How to Brief AI So Your Marketing Sounds Like You

If your AI-drafted emails and generated images keep coming out generic, the fix is not a better model or a cleverer prompt. It is a structured account brief: your positioning and category language, your offer, your proof and real numbers, your voice rules, the segment you are writing to, and your own customer records feeding the draft.

That is the whole argument. AI content quality is a data and briefing problem. Most teams are asking a model to write for a business it has never been told anything about, then judging the model when the output reads like everyone else's.

Why generic output is expensive now

Generic is not a cosmetic problem. It costs trust.

Buyers who read a message as machine-made discount it, and the common tells are the same four every time: it arrives too fast to be human, it reads too formal or robotic, the tone is off, or it is generic and repetitive. Read that as a defect list. Every one of those is a briefing failure, not a model limitation.

Meanwhile supply is moving the opposite way. More teams are producing more AI output every quarter, so the volume of average content rises while the tolerance for it falls. The gap between those two lines is where briefing quality decides who wins.

Worth holding onto: people are also bad at detection, and human work regularly gets flagged as machine-made. So the penalty is not attached to whether a machine helped. It is attached to whether the work reads like nothing in particular. That is a style and substance problem you can fix.

A prompt carries a request. A brief carries the business.

A prompt says what you want done. A brief says what is true.

The distinction matters in practice: instructions govern how the model behaves, and reference material holds what is true about your business. Concrete, situation-specific instructions of the "when X happens, do Y" form beat vague description, and short bullets beat long paragraphs.

"Friendly but professional" is not an instruction. It is a mood. The model has to invent the business around it, and the average of every business on the internet is exactly what you get back.

The account brief

Five parts. Write them once, keep them in one place, and make every draft start from them.

1. Positioning and category language. What category you claim, the words you use for it, and the words you refuse. We hold a category line at Payani: AI-native marketing operating system, never AI CRM, never chatbot, never marketing automation tool. Your list will differ. The point is that the naming decision is upstream of every sentence, and if the model does not have it, it will pick the most common label in its training data, which is your competitor's label.

2. The offer. What you sell, to whom, at what commitment, and what the next step is. A draft that closes with the wrong call to action was never a writing problem.

3. Proof and real numbers. This is the part almost every brand-voice framework skips, and it is where AI content stops being bland and starts being wrong. Name your actual customers, actual results, actual claims you are allowed to make, and the claims you are not. Without a proof set, a model will produce plausible numbers. Plausible numbers are the fastest way to lose a deal in legal review.

4. Voice rules as rules, not adjectives. Banned phrases. Sentence length. Whether you use exclamation marks, emoji, or a word like "transform." Whether you open with a question. Concrete constraints travel; adjectives do not.

5. The segment. Who this specific piece is for, what they already believe, and what they are deciding. A brief written to "our audience" produces copy written to nobody.

Personalisation comes from records, not adjectives

Here is the gap in every brand-voice framework currently circulating: they brief tone and stop. Voice without business substance still reads generic, because the substance is what makes a message feel written for one person.

What makes a message feel understood is a clear, relevant answer, and a sender who remembers details from past interactions, and a recommendation with a reason attached. Natural tone finishes behind all of them. Memory and relevance do not come from a prompt. They come from records.

So the brief has to include the account, not just the style guide. Who the contact is. What pages they viewed and how often. What stage the deal is in and how long it has been there. What they replied to last time. Which campaigns actually produced meetings. A model with access to that writes the sentence you would have written. A model with only adjectives writes a sentence about synergy.

This is the design principle behind Ask Payani sitting across every screen in Payani: the drafting layer reads the same contacts, companies, deals, website activity and campaign history the rest of the system runs on. Brand voice and account knowledge feed email marketing, articles and Creative Studio, so a draft starts from the business rather than a blank box. The intelligence is only as good as the record underneath it, which is the honest version of this argument and the reason the CRM and the content engine belong in one place.

Brief images and video the same way

Almost all the advice on this subject is text-only, and it shows in the creative. Stock-feeling visuals cost you credibility whether or not a model made them.

Three things belong in every image and video brief:

• Brand colours and treatment, as actual values. Ours are #144e63, #ded6c4, #c0491f, #f7f4ed and #1c1b19.
• Product truth. What the product actually looks like, what the screen actually shows, who actually uses it. Generated interfaces that do not exist are the visual equivalent of an invented statistic.
• Placement. A LinkedIn still, an email header and a paid social video are different jobs with different crops, different text load and different first frames.

The review loop

Briefing gets you a good first draft. The loop is what keeps you honest.

• Claim check against what ships. Every number, integration and capability in the draft must exist in the product today. No timelines, no coming soon.
• Edit live, in plain language. Fix the draft in the system where the voice rules live, so corrections compound into the brief instead of evaporating into a chat window.
• Read for the defect list. Too formal, tone off, generic, repetitive. Those four cover most of what gives AI drafts away.
• Measure on replies, meetings and pipeline. Open rates are a weak signal in an era of privacy proxies and inflated opens. Judge drafts on whether a human replied and whether a deal moved.

The teams getting real leverage from AI are not prompting better. They are briefing the business properly once, connecting the records, and then reviewing against what is true. That is a system problem, and it has a system answer.

Frequently asked questions

Why does AI content sound generic even when I give it a detailed prompt?

Because a prompt carries a request while a brief carries the business. Adjectives like "friendly but professional" give the model a mood, not facts, so it fills the gap with the average of everything it has read. OpenAI's own guidance separates instructions, which govern behaviour, from knowledge files, which hold the reference material the model draws from, and recommends concrete "when X happens, do Y" instructions over vague description.

Does it actually hurt to publish content that reads as AI-generated?

Yes. In a December 2025 Klaviyo and Datalily study of 8,000 consumers across eight markets, 7% said visible AI-generated marketing makes them trust a brand more while 32% said it makes them trust the brand less. Emplifi data cited by eMarketer found 52% would stop buying after an inauthentic experience and 91% expect brands to disclose AI use in marketing.

What should go into a brand voice brief that most templates leave out?

Proof and offer. Most frameworks brief tone and stop, which is why AI drafts come out bland and occasionally wrong. Include your actual customers, actual results, the claims you are allowed to make and the claims you are not, plus what you sell and what the next step is. Without a proof set, a model will produce plausible numbers, which is the fastest way to fail legal review.

Is tone the most important thing for making AI content feel personal?

No. What makes people feel understood by AI is a clear, relevant answer, a sender that remembers details from past interactions, and a recommendation with a reason attached, with natural tone finishing behind all of them. Relevance and memory outrank tone, and both come from customer records rather than from writing instructions.

How should I judge whether AI-drafted campaigns are working?

On replies, meetings and pipeline rather than open rates. Open rates are a weak signal given privacy proxies that inflate them. Pair that with a claim check on every draft, confirming that every number, integration and capability mentioned exists in the product today.