Agentic AI Wiki
Playbooks / Domain Playbooks / Sales & GTM Agents

Sales & GTM Agents

Y5
Playbook · Domain Playbooks

Sales & GTM agents: the failure mode is automated spam at scale.

A sales agent that researches accounts, personalizes outreach, and sends at scale is the agent most likely to damage a brand the fastest: it does not crash, it does not error — it confidently sends, and the blast radius is your reputation and your domain deliverability. This playbook is the opinionated version: where the value actually is, why compliance and consent are not legal's problem to bolt on later, why human sign-off scales with reach, and why "personalized at scale" is one bad prompt away from "spam at scale."

STEP 1

The value is in research and personalization, not in send volume.

The instinct is "an agent that sends more emails." That is the failure framing. The durable value of a GTM agent is the work a rep does badly at scale: deep account research, finding the real trigger and the right contact, drafting genuinely relevant outreach, and keeping CRM truth current. Volume is the cheap part and the dangerous part. Decompose the job and you find the high-value, low-risk portion is everything up to but not including the send. Optimize for relevance per touch, not touches per day — a GTM agent measured on volume will reliably produce volume, and volume of irrelevant outreach is the product nobody wants.

STEP 2

Compliance and consent are a precondition, not a post-filter.

Outreach lives inside CAN-SPAM, GDPR/PECR, suppression lists, and per-region consent rules. These are not a wrapper you add after the agent drafts — they are a gate the agent's audience is filtered through before it ever composes, because a perfectly written email to someone who must not be contacted is still a violation and a deliverability hit.

# consent + suppression are an upstream gate, not a post-filter
if not consent_ok(contact, region) or contact in suppression_list:
    return skip(reason="not_contactable")   # never reaches drafting
draft = model(research(account), template, contact)

The catastrophic GTM failure is not a clumsy email — it is a high-volume run against an unfiltered list: domain reputation tanks, you land in spam folders for legitimate mail too, and the damage outlives the campaign. The list filter must be upstream and fail-closed, not a hopeful instruction in the prompt.

STEP 3

"Personalized at scale" is one bad prompt from "spam at scale".

The same machinery that sends a relevant, well-researched note to 50 right-fit accounts sends a generic, slightly-wrong note to 50,000 wrong ones if the targeting or the template regresses — and it does it fast, confidently, with no error. Personalization is not a tone setting; it is evidence that the agent actually understood the account. Require every outreach to cite the specific, verifiable reason this contact is being approached now; an outreach that cannot name a real trigger is spam wearing personalization's clothes and must not send.

STEP 4

The right autonomy: human sign-off scales with reach.

Autonomy should be inversely proportional to how many people a mistake reaches and how irreversible it is. A drafted email a rep reviews and sends is low-risk — the human is the gate. A fully autonomous send to thousands is the configuration where one regression is a brand event. The pattern that fits: the agent runs research, targeting, and drafting fully autonomously; a human approves the batch (or a representative sample plus the targeting logic) before send; and an automated send is permitted only within tight, pre-approved volume and segment limits with a kill switch on reply-sentiment and bounce-rate spikes.

Approve the targeting logic and a sample, not every individual email — reviewing 5 representative drafts plus the segment definition catches a regression that reviewing 500 individually would not, and it is the only review that scales with the agent.

STEP 5

The eval signal: relevance and harm, not open rates alone.

Open and reply rates are gameable and lagging; a subject line can lift opens while the body burns the relationship. The eval that steers a GTM agent grades, on a sample of drafted (pre-send) outreach:

  • Relevance / personalization grounding — does each draft cite a real, verifiable trigger specific to this account, judged against the source, not on how personalized it sounds.
  • Compliance — audience correctly filtered for consent and suppression; opt-out present; claims about the prospect accurate, not fabricated to flatter.
  • Spam-risk — would a human mark this as spam; weighted heavily negative, because the cost of a spammy send is borne by every future legitimate email from your domain.
STEP 6

The sales & GTM checklist.

  • Optimize relevance per touch, not touches per day; value is in research/personalization, not send volume.
  • Consent and suppression filtering is an upstream, fail-closed gate before drafting — never a post-filter or prompt instruction.
  • Every outreach must cite a real, verifiable trigger; no nameable trigger ⇒ do not send.
  • Autonomy inversely scales with reach: rep-reviewed sends are low-risk; mass autonomous send needs pre-approved volume/segment caps.
  • Humans approve targeting logic plus a sample, not every email; that is the review that scales.
  • Kill switch on bounce-rate and reply-sentiment spikes; domain deliverability is monitored.
  • Eval grades relevance grounding, compliance, and spam-risk on pre-send drafts; spam-risk weighted heavily.

The honest tradeoff: a GTM agent's leverage is reach, but reach is exactly what converts one bad prompt into thousands of spam complaints and a poisoned sending domain — so you cap volume and keep a human on the targeting precisely because the thing that makes it valuable is the thing that makes it dangerous.