AI Personalization SOP
From raw prospect data to QC-approved personalization lines ready to inject into any cold email sequence. First run: 60 to 90 min. Subsequent runs: 15 to 20 min.
Before You Start
What this SOP produces: output, time, and prerequisites
Output: QC-approved AI personalization lines mapped to contact records, ready to inject as custom variables in cold email sequences. Time: 60 to 90 min for first-run setup; 15 to 20 min per subsequent batch once prompt and columns are locked.
Enriched lead list with first name, last name, job title, company name, domain, and at least one live signal field (LinkedIn post, job change under 90 days, tech in use, or funding event). Access to Clay or Lyne.ai / Smartwriter.ai, plus Instantly or Smartlead with at least one active mailbox.
Workflow Overview
The 5-step AI personalization SOP at a glance
| Step | Action | Tool | Output |
|---|---|---|---|
| 1 | Assemble and validate personalization inputs | Clay / Lead database | Structured input columns per contact |
| 2 | Write and scope the AI prompt | Clay AI column / OpenAI | Tested, bounded prompt template |
| 3 | Run bulk generation and export raw output | Clay / Lyne.ai / Smartwriter.ai | Raw personalization CSV, one line per contact |
| 4 | QC filter: score, sample, and reject unsafe lines | Lavender / manual review | Approved output batch, rejection log |
| 5 | Inject approved lines into sequence as custom variable | Instantly / Smartlead | Live campaign with safe, field-mapped personalization |
Step by Step
Complete AI personalization SOP: inputs to QC to safe output
- Step 1: Assemble and validate your personalization inputs
Confirm each contact has job title, company name, domain, and one live signal field (LinkedIn post, job change under 90 days, tech in use, or funding event). Remove any contact where the signal is blank or older than 90 days before running.
- Step 2: Write and scope your AI prompt
Build the prompt with four components: role instruction ("Write a one-sentence opening line"), scope limit ("Do not infer revenue or headcount"), output format rule ("Output the line only, no greeting"), and fallback ("If signal is empty, output SKIP"). Test on 10 contacts and fix if more than 2 of 10 fail on accuracy, tone, or unverifiable claims.
- Step 3: Run bulk generation and export raw output
In Clay, add an AI column using the tested prompt, map input fields, and run on the full list. Export to a CSV named with the batch date and prompt version. Do not inject anything at this stage.
- Step 4: QC filter β score, sample, and reject unsafe lines
Filter out rows with the SKIP token, lines over 35 words, and any unverifiable claim (inferred revenue, fabricated product name, wrong title). Then manually review a random 10% sample: if more than 5% fail, stop and fix the prompt before proceeding. A rejection rate above 20% is a prompt problem, not a data problem.
- Step 5: Inject approved lines into your sequence as a custom variable
In Instantly or Smartlead, create a custom variable (e.g., {{personalization}}), import the QC-approved CSV, map the column, and verify rendering on 3 test contacts before activating. Position the variable in sentence one, not sentence two.
A prompt producing fabricated claims at a 10% rate means 50 bad emails in a batch of 500. The reputational cost far exceeds the time saved by skipping the test.
Common Failures
What breaks in the AI personalization SOP and how to fix it
Most failures trace back to stale input data or an under-scoped prompt. Both are preventable at Steps 1 and 2.
Tool Stack
AI personalization SOP: which tool fits each step
Four tools cover the full SOP. Match each to the step it handles rather than combining multiple tools in a single step.




Clay for Steps 1, 2, and 3 paired with Lavender for Step 4. Lyne.ai and Smartwriter.ai substitute for Step 3 when Clay is not in the stack. See the AI Personalization directory for a full comparison.
SOP running? Protect deliverability before you scale volume.
The AI Personalization and Deliverability guide covers the specific spam triggers that AI-generated copy introduces and how to avoid them before scaling send volume.