How agencies brief AI agents on do-not-say lists per client without killing personalization

AI outbound fails in public when it invents a partnership, names a competitor the client hates, or promises a metric nobody approved. The fix is not "turn personalization off." The fix is a per-client do-not-say brief that blocks unsafe claims while leaving room for evidence-based personalization.

Agencies already keep voice guides. A do-not-say list is the sharper sibling: explicit bans, scoped claims, and a pre-send check that fails closed on banned phrases. GTM writes from prospect profile and signals in the client voice, with a compliance gate on send. The brief below is how ops teams feed that voice without turning every email into a bland template.

What belongs on a do-not-say list

Keep it short enough that writers and models can follow it. Typical agency buckets:

• Legal / regulated claims the client is not cleared to make
• Competitor names (ban, allow-list, or "never disparage" rules)
• Customer names and logos not cleared for reference
• Pricing, discounts, or SLA numbers not on the approved sheet
• Fake personalization (invented job changes, funding rounds, blog posts)
• Internal codenames and unreleased product names
• Guarantees ("we will book you X meetings") unless sales approved the exact wording

If a line cannot be traced to an approved source, it is a do-not-say until someone adds the source.

Brief structure that survives handoff

One page per client workspace:

1. Voice in five bullets (tone, sentence length, CTA style)
2. Approved claims with source links or doc anchors
3. Do-not-say list as exact phrases plus concept bans
4. Personalization allowed fields (title, company news with URL, mutual connection, recent post)
5. Personalization forbidden patterns (flattery about appearance, guessing revenue, implying a relationship that does not exist)
6. Escalation who can approve an exception for one campaign

Store it in the client workspace, version it, and date the last review. Slack pins rot.

How to keep personalization alive under the bans

Personalization should come from verified prospect evidence: role, company, a public post, a hiring page, a funding announcement with a link. Ban unsupported color, not specificity.

Practical rules:

• Prefer one concrete observation over three vague compliments.
• Require a source URL or CRM field provenance for any company-specific claim.
• If the evidence field is empty, fall back to ICP-level copy. Do not invent a hook.
• Separate merge tokens (safe fields) from freeform AI sentences (must pass the do-not-say filter).

That split is how you stop a {{company}} token from becoming "loved your Series D with MegaFund" when no Series D exists.

Enforcement that does not become endless QA

Layer the controls:

1. Prompt / guidance block loads the client do-not-say list at generation time.
2. Post-generation filter blocks banned phrases and competitor mentions before human review.
3. Human QA sample on new clients and after brief changes (for example first 20 sends, then spot checks).
4. Send gate still requires compliance verdict and unsubscribe URL. Copy guardrails do not replace legal opt-out controls.

Auto-reject is for clear bans. Ambiguous claims go to a human queue. Do not make reviewers re-litigate "sounds respectful" without the brief open.

Agency multi-client traps

• Cloning Client A's brief into Client B because "they are both SaaS"
• A global agency ban list that hides client-specific bans operators need
• Updating the sales deck but not the do-not-say list, so the agent keeps pitching retired messaging
• Letting personalization pull from stale enrichment and treating it as fact

Client-isolated workspaces help only if the brief lives inside the workspace and is required before supervised or autopilot send modes go live.

Monthly refresh ritual

Once a month per active client:

• Remove bans that no longer apply
• Add new bans from sales objections and legal notes
• Re-test three sample generations against the list
• Confirm approved claims still match the live site and deck

A do-not-say list that never changes becomes either ignored or overly broad. Keep it tight, enforced, and tied to real evidence fields so personalization stays specific without becoming fiction.

Sources: OutboundOS prompt guidance, Dureach AI personalization guardrails, Outsales AI email writer rules, mm-ais AI SDR compliance guardrails, GTM product knowledge.