Macro vs Help Center Policy Audit: Find Conflicts Before AI Launch

Start by mapping every high-volume customer intent to its sources across three layers. The public help center is what customers see and what the AI will likely retrieve first. The macro library is what agents paste into tickets, often with faster edits and less review. The SOP or policy document is the internal rulebook that may contain exceptions, approvals, and workflows not written for customers.

Each intent should have one canonical source. If it has two or more that disagree, the intent is not ready for AI automation.

Not every contradiction matters equally. A mismatch in branding tone is low risk. A mismatch in refund window, eligibility rule, cancellation policy, data retention, or account recovery is high risk. Focus the audit on topics where a wrong answer creates customer harm, legal exposure, or financial loss.

Common high-risk contradictions include: the macro says full refund while the help center says case-by-case; the SOP allows a manager override while the public policy does not mention it; the macro promises a credit while the billing system cannot issue one.

The conflict table is the core audit artifact. Each row names the customer intent, the conflicting sources, the policy owner, the customer-safe answer, the resolution action, and the retest trigger. This turns a vague worry into a named, owned, timed fix.

Without a conflict table, teams often discover contradictions only after a customer complains or after the AI gives a wrong answer at scale.

When sources conflict, someone must choose the canonical answer. This is a business decision, not a model decision. The policy owner, legal reviewer, or support lead should choose the customer-safe version and retire or rewrite the conflicting source.

Do not let the AI reconcile contradictions. The AI may average conflicting sources into a middle ground that satisfies no one and violates policy.

The audit should end with a launch decision for every conflicting intent. Approved means the conflict is resolved and the canonical source is current. Restricted means the intent is answerable only with additional context. Blocked means the conflict is unresolved and the intent should stay out of AI scope. Source-fix-needed means a fix is in progress. Human-only means the intent requires judgement even when sources agree.

This launch map becomes the boundary for vendor configuration, test sets, and QA sampling.

Every blocked or source-fix-needed intent should become a work item. The backlog should include the exact source to rewrite, the owner, the deadline, the customer-safe wording, and the retest prompt. Sort by launch impact rather than document count.

A single resolved refund contradiction can unlock a high-volume intent. Twenty fixed typos do not unlock any intent.