AI for Work: Using AI for email drafting

Using AI for Email Drafting

What This Is

AI email drafting uses large language models (LLMs) to generate, refine, or personalize professional emails—saving time while maintaining tone, clarity, and relevance. It matters because emails are a daily bottleneck: professionals spend 28% of their workweek on email (McKinsey), and poorly written messages waste time, damage relationships, or create misunderstandings. Example: A sales rep uses AI to draft a follow-up email to a prospect, adjusting tone (friendly vs. formal) and including key product details pulled from a CRM—cutting drafting time from 15 minutes to 2.

Key Facts & Principles

• Tone calibration: AI can match the style of your email (e.g., concise and direct for executives, warm and detailed for clients). Example: Prompt: "Write a polite but firm email to a vendor who missed a deadline. Use a professional tone, no apologies, and propose a new timeline."
• Context window: LLMs process a limited amount of text (e.g., 4K–128K tokens). Long emails or threads may get cut off. Example: If pasting a 10-email thread, summarize key points first or use a tool with a large context window (e.g., Claude 3).
• Personalization tokens: Use placeholders (e.g., {recipient_name}, {project_name}) to dynamically insert details. Example: "Hi {first_name}, I noticed your team’s work on {project}—here’s how we can help."
• Hallucination risk: AI may invent details (e.g., fake deadlines, incorrect names). Always verify before sending. Example: If AI drafts, "Per our call on Tuesday," but you never spoke, delete or correct it.
• Structured prompts: Specify purpose, audience, tone, and key points to avoid generic outputs. Example: "Draft a 3-sentence email to my manager requesting budget approval for a new tool. Include: (1) ROI data, (2) urgency, (3) next steps."
• Iterative refinement: Treat AI as a first draft tool, not a final product. Edit for nuance, brevity, or cultural fit. Example: AI might overuse "I hope this email finds you well"—delete or replace with something fresher.
• Privacy boundaries: Avoid pasting sensitive data (e.g., financials, PII) into public AI tools. Use enterprise-grade LLMs (e.g., Microsoft Copilot, Google Workspace AI) or redact details. Example: Replace "Our Q3 revenue was $2.4M" with "Our Q3 revenue met targets."
• Call-to-action (CTA) clarity: AI often buries the ask. Explicitly prompt for a clear CTA. Example: "End the email with: ‘Please reply by EOD Friday with your availability for a 15-minute call.’"
• Cultural sensitivity: AI may default to Western norms. Specify adjustments for global audiences. Example: "Avoid humor or idioms—this recipient is based in Japan and prefers formal language."

Step-by-Step Application

1. Define the goal
2. Ask: What’s the email’s purpose? (e.g., request, update, apology, pitch).

3. Example: "I need to decline a meeting request but keep the door open for future collaboration."

4. Gather context

5. Pull key details: recipient’s role, past interactions, deadlines, or data points.

6. Example: "The recipient is a potential client who asked for a demo last month. Our last email had low engagement."

7. Write a structured prompt

8. Include:
   • Purpose (e.g., "Follow up on a stalled deal").
   • Audience (e.g., "VP of Marketing at a SaaS company").
   • Tone (e.g., "Confident but not pushy").
   • Key points (e.g., "Mention their pain point: low user adoption. Offer a case study.").
   • CTA (e.g., "Ask if they’d like to reschedule the demo.").

9. Example prompt: > "Draft a 4-sentence email to a VP of Marketing at a SaaS company. Purpose: Follow up on a stalled demo request. Tone: Professional and helpful. Key points: (1) Reference their challenge with user adoption, (2) Share a case study where we improved adoption by 30%, (3) Offer to reschedule the demo. CTA: Ask if they’d like to pick a new time this week."

10. Generate and refine

11. Run the prompt in your AI tool (e.g., Gmail’s "Help me write," Outlook Copilot, or a standalone LLM).
12. Edit for:
    • Brevity (cut fluff; aim for 3–5 sentences).
    • Clarity (replace jargon with plain language).
    • Tone (does it sound like you?).

13. Example edit:

    • AI draft: "I hope this email finds you well. I wanted to circle back regarding our previous conversation about your user adoption challenges..."
    • Revised: "Following up on your demo request—we’ve helped similar SaaS teams boost adoption by 30% (case study attached). Would you like to reschedule for later this week?"

14. Add personal touches

15. Insert specific details the AI couldn’t know (e.g., inside jokes, recent news about the recipient).

16. Example: "Congrats on the recent funding round—excited to see where [Company] goes next!"

17. Verify and send

18. Double-check:
    • Facts (dates, names, data).
    • Links/attachments (do they work?).
    • Tone (read aloud to catch awkward phrasing).
19. Example: If AI mentions a "recent call," but you haven’t spoken, replace with "our email exchange."

Common Mistakes

• Mistake: Using AI for every email without editing.

• Correction: Treat AI as a co-pilot, not a replacement. Always review for tone, accuracy, and personalization. Why? Over-reliance leads to generic, impersonal emails that hurt relationships.

• Mistake: Ignoring the recipient’s communication style.

• Correction: Adapt the AI’s output to match the recipient’s preferences (e.g., short vs. long emails, formal vs. casual). Why? A CFO may ignore a chatty email; a startup founder might dismiss a stiff one.

• Mistake: Pasting confidential data into public AI tools.

• Correction: Use enterprise-grade AI (e.g., Microsoft 365 Copilot) or redact sensitive info. Why? Public LLMs may store or leak data.

• Mistake: Overloading the prompt with too many details.

• Correction: Break complex emails into multiple prompts (e.g., draft the body first, then the subject line). Why? LLMs perform worse with overly long or vague instructions.

• Mistake: Assuming AI understands context from past emails.

• Correction: Summarize key points in the prompt. Why? Most LLMs don’t retain memory between sessions.

Practical Tips

• Create email templates for common scenarios (e.g., meeting requests, follow-ups, apologies) and customize with AI. Example: Save a "Cold outreach to prospects" template and tweak it per recipient.
• Use AI for subject lines: Prompt "Generate 3 subject line options for an email about [topic]. Make them concise and intriguing." Example: For a delayed project update: "Quick update on [Project] timeline" vs. "Heads up: [Project] delay (here’s why)."
• Batch draft emails: Use AI to generate multiple versions of an email (e.g., formal vs. casual) and pick the best fit.
• Set up a "review buddy" system: Have a colleague quickly scan AI-drafted emails for tone or errors before sending. Example: "Does this sound too aggressive?"

Quick Practice Scenario

Scenario: You’re a project manager emailing a client about a delayed deliverable. The client is usually direct and dislikes vague updates. You need to:
1. Acknowledge the delay.
2. Explain the reason (a key team member is out sick).
3. Propose a new timeline.
4. Offer a concession (e.g., a discount on the next phase).

Question: What’s a concise, client-friendly prompt to generate this email?

Answer:

"Draft a 4-sentence email to a client about a project delay. Tone: Direct and transparent. Key points: (1) Acknowledge the delay, (2) Explain the reason (team member illness), (3) Propose a new timeline (2 weeks later), (4) Offer a 10% discount on the next phase. Avoid apologies—focus on solutions."

Explanation: The prompt specifies tone, structure, and key details while avoiding fluff (e.g., "I hope you’re well"), which the client would likely ignore.

Last-Minute Cram Sheet

1. AI email drafting = co-pilot, not autopilot—always edit.
2. Tone > perfection: Match the recipient’s style (e.g., formal for executives, casual for peers).
3. Prompt structure: Purpose + audience + tone + key points + CTA.
4. Hallucination trap: Verify names, dates, and facts before sending.
5. Context window limits: Summarize long threads or use tools with large windows (e.g., Claude 3).
6. Personalization tokens: Use {placeholders} for dynamic details (e.g., {recipient_name}).
7. Privacy rule: Never paste PII or confidential data into public AI tools.
8. CTA clarity: End with a specific, actionable ask (e.g., "Reply by Friday").
9. Cultural sensitivity: Adjust tone/idioms for global audiences.
10. Batch drafting: Generate multiple versions (e.g., short vs. long) and pick the best.