Draft to Delivered: The Three-Pass Email Workflow
The tools above are the inventory. The workflow that turns them into emails worth sending has three passes, and skipping any of them is where AI email goes wrong.
Pass one: the brief, not the button. The single biggest quality difference in our testing was not between tools. It was between users who hit "generate" with a vague instruction and users who briefed the AI like a colleague: who the recipient is, what the email must achieve, what tone fits the relationship, what action it should end with. The recipient-brief framework in the next section makes this concrete; the point here is that the thirty seconds of briefing replaces the ten minutes of fixing a generic draft, every time.
Pass two: the human read that earns the send button. Every AI draft gets one deliberate pass before sending, and the pass has a specific job list: verify any facts, dates, names, and numbers the AI included (it invents these with total confidence), inject the one specific detail that proves a person wrote this (the reference to last week's call, the genuinely personal line), and check the emotional register against the actual relationship, because AI defaults to a politeness that reads as distance with people you know well. The test that works: would you be comfortable if the recipient knew exactly how this email was produced? If yes, send. If you would wince, the draft needs more of you in it.
Pass three: the delivery layer most people skip. What happens after "send" is also an AI problem now. Subject lines decide opens, so generating five variants and picking against your audience (or A/B testing them inside HubSpot-class platforms) beats accepting the first suggestion. Send timing is predictable enough that platform AI optimises it from recipient behaviour. And for volume senders, deliverability is its own discipline: the spam-filter ecosystem is itself AI, which is why the Instantly-class infrastructure (warm-up, variation, sender rotation) exists. For everyday senders, the takeaway is lighter: a strong subject line and a sensible send time are free response-rate improvements sitting one prompt away.
The pattern across all three passes: AI compresses the typing, the human supplies the truth and the relationship, and the delivery layer makes sure the work actually gets read.
The Recipient Brief: A Prompt Framework for Emails That Land
Email prompts fail in a specific way: they describe the email the sender wants to write instead of the recipient it needs to land on. The fix is a six-question brief (the journalist's five Ws and an H, pointed at your reader), answered before generating anything.
Who is reading this? Role, relationship to you, and what they are dealing with. "A small business owner struggling with lead generation" produces a different and better email than "a prospect".
What must it achieve? The core message and the single action you want. One ask per email; AI happily writes three and the response rate pays for it.
Why should they care? The benefit in their terms, not your feature list. This is the question most skipped briefs miss, and its absence is exactly what makes AI email smell like AI email.
When are they reading it? Context and urgency: a renewal deadline, a just-announced funding round, a Monday-morning inbox versus a Friday afternoon one.
Where will it be read? A phone screen between meetings argues for three sentences; a considered desktop read can carry a structured pitch.
How should it sound? Tone, anchored to the relationship: "professional but warm, like a vendor they already trust", not just "friendly".
Assembled, the brief becomes the prompt: "Write a concise, benefit-driven email to a small business owner struggling with lead generation. Introduce our CRM as the fix, leading with automated lead nurturing and conversion improvement, in a professional but empathetic tone. They will likely read it on a phone. End with one ask: a 15-minute demo." Compare that with "write a sales email" and you have the entire gap between AI drafts that get edited and AI drafts that get deleted.
Two upgrades once the basic brief is habit: paste your three best past emails to that audience into the prompt as voice reference, and for stakes that justify it, add "then critique this draft as the recipient: what would make them ignore it?" The AI's self-critique pass catches the generic phrasing its first pass produced.
Where AI Email Actually Moves Numbers
The improvement claims in this category deserve scepticism, so rather than recite a vendor case study with conveniently round percentages, here is the pattern our testing and the published reply-rate data consistently support: where AI email assistance produces measurable lift, and where it only produces volume.
The follow-up gap is the highest-value target. The most reliable revenue leak in any sales or client operation is not bad first emails. It is the follow-ups that never get sent because reps are busy and humans forget. AI closes this gap structurally: drafts generated from the actual interaction history (the call notes, the previous thread, the expressed interest), reviewed and lightly edited by the rep, sent on schedule. The lift comes less from better prose than from the messages existing at all, personalised enough to deserve a reply. This is the pattern behind every credible AI-email success story we examined, and it is replicable by anyone with a CRM and twenty minutes of setup.
Personalisation lifts replies only when it is real. Emails referencing genuine, verified, recipient-specific signals outperform generic sends meaningfully; emails with AI-generated pseudo-personalisation ("I noticed your company is doing great things") perform no better than templates and increasingly worse, because recipients have learned the pattern. The variable that predicts lift is not whether AI wrote the email. It is whether the email contains anything true and specific.
Subject lines and timing are the cheap wins. Testing AI-generated subject line variants and letting platform AI pick send times are the lowest-effort improvements with consistently measurable open-rate effects, available inside tools most senders already pay for.
And the honest ceiling: AI cannot make people want what the email offers. Teams that deployed AI email expecting it to fix a weak offer or a poorly targeted list got more emails, not more replies. The technology amplifies the quality of what feeds it (the list, the offer, the genuine signal), which has been the refrain of this entire guide series for a reason.
Use Case Scenarios
If you are a knowledge worker drowning in routine email, start with Gemini in Gmail or Copilot in Outlook. Add Grammarly Premium at $12/month for tone control. This stack costs nothing extra if you already have a Workspace or Microsoft 365 licence, and it handles ninety percent of professional email.
If you are a salesperson doing outbound, Lavender at $29/month plus your existing Gmail or Outlook is the right setup. The response rate improvement pays for the subscription almost immediately if you are sending more than a few cold emails a week.
If you are an executive or founder where email is a major time drain, Superhuman at $30/month is genuinely worth considering. The combination of fast UI, AI drafting, and Auto Summarize on long threads can recover an hour a day.
If you are running marketing campaigns, do not subscribe to a separate AI email tool. Use the AI features inside whatever marketing platform you already have. HubSpot, Klaviyo, ActiveCampaign, Mailchimp, and Customer.io all have native AI now, and the campaign performance data they have access to makes their drafts more relevant than any standalone writer.
If you are an agency running cold outreach for clients, Instantly at $37/month plus Lavender for the senior reps is the standard stack. Add Apollo or Clay for enrichment.
If you are someone who writes one or two important emails per week where the wording really matters, you do not need an email-specific tool. Claude Pro or ChatGPT Plus at $20/month covers it and gives you a general AI assistant for everything else.