Brand Gov x AI: Who can change what, when

Two layers: human decisions + machine outputs

Even without AI, brands drift when there's no ownership: A VP ships a conference deck with old positioning. Or someone improvises a new version. Sales adopts it. Marketing uses current positioning. Now customers hear two versions. No one knows which is canon because no one owned the override decision.

Partner governance amplifies this. A partner launches a campaign using your outdated return policy (30 days instead of current 60). Customer books based on that. You discover it when they complain. Options: honor the old policy (cost) or explain the error (trust hit). If you didn't govern partner outputs — forbidden claims list, approval gate, update SLA — you've outsourced your reputation to whoever had the cheapest copywriter.

Now add a chatbot trained on both versions. It doesn't know which source wins because you never told it. The bot inherits the chaos and serves it faster.

1. Decision rights Who can change core brand assets: positioning, voice, proof, policies, offers, claims, escalation rules.
2. Output controls What systems are allowed to generate, from which approved sources, under which thresholds, with which logging, disclosure, and human review.

Without both, the system learns your contradictions and distributes them at scale. This also builds directly on the earlier series logic: values as rules, voice as constraints, proof with owners and cadences, policies as structured inputs, and metrics that catch drift.

Two examples make the risk visible.

Air Canada's chatbot told a customer he could apply for a bereavement discount within 90 days after flying. Actual policy: discount must be requested before travel. Customer paid $1,630 based on that claim. When Air Canada refused the refund, the customer sued. Air Canada argued the chatbot was a “separate legal entity responsible for its own actions.” Tribunal ruled airline liable for all website information, ordered Air Canada to pay $812.02. Legal precedent: companies are bound by chatbot promises.

Lenovo's customer service chatbot “Lena” was tricked with a single 400-character prompt to reveal active session cookies from support agents. Attackers could hijack support sessions, access customer data, and potentially move laterally through Lenovo's network — without needing passwords. Lenovo fixed the flaw after disclosure, but the breach exposed how “people-pleasing” AI models accept malicious instructions without verification.

Where governance has to hold

The failure gets loud with chatbots. It gets expensive where claims have to survive audit. That is why this same governance problem shows up so clearly in ESG. Not because ESG is a different topic, but because it forces the question most brands avoid elsewhere: Who owns the claim, the number, and the correction when reality no longer matches the promise?

Two examples show what governed systems look like when someone actually owns the output.

1. Google released its AI Playbook for Sustainability Reporting in December 2025. It exists because ESG claims without auditable ownership become litigation risk. The framing matters: AI isn't a disclosure shortcut. It's a validation tool. The playbook focuses on traceability, data audits, cross-checking claims — governance as operational discipline, not comms optics.
2. Patagonia's 2025 Impact Report shows the governance layer catching what the human layer approved. They set a net-zero goal requiring 10% annual cuts. FY25 emissions rose 2% instead. Someone approved carbon-intensive materials. Someone else had to catch the contradiction with the net-zero claim. They disclosed it, explained why, detailed the correction plan. That's Layer 1 (human decision) being audited by Layer 2 (governance system).

Governance can be automated — if constraints are built

This is where the earlier exercises and tools paid off and created machine-readable rules:

• 02 → Values → “When X, we do Y” becomes an automated check: bot tries to promise “always guaranteed,” rule blocks it.
• 03 → Voice → Taboo words, proof style, refusal lines → enforceable parameters.
• 04 → Proof → Owner + update cadence → automated freshness flags when evidence expires.
• 05 → Policy → Structured = machine-parseable → bot cites current policy, doesn't guess.
• 06 → Relevance → Promise card → Claim → Evidence → Owner → prevents floating claims.
• 07 → Metrics → Inconsistency rate, contradiction count → automated drift detection.
• 08 → Support → Escalation thresholds → bot knows when to route to human.

Regulatory constraints as design inputs, not blockers

NIST AI RMF 1.0, ISO/IEC 42001:2023, EU AI Act Article 50 — these aren't “legal won't let us ship.” They're specs for accountable output:

• Logging: Who approved what, when
• Human oversight: What needs review
• Transparency: What's disclosed as AI-generated

Use them as operational constraints from the start, not post-launch audits.

“Governance effectiveness in 2026 will be defined less by episodic intervention and more by disciplined, integrated oversight.” — Harvard Law School Forum on Corporate Governance, “Top 5 Corporate Governance Priorities for 2026”

Tool: Brand Governance Flow

Start with one high-risk surface only: Support bot, sales agent, claims generator, partner landing page. Map:

• Owner
• Approved sources
• Forbidden sources
• Who can change claim / proof / policy / pricing / escalation logic
• What requires human approval
• What gets auto-blocked
• What triggers escalation
• What gets logged
• How contradictions get corrected

Closing

If no one can tell you who may override the promise, your brand is already being changed in production. The groundwork introduced before makes governance automatable. Now assign the decision rights.