Email & Message Composition

Sonnet 4 preserves a distinct sender voice and personalizes without sounding robotic. Haiku 4 is 5x cheaper and good enough for transactional (receipt, notification, reminder) flows where personalization is shallow. GPT-4o is a close second for English; Sonnet edges ahead on tone preservation across follow-ups.

Alternatives: GPT-4o, Gemini 2.5 Pro

This runs on every message and is latency-sensitive. Flash at 200-400ms TTFT and 5-20k tokens out per dollar is perfect. For structured intent labels (10-20 categories), a fine-tuned Flash or Haiku beats zero-shot by 5-10 points.

Alternatives: GPT-4o-mini, Rule-based classifier

Tone drifts with relationship stage — formal on first touch, casual after 5 exchanges. The cheapest reliable method is to include the last 2-3 messages from the thread as tone anchors plus 3-5 paragraph exemplars of the sender's writing. Fine-tune only for high-volume individual senders (sales reps, CS managers).

Alternatives: Static per-segment config, Fine-tuned per-sender model for high-volume senders

Multi-variant generation costs 2-3x more but lets the scorer pick the best and provides signal for a learned bandit. Temperature 0.7 gives genuine diversity; too low and variants collapse to the same message, too high and they drift off-topic. For transactional flows (receipts), single variant at 0.3 is fine.

Alternatives: Single variant at temperature 0.3, 5+ variants with bandit selection

Static checks catch obvious issues (ALL CAPS, excessive exclamation, spam-trigger phrases, broken links). Provider reputation monitoring catches domain or IP issues. Both are cheap; skipping either costs you 10-30 percent of deliverability on cold outbound.

Alternatives: SpamAssassin, Litmus, Custom blocklist

Reply rate is the only metric that matters for outreach; open/click are weaker proxies. Feed the outcome of every message back to the scorer as training signal. Over 3-6 months, this closes the loop and the scorer's rankings start correlating with real reply rates.

Alternatives: No feedback (pure prompt engineering), Manual A/B tests, RLHF on reply outcomes

Personalization depth vs generation cost

Heavy personalization (recent product behavior, mutual connections, firmographic context) lifts reply rates 2-4x on cold outreach but triples the context retrieval cost and increases input tokens. For transactional messages (receipts, reminders) skip deep personalization — it does not move the needle. For high-value outreach (enterprise, >$50k ACV), personalize hard; for long-tail nurture, template-heavy with light personalization is more efficient.

Multi-variant vs single-variant generation

Generating 3 variants at temperature 0.7 and scoring picks the best message but costs 3x as much. For cold outreach where reply rates are the KPI, the extra cost is easily justified. For transactional flows with low variance in outcome, single variant at low temperature wins. Start multi-variant, measure whether the scorer's 'best' actually wins on reply rate; graduate to single variant when it does not.

Human approval vs fully automated send

Human approval catches embarrassing hallucinations (wrong company name, wrong exec title, off-tone) but adds hours-to-days of latency and caps volume at human throughput. Route by stakes: auto-send for low-stakes transactional, human-approve for executive recipients, regulated industries, and first-touch enterprise. A 5-10 percent sample of auto-sent messages should still route through a human for ongoing QA.

Hallucinated names, titles, or company references

Mitigation: Restrict the composer to use only entities explicitly present in the retrieved context. Post-generation, extract every proper noun and verify against the CRM record and retrieval context. Reject messages with unverifiable entities. Track hallucination rate per model version; a spike usually means a bad context retrieval.

Tone mismatch — cold message reads like a warm follow-up or vice versa

Mitigation: Make the tone profile explicit and structured (formality 1-5, directness 1-5, length preference). Include 2-3 example messages at the target tone in the system prompt. Run a tone-consistency LLM judge; reject variants scoring below threshold. Fine-tune per high-volume sender when scale justifies it.

Sender reputation damage from low-quality generated outreach

Mitigation: Warm up new domains with transactional volume before cold outreach. Monitor bounce, complaint, and unsubscribe rates per domain and per sender. If complaint rate exceeds 0.1 percent or bounce rate exceeds 2 percent, pause that sender and investigate. Never send generated cold outreach from a domain also used for transactional.

Prompt injection via thread history (recipient includes instructions in their reply)

Mitigation: Treat thread history as untrusted input. Strip or neutralize patterns that look like instructions. Use structured output so the composer can only emit subject/body/cta fields. Run a prompt-injection classifier on inbound thread content before including it in the composer prompt.

Generated message triggers spam filters despite passing static checks

Mitigation: Monitor provider-side reputation metrics (SendGrid, Postmark dashboards). Track inbox placement using Litmus or seed-list testing. Feed back deliverability outcomes to the scorer. Maintain a deny-list of phrases that correlate with spam-folder placement for your specific sending pattern.