An AI influencer is not an AI video with a person in it. It is a persistent identity: same name, face, voice and opinions across hundreds of posts, building a relationship with an audience. The video is easy. The identity is the work.
Influencer or presenter?#
Three things separate a persona from a presenter:
- Continuity. Post 1 and post 200 show the same, recognisable person.
- Point of view. The character holds opinions and would refuse some scripts.
- Accumulation. Post 40 is worth more than post 4 because of what came before.
If you only need someone to say a claim in an ad, you need a presenter, not a persona. See the AI UGC generator guide for that route.
Step 1: write a one-page spec#
Write this before you generate anything. One page; a long character bible does not survive a production schedule.
- Name. Short, easy to say and search, and not a real person’s.
- Age. Specific. “Late twenties” is a decision; “young” is not.
- Role. A job or role that gives them a reason to talk about your category.
- Setting. A kitchen, desk, car or gym. A fixed setting builds recognition.
- Stance. One sentence of opinion on the category.
- Five phrases they use.
- Three things they would never say. This stops drift when several people write for them.
Step 2: name and description#
The description is an input to every script, so vague in means vague out.
Bad: “Friendly and relatable, loves our product.”
Better: “29, works front of house at a busy restaurant, has tried every scheduling app and hates them all, talks fast, shows rather than explains.”
In Looptake an AI influencer has a name, description, niche, look and voice. Fill every field. A blank field is a guess.
Step 3: the reference image#
The image is what continuity is made of. Every video pulls from it.
- Neutral expression, not mid-smile or mid-word.
- Even, ordinary light. Studio light reads as an ad.
- Head and shoulders, facing forward.
- A background you can live with for a year, or a plain one.
Never upload a photo of a real person you have not cleared. Once you choose the image, freeze it. A new image is a new character wearing the old name.
Step 4: preview and check#
Generate a preview before you commit. Check:
- Does it look like the same person from different angles?
- Does it read as the age you set? Generated faces often read younger.
- Would your buyer take a recommendation from this person? That is not the same as attractive.
If any answer is no, change the image and preview again.
Step 5: voice#
In short-form, voice carries identity more than the face, because viewers hear the second sentence before they study the face. Match pace to the character. Generate in the audience’s language rather than translating afterwards. Then freeze the voice too.
Step 6: the first scripts#
Write the first ones yourself to calibrate. Use this shape:
- Open in the middle. No self-introduction. Nobody cares yet.
- One claim.
- One concession. Something true and mildly negative. It makes the persona sound like a person with a view.
- Close on the situation, not the product.
Aim for 60 to 110 words for a 20 to 30 second read. Read it aloud at the character’s pace. More shapes in short-form script templates.
Step 7: own account or brand account#
Decide now. You cannot move a back catalogue later.
- Own account. The persona is a creator. Better follow behaviour, longer path to purchase, and a disclosure duty when your product appears.
- Brand account. The persona is your presenter. Shorter path, lower ceiling on follows.
Do not split a new cadence across both. A full cadence on one account beats half on two. Do not run a persona to hide who you are. In Looptake a persona can have its own connected accounts, kept separate from the brand’s on the calendar and in analytics.
Step 8: the first ten videos#
Make ten in one sitting before you publish any.
- Same setting, voice and framing in all ten.
- Ten subjects from one source. Real support questions work well.
- Read all ten together. Drift is invisible one by one and obvious in a batch. Fix it in the description.
- Schedule them over two weeks, not two days.
Then leave the character alone for thirty videos. Each change to name, look, voice or format resets recognition. With auto-post to TikTok and Instagram, the schedule runs without daily attention.
Cadence#
Talking video is the most expensive format to render, so a daily persona-only schedule is the wrong design. Let the persona appear two or three times a week on posts that need a voice, and carry the daily floor with cheaper formats. Credits pay only for AI media; scripts, editing, scheduling and posting are free (pricing).
Rules you inherit#
Can: have opinions, recommend your product, appear in paid ads you own, speak in any supported language.
Cannot: be presented as a real named person, a real customer or an independent reviewer; claim results that did not happen; use a real person’s face or voice without permission.
Two disclosure duties apply together. Platforms label synthetic media and you must follow their current rule. Advertising law requires you to disclose the relationship when the persona promotes your product, the same as for a human creator. Put both where viewers see them, in the persona’s own voice.
Do not fake a life#
Give the persona a personality, not a biography. “Works front of house at a busy restaurant” is a character. “Sarah, 29, from Leeds, verified buyer since March” is a fake customer. A fake backstory breaks platform rules and advertising law, and when it is found, every post is reframed as a lie. An openly synthetic persona has a ceiling. A lying one has a cliff.
Signs it is working#
After thirty posts, look for:
- Comments that name the character’s traits, not the handle.
- Return viewers across several posts.
- Questions asked of the character, not about the product.
- Profile visits per post that hold steady as volume rises.
If none appear, change the image and description and run another thirty, or switch to a rotating presenter roster. Rebuilding early is cheap; rebuilding after two hundred videos throws away the only thing a persona builds.