The single biggest lever for getting AI to write emails that sound like you is a structured voice brief. Not a style guide, not a one-line prompt instruction, but a specific document that gives the AI enough information about your tone, vocabulary, structural habits, and anti-patterns to replicate the way you actually write. Without it, every AI draft sounds like a slightly different brand. With it, you get a usable first draft that needs editing, not a total rewrite.
Key Takeaways
- A voice profile isn't a style guide. It's a structured brief built specifically for AI context, containing tone descriptors, vocabulary lists, annotated writing samples, structural patterns, and explicit anti-patterns.
- Five to ten annotated writing samples are enough. Quality of annotation matters more than volume. The AI needs to understand why a piece works, not just see the finished output.
- Testing is binary. If someone who knows your brand can't tell whether you or the AI wrote a given paragraph, the brief is working. If they can, it isn't.
- Most generic AI output is a briefing failure, not a tool failure. The AI is doing exactly what you asked. If you didn't tell it how you sound, it defaulted to how everyone sounds.
What Does a Voice Profile for AI Actually Contain?
A brand voice document that sits in a Google Drive folder and describes your tone as "professional yet approachable" isn't a voice profile. That phrase means nothing to an AI. It's so broad that it describes every B2B SaaS company that has ever existed, and the AI will treat it accordingly.
A voice profile built for AI context contains specific, pattern-level information the model can act on. From my experience building Digiwell's own voice system, I've learnt that the difference between a usable voice profile and a decorative one comes down to specificity. The AI doesn't need to know that your brand is "warm." It needs to know that you start most emails with a concrete observation rather than a question, that you use contractions always, and that you never use the word "leverage" as a verb. The Content Marketing Institute's research on brand voice and AI confirms this. Teams that provide structured voice context get dramatically better first drafts than teams that rely on general tone descriptions (source: contentmarketinginstitute.com/articles/brand-voice-ai-content/).
The 5-Part Brief Structure That Actually Works
When I built the voice brief we use internally at Digiwell, I tested dozens of formats before settling on a five-part structure. Each part addresses a different dimension of voice that AI tends to flatten when left to its defaults.
- Tone descriptors with calibration. Don't just say "conversational." Say "conversational the way a peer would explain something at a coffee meeting, not conversational the way a brand tries to sound relatable on social media." Give the AI a spectrum. "More direct than encouraging. More specific than inspirational. More peer than mentor."
- Vocabulary and phrase lists. Two columns: words and phrases you use deliberately, and words and phrases you never use. This is where you catch the generic language before it enters the draft. If you never say "streamline" or "empower," list them as anti-vocabulary. If you always say "the system underneath" or "the leak," list those as signature phrases.
- Annotated writing samples. This is the most important part and the one most people skip. Pull five to ten of your best emails or content pieces and annotate them. Mark why specific paragraphs work. "This opening works because it starts with a real moment, not an abstract claim." "This CTA works because it asks a reflective question." The annotations teach the AI the reasoning behind your voice, not just the surface pattern.
- Structural patterns. Document your habits. Do you write in long flowing paragraphs or short punchy ones? Do you use bullet lists or avoid them? Do you open with a story, a data point, or a direct statement? How long are your sentences on average? These structural signals are half of what makes writing recognisable, and most briefs ignore them entirely.
- Anti-patterns. Explicitly list what you don't do. No rhetorical questions as openers. No three-word sentence fragments for emphasis. No em dashes. No exclamation marks. No "I hope this email finds you well." Anti-patterns are as important as patterns because AI defaults to the most common conventions unless you tell it not to.
| Brief Component | What It Tells the AI | Common Mistake | |---|---|---| | Tone descriptors | How to calibrate register and warmth | Using vague words like "friendly" without examples | | Vocabulary lists | Which words to reach for and which to avoid | Only listing positive preferences, skipping anti-vocabulary | | Annotated samples | Why your best writing works, not just what it looks like | Providing raw samples without any annotation | | Structural patterns | Paragraph length, opening style, transition habits | Assuming the AI will infer structure from samples alone | | Anti-patterns | What your brand explicitly doesn't do | Leaving this section out entirely |
How Many Writing Samples Does AI Need?
From what I've seen, five is the minimum and ten is the sweet spot. Beyond ten, you hit diminishing returns unless your writing varies significantly across formats. The key isn't volume. It's annotation quality and consistency within the sample set.
Pick samples that represent your best work, not your most recent, and choose from the same format you're asking the AI to produce. If you want AI to write marketing emails, brief it with your best marketing emails. Feeding it blog posts and expecting email-quality output is like showing someone your oil paintings and asking them to replicate your watercolour style. The HBR research on effective AI prompting reinforces this. Context specificity is the single strongest predictor of output quality (source: hbr.org/2023/06/how-to-write-good-prompts-for-chatgpt).
How to Test if the Brief Is Working
Testing is simpler than most people make it. Generate three to five email drafts using your voice brief, then show them alongside emails you actually wrote to someone who knows your brand well. Ask one question: which ones did you write? If they can't reliably tell, the brief is working.
If they spot the AI drafts immediately, ask what gave it away. Their answers tell you exactly what to adjust. Usually it's something specific: "You never open with a question," or "This sounds too polished, you're usually more blunt." Those observations go directly into your anti-patterns and structural patterns sections.
I run this test quarterly at Digiwell. Each round surfaces one or two adjustments that tighten the brief further, and that compounds over time. After three rounds of refinement, our AI drafts need about 20 percent editing rather than the 60 to 70 percent they needed when we started.
What Makes AI Sound Generic in Email Copy?
The leak is almost always in the brief, not in the tool. From my experience auditing email programs, here are the patterns I see most often when AI output sounds like it could belong to any brand.
No anti-patterns defined. If you don't tell the AI what not to do, it defaults to the most common patterns in its training data. That means opening with a question, using filler transitions like "but here's the thing," and closing with "ready to get started?" The AI will reach for generic marketing conventions unless you explicitly block them.
Samples without annotation. Dropping five emails into a prompt and saying "write like this" gives the AI surface-level mimicry. It might match sentence length, but it won't capture the reasoning behind your choices. Annotations teach it the difference between copying your style and understanding your style.
Tone words without calibration. "Professional but casual" isn't actionable. Calibrated descriptors like "the directness of a Slack message with the substance of a strategy memo" give the AI something concrete to aim for.
Ignoring structural habits. Voice isn't just word choice. It's paragraph length, transition style, where you place your CTA, whether you use headers as questions or statements. Two writers can use identical vocabulary and still sound completely different because of structure. Litmus's research on AI in email marketing highlights this structural dimension as a key differentiator in brand-consistent output (source: litmus.com/blog/ai-email-marketing).
How to Build Your Voice Brief in Two Hours
If you don't have a voice brief yet, you can build a working first version in a single sitting. Pull your ten favourite emails or content pieces, the ones where you thought "yes, that sounds exactly like me." Read each one and annotate two to three things that make it distinctly yours. Then look across all ten and extract the recurring structural habits, vocabulary, and tone patterns. Build your anti-vocabulary list by scanning competitor copy for overused words you want to avoid. Finally, write calibrated tone descriptors in comparative terms: "More [X] than [Y]."
The whole process feeds directly into the Context Layer that powers every AI workflow in your email ops stack. If you've already built a context layer, the voice brief slots in as the centrepiece. If you haven't, this is a strong place to start.
Frequently Asked Questions
How do I stop AI from sounding generic in my emails?
Build a structured voice profile that includes annotated writing samples, explicit anti-patterns, calibrated tone descriptors, and vocabulary lists. Generic output happens when the AI has no specific voice context to work with, so it defaults to the most common patterns in its training data. Giving it a detailed brief shifts the output from "sounds like marketing" to "sounds like you."
What should a brand voice brief for AI include?
Five components: calibrated tone descriptors, vocabulary and anti-vocabulary lists, five to ten annotated writing samples showing why each piece works, documented structural patterns like paragraph length and opening style, and an explicit list of anti-patterns covering what your brand never does. Each component addresses a different dimension of voice that AI tends to flatten without guidance.
Can AI match a personal brand voice for email?
Yes, but only with a structured voice brief. AI is very good at pattern replication when given enough specific information. From what I've seen, teams that invest two hours in building a proper voice profile get AI drafts that need 20 to 30 percent editing. Teams that skip the brief get drafts that need 60 to 70 percent editing, which often makes the AI slower than writing from scratch.
How many writing samples does AI need to learn my voice?
Five is the minimum, ten is the sweet spot. Beyond ten, returns diminish unless your voice varies significantly across content types. The critical factor is annotation quality, not sample volume. An annotated sample that explains why a paragraph works teaches the AI far more than three unannotated samples pasted into a prompt. Choose samples from the same format you want the AI to produce, and pick your best work rather than your most recent.
Read Next
- The Complete Guide to Email Operations and AI Workflows covers how the voice brief fits into the broader email ops stack and the Context Layer that powers every AI workflow.
- AI-Assisted Newsletter Workflow walks through the full production cycle with AI touchpoints at each stage, including where the voice brief enters the workflow.
- A Prompt Library for Email Marketing Teams gives you twelve copy-paste prompts organised by use case that pair directly with your voice brief.
Want Help Building Your Voice Brief?
If your AI email drafts still sound like they could belong to any brand, the brief is the first thing to fix. It's the foundation every other AI workflow in your email program depends on.
Our free audit reviews your current email voice and AI workflow, identifies where the brief is falling short, and gives you a prioritised action plan to fix it.