Today’s plays › All plays › Cold email
Have AI score every AI-drafted email from 1 to 10 and send anything under 7 to a person before it goes
Picked and checked by Olga Pearson, SalesUpLevel
Get tomorrow’s play in your inbox. Free ↓
Reported in Scale Venture Partners' write-up: Snowflake's reply rate went from 0.5% on an earlier 20,000-prospect push to 7.6% with the AI workflow (Scale's LinkedIn post calls it a 7.6% positive reply rate), and the team booked more than 2,000 meetings from 55,000 prospects.
AI can draft a good email eight times out of ten. At thousands of prospects, the other two add up to hundreds of emails that hurt you.
Jeff Long runs GTM operations for Snowflake's 300-person SDR team. When an SDR adds a prospect to a sequence in Outreach, Workato pulls their own data (product usage, account details) and outside data (company news, contact details), runs a web search, finds similar Snowflake customers, and has Claude draft the email. A separate check scores each draft from 1 to 10 for accuracy, tone and relevance, and flags anything under 7 for an SDR to review. Each task gets its own short prompt. They tested it with 10 SDRs, each putting half their prospects in the AI sequence and half in the usual one.
By Jeff Long (Snowflake), in a Scale Venture Partners write-up, Senior Technical GTM Operations Manager, Snowflake (at the time of the talk) · How We Scaled This (Scale Venture Partners), from Scale's GTM AI Summit talk ↗ · Jan 14, 2026
Run it before your 10am
- Start the workflow from one SDR action, like adding a prospect to a sequence.
- Pull your own data and outside data automatically, so no SDR researches by hand.
- Give each task its own short prompt: research, draft, follow-ups, call script.
- Add a second AI check that scores each draft 1 to 10 for accuracy, tone and relevance.
- Send anything under 7 to an SDR; only drafts that pass go into the sequence.
- Pilot with 10 SDRs: half their prospects on the AI sequence, half on the usual one, with a shared chat for problems.
You are checking a cold email before it is sent. Prospect facts we trust: [name, title, company, recent news, product usage]. Draft: [paste]. Score it 1 to 10 on each: accuracy (every claim matches the facts above), tone (sounds like a person, not a template), and relevance (right for this person's role). List any claim not supported by the facts. Give an overall score. If any score is below 7, say exactly what to fix.
If you lead a team
- Pilot with a small group and a half-and-half split, so you have a real comparison.
- One prompt per task beats one giant prompt.
Watch out for
- The talk was at Scale's GTM AI Summit in late 2025 and the write-up is from Jan 2026; tools and speeds will have changed.
- The 0.5% baseline is from an earlier campaign a few years back, not a side-by-side test, and the sources do not define reply rate the same way.
- This is a big team with an ops engineer and an orchestration tool; start smaller.
- Check your company's rules before you send customer data to an AI model.
- What I'd have my team try first: run the scoring prompt on 20 AI drafts you were about to send and read every one scored under 7.
Source: 'How Snowflake 15x'd replies with AI orchestration' by Anna Held, How We Scaled This (Scale Venture Partners), Jan 14, 2026, based on Jeff Long's talk at Scale's GTM AI Summit; Scale's LinkedIn post (Jan 2026) gives a 7.6% positive reply rate. Results are as reported; we did not measure them.
Get tomorrow’s play before your first call
Free. One sourced play a day.