Here is the number that should reframe how you think about AI governance. When Gartner predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027, one of the three causes it named was not technical at all: inadequate risk controls. Governance failure is not a compliance footnote to the AI story. It is one of the main ways the story ends badly.

And yet most marketing teams currently sit at one of two bad extremes. Either governance means a forty-page policy nobody has read, written by people who have never shipped a campaign, functioning mainly as a reason to say no. Or it means nothing at all: tools adopted individually, no shared rules, and everyone quietly hoping the confident output is correct.

There is a workable middle, and it fits on one page. This piece covers what actually goes wrong when marketing uses AI ungoverned, the evidence and the precedents, and a five-rule standard your team could adopt this week.

What Actually Goes Wrong: Four Risk Families

Strip away the hypotheticals and the real incidents cluster into four families. Each one has already produced public consequences.

1. Your AI's words are legally your words. The precedent every marketer should know is Moffatt v. Air Canada (2024 BCCRT 149), where a Canadian tribunal found the airline liable in negligent misrepresentation after its website chatbot invented a bereavement fare policy that did not exist. The company's defence, that the chatbot was effectively a separate entity responsible for its own statements, was dismissed as "a remarkable submission": the chatbot was simply part of the website, and it made no difference whether the misleading information came from a static page or a chatbot. That principle travels. Every chatbot, AI-drafted FAQ, automated reply and agent-generated claim your marketing publishes carries the brand's full liability, and "the AI said it" is not a defence anyone has successfully run.

2. AI fabricates, fluently, and reassures you while doing it. The most famous example came from law rather than marketing: in Mata v. Avianca (2023), a US federal court sanctioned two lawyers and their firm, with a $5,000 penalty and letters of correction to every judge falsely named, after they filed a brief citing six judicial opinions their AI tool had simply invented. The detail marketers should sit with: when the lawyer grew suspicious and asked the tool directly whether the cases were real, it assured him they were. The court's framing is the governance lesson in one line: there is nothing inherently improper about using an AI tool for assistance, but the gatekeeping duty to verify accuracy stays with the human. The marketing equivalents are quieter but constant: statistics that do not exist, capabilities your product does not have, claims your evidence cannot back. The research explains why this keeps catching capable people: in the Harvard-led field experiment with BCG consultants, professionals using AI on a task just beyond its competence were 19% less likely to reach correct answers than colleagues without it, because fluent, confident output is precisely what disarms scrutiny. Fabrication risk is not a model flaw you can wait out. It is a workflow property you have to design for.

3. What goes in can get out. Several large enterprises have restricted or banned general-purpose AI tools after employees pasted confidential material, source code, strategy documents, unreleased plans, into consumer chatbots; Reuters documented the pattern, and the corporate bans that followed, as early as 2023. For marketing specifically, the sensitive inputs are customer data, unannounced launches, pricing strategy, and anything under NDA with partners or agencies. The risk is twofold: exposure through the tool itself, and the simple fact that once confidential material leaves your controlled environment, you no longer decide where it lives.

4. Brand integrity erodes at machine speed. The quieter, compounding risk: off-brand output published at a volume no reviewer sees. A voice that drifts generic because every competitor uses the same models. Claims slightly too strong for your evidence, a hundred times a month. Third-party IP reproduced without anyone checking. None of these makes headlines like a tribunal ruling, but the liability and sanctions cases both started the same way: output nobody with accountability actually read.

Four families, one pattern: every incident was preventable with rules a marketing team could write in an afternoon. Which is exactly the point of what follows.

The One-Page Standard: Five Rules

Governance sized for a marketing team is not a policy document. It is five rules, written down, agreed once, and revisited quarterly. Here they are, with the reasoning each one carries.

Rule 1: Map the zones. Divide your AI use into green, amber and red, along two axes: how sensitive the input is, and how exposed the output is. Green: internal, low-sensitivity work (brainstorming, summarising public material, first drafts of internal documents), use freely. Amber: anything customer-facing or data-touched, use with the named-human sign-off below. Red: never (see Rule 4). The map matters more than the categories: teams with a clear green zone adopt faster, because nobody wastes judgement on permissions the map already grants. Good governance is an accelerant, not a brake.

Rule 2: A named human signs off anything public. Anything that leaves the building, ads, posts, emails, chatbot behaviours, sales collateral, claims of any kind, gets approved by a person whose name attaches to it. Not "reviewed by the team": named. This is the direct lesson of the liability cases, and it is the gatekeeping duty the Avianca court described made operational. It also encodes a principle from our handover rules: the machine produces, the human decides, and anything that would need your name on it in front of a stakeholder is a decision.

Rule 3: Claims need sources, whoever drafted them. Every statistic, capability claim and comparison in AI-drafted content gets traced to a source before publication, exactly as if a new hire had drafted it. This is the anti-fabrication rule, and it is cheap: it usually takes minutes, and it catches the invented statistic before it becomes the invented tribunal exhibit. If a claim cannot be sourced, it does not ship. And remember the Avianca detail: asking the tool itself whether its output is accurate is not verification, because the tool will reassure you.

Rule 4: Nothing sensitive goes in. One list, visible to everyone: customer personal data, unreleased plans, pricing strategy, partner-confidential material, credentials. Nothing on the list enters an external AI tool, in any form, for any deadline. Where your organisation provides enterprise AI environments with data protections, route sensitive work there; the rule is about uncontrolled tools, not about avoiding AI.

Rule 5: Log it, and know who to call. Keep a simple register of which AI tools the team uses and for what, and agree the escalation path before you need it: who gets told when an output was wrong in public, and who can pause an automated system. For anything agentic, this connects to the test we set out in our evidence review of AI agents: if you cannot see and correct what it did, it does not get autonomy. Visibility is the difference between an incident and a crisis.

That is the whole standard. One page, five rules, and a quarterly half-hour to revisit the zone map as tools and use cases change. It will not satisfy a regulator in a high-compliance industry on its own, and it is not meant to; it is the marketing team's working layer, designed to be followed rather than filed.

The Judgement: Sized to Risk, Owned by Marketing

Two closing calls, because governance is ultimately a judgement exercise, not a template.

First, size it to your actual risk, and keep sizing it. A ten-person team drafting blog content needs the five rules and nothing more. A team running customer-facing chatbots and agentic outreach needs the same five rules plus real monitoring, because their amber zone is bigger and their mistakes propagate faster. The variable is not company size. It is the cost of a wrong call reaching the public, which is the same calibration logic that governs how much autonomy an agent gets in the first place.

Second, own it; do not outsource it to legal. Legal review matters at the edges, but a governance standard written without marketing judgement becomes the forty-page blocker, and one written without legal awareness becomes the tribunal exhibit. The five rules work because they encode marketing judgement (what is actually risky in our work) in a form other functions can trust. Teams that write their own standard also follow it, which is the only metric of governance that matters.

The throughline, as ever: the tools will keep changing, the incidents will keep making case law, and the durable capability is the judgement to decide what needs a rule, what needs a human, and what can run free. Evidence informs. Judgement decides.

Ready to Put the Fundamentals Under Your AI?

Governance is what protecting the brand looks like in the AI era, and protecting the brand is a fundamental, not an afterthought. That sequencing, fundamentals first, then AI as a deliberate layer, is how the FP Collectiv curriculum is built: the B2B Marketing track for the fundamentals, and the Marketing and AI series, starting with Marketing and AI Foundations, for using AI with exactly this kind of judgement.

Sources

Gartner press release (June 2025): more than 40% of agentic AI projects will be cancelled by end of 2027, citing escalating costs, unclear business value and inadequate risk controls.

Moffatt v. Air Canada, 2024 BCCRT 149 (British Columbia Civil Resolution Tribunal, February 2024): airline liable in negligent misrepresentation for its website chatbot's invented bereavement fare policy; the separate-entity argument rejected. See also the McCarthy Tétrault analysis.

Mata v. Avianca, Inc., Opinion and Order on Sanctions (S.D.N.Y., 22 June 2023): lawyers and firm sanctioned, with a $5,000 penalty and corrective letters ordered, for filing six AI-fabricated judicial opinions; the tool assured counsel the fake cases were real when asked.

Reuters (August 2023): reporting on workplace adoption of consumer AI tools and the corporate restrictions that followed confidential-data concerns.

Dell'Acqua, F., et al. (Harvard Business School Working Paper 24-013, 2023), "Navigating the Jagged Technological Frontier": consultants using AI on a task beyond its competence were 19% less likely to produce correct solutions.