Client voice briefing for AI outbound agents
If your AI outbound agent only gets a product one-liner and a tone adjective like "professional but friendly," it will invent the rest. Agencies cannot afford that. Off-brand claims land in the client's inbox and on the client's domain.
A workspace brief is the antidote. It is not a brand PDF dumped into a prompt. It is a short, structured packet the agent can retrieve every time it writes: offer, proof, tone, and hard stops.
GTM writes from prospect evidence plus client voice. The brief is how you keep that voice honest across every client workspace.
Why templates fail and briefs work
Templates freeze a sentence shape. Briefs freeze the facts and constraints. When the prospect changes, a template still sounds like a template. A brief lets the agent assemble a new message from approved materials without inventing metrics, customers, or promises.
Market research on agent voice guides in 2026 keeps landing on the same pattern. Agents need examples they can imitate, phrases to avoid, claims they may make, claims that need evidence, and moments that require human review. AgentCeres puts it bluntly: treat the model as a drafter, not a publisher. Metaflow's brand knowledge layer framing adds structure: identity, voice, offers, evidence, prohibitions.
Agencies need that structure per client, not once for the agency brand.
What belongs in the workspace brief
Keep it short enough that someone will maintain it. One working page beats a 40-page deck nobody updates.
Offer. What the client sells in plain language. Who it is for. Who it is not for. Primary CTA for outbound (usually a booked conversation, not a feature dump). Pricing language only if sales has approved exact words.
Proof. Approved facts only: named features that exist, cleared quotes, public metrics the client will stand behind. If a number is not in this section, the agent may not invent one. "Cite or cut" is the rule.
Tone. Concrete rules, not vibes. Sentence length. Whether first person is "we" or "I." Humor ceiling. How direct the ask can be. Pair "sounds like" with "does not sound like" examples from real client writing. Positive samples alone push models toward generic marketing mush.
Forbidden claims. Guaranteed ROI. Invented case studies. Competitor digs the client did not approve. Regulated categories the client is not cleared to discuss. Words that trigger legal review. Anything that would fail a compliance gate for honesty.
Audience notes. What this ICP already knows, distrusts, and ignores. Agency clients often sell to other operators. The brief should say so.
Escalation. When the agent must stop drafting and ask a human: discounts, legal language, sensitive replies, anything outside the proof section.
How to gather it without a two-week brand workshop
Pull from what already exists.
- Ask the client for three emails they are proud of and three they would never send again. Annotate why.
- Export the current landing page offer and pricing page language. Mark what is approved for cold outbound. Often the website overclaims relative to what sales will say on a first call.
- List five claims that have caused trouble before (inflated timelines, "AI replaces your team," named logos without permission). Put them in the forbidden list.
- Name the owner. Briefs rot when "marketing" owns them and outbound never sees updates.
Version the brief. When the offer changes, bump the version and retire old sequences that still cite dead claims.
Wiring the brief into the write step
A brief sitting in Drive does nothing. The write step must load it every time.
Practical pattern:
- Load identity + offer + prohibitions for every draft.
- Load evidence only for claims the draft actually needs.
- Fail closed on missing proof: if the draft wants a number and the brief has none, delete the number.
- Keep a human gate on first live sends for a new client, and on any copy that touches pricing or guarantees.
GTM's motion already expects writing from prospect profile and signals inside a client-isolated workspace. The brief is the client half of that grounding. Prospect evidence answers "why this person." The brief answers "what we are allowed to say as this client."
Common failure modes
Adjective soup. "Bold, empathetic, innovative." Agents ignore this. Replace with examples.
Proof without provenance. A spreadsheet of "wins" with no source is invitation to hallucinate. Prefer short, sourced bullets.
One brief for every client. Multi-ICP agencies need isolation. Client A's voice leaking into Client B's sequence is a trust failure, not a copywriting quirk.
No forbidden list. Models fill silence with confident fiction. Empty prohibitions are a bug.
Skipping review forever. Supervised sending exists for a reason. Autopilot on brand-risk copy is how agencies lose clients.
A one-page checklist before first AI writes
- Offer and CTA approved by the client sponsor
- Proof list with sources
- Sounds-like / does-not-sound-like pairs
- Forbidden claims and banned phrases
- Escalation rules for pricing and legal
- Owner and last-updated date
- Sample messages reviewed by a human before live volume
If any line is blank, do not let the agent send. Drafting is fine. Shipping is not.
Sources
- Brand Knowledge Layer for AI Agents (Metaflow)
- How to Keep AI Marketing On-Brand and Accurate (AgentCeres)
- Your AI Agent Needs a Voice Guide (MarketMai)
Want outbound agents that write from client voice and prospect evidence inside isolated workspaces? Join the GTM waitlist.