves-302661849.html
BCG. “How Leaders Build an AI-First Cost Advantage.” March 27, 2026. https://www.bcg.com/publications/2026/how-leaders-build-an-ai-first-cost-advantage
McKinsey & Company. “The State of AI in 2025.” November 2025. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
Gartner. “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027.” June 25, 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
SmarterX. “2026 State of AI for Business.” April 2026. [Industry survey — secondary reference]
About Kim
Kim Celestre is a strategic advisor and executive coach who helps marketing leaders navigate AI transformation without eroding judgment, trust, or human value. Her work is grounded in AIGP-certified responsible AI expertise, executive coaching, and 25 years of Silicon Valley marketing leadership, including 4 years as a Forrester industry analyst.