personalized-outbound-ab-test

Run this before you let an agent write outbound at scale. It produces a measured answer to "does personalization actually beat our generic message?" and a holdout that receives the winner.

The play

1. Split the list into three. Control (e.g. 100) gets the generic baseline. Personalized (e.g. 100) gets researched messages. Holdout (the rest) waits for the result, then gets the winner. Without a control you are not testing, you are just spending.

2. Build the do-not-contact list first. Exclude anyone already in an open thread, in the CRM as an active deal, or recently sequenced. Skipping this is how you send a "nice to meet you" to a customer. Do this before enrichment so you never pay to research someone you cannot contact.

3. Enrich only the Personalized arm. Pull current role, headline, employer, and recent activity per lead. The control arm needs nothing, so half your enrichment spend disappears.

4. Write one message per lead, then gate it. Every message must pass all six checks or it silently drops to the baseline:

   • Account-safe — no links, no phone numbers, no mass-template feel.
   • Factually grounded — every personal reference traces to retrieved data. Zero invented facts.
   • Human — no AI tells, in the sender's own rhythm.
   • Respectful — warm, no creepy over-familiarity.
   • Rule-compliant — length cap, CTA placement, banned vocabulary, whatever the sender set.
   • Recognizably theirs — the hook is something only that person would recognize as about them.

5. Re-verify programmatically. Do not trust the model's self-report. Recompute length, dash counts, CTA position, name casing, and banned phrases after generation. Anything that fails gets fixed or dropped to baseline.