AI influencersInstagram·October 13, 2026·5 min read
How to create an AI influencer: what the job actually is
Making one beautiful picture of a person who does not exist is now within anyone's reach. A million-follower account that sells something does not follow from it. The distance between the two is measured in post counts, in how consistent the character stays, and in how many platform refusals you survive. Below are four working examples and our own invoices.
This is not a "top 10 virtual models" listicle. These are four profiles opened on 13 October 2026. Everything below is visible in the profiles themselves: post count, follower count, bio, link. What is not visible there is not claimed here.
Four profiles, captured on 13 October 2026. Everything discussed below is visible right here: the post count, the follower count, the AI-generated label and what sits in the link.
@lilmiquela, 2.2MThe oldest living example. 1,500 posts, a bio that reads "23 / L.A / Robot", a music release in the link. The character has been run for years, and the volume shows it.
@mother.satori, 1.1MThe health niche: "US based Holistic Expert" in the bio, a shop for her own bitters in the link. 1,173 posts, nearly all video. She sells a product, not ad slots.
@naina_avtr, 348K"Actor, Influencer, Fashionista", two 2024 industry awards in the bio, links to YouTube Shorts. Lifestyle and fitness, which is an audience rather than a product.
@kyraonig, 247KFewest followers and fewest posts: 138. But the pinned highlights carry brand and TV-show names, and the bio lists an email for collabs outright.
lilmiquela1,500 posts
2.2M
mother.satori1,173 posts
1.1M
naina_avtr730 posts
348K
kyraonig138 posts
247K
Followers and the posts behind them, as of 13 October 2026. Look at the second row: 1,173 posts for 1.1 million. That is not one generation, it is a production line that ran for months.
The picture is not the product. Consistency of the character and sheer volume are — the things a human creator gets from living their life, and an AI character has to manufacture every single day.
How many posts sit behind a million
Between 138 and 1,500. The low end is not "nailed it on the first try": Kyra, at 138 posts and 247K followers, carries noticeably more brand highlights than the million-follower accounts, which means her money arrives through collaborations rather than reach. The two million-follower accounts have over a thousand posts each behind them.
The practical conclusion is simple and unwelcome. An AI character removes the constraint of having to look, travel and film yourself. It does not remove the volume constraint. You need a stream of posts in which the face, the voice and the manner repeat every time. So the first thing worth testing is not "will it render beautifully" but "will the same face come back tomorrow".
What it costs in models
Our numbers now, not estimates from someone else's roundup. We keep our own ledger of model calls, and this is what it says.
Text is almost free.One post rewritten in a persona's voice costs us about $0.0096, measured from our own call table. A thousand posts is under ten dollars.
Video is three orders of magnitude dearer.By the providers' published rates at the time of writing: video from a reference runs $0.05 to $0.16 per second, replacing the person in frame runs $0.30 per second at 768p and $0.45 at 1080p. A five-second 1080p clip is about $2.25.
The invoice beats the button.Our real month came in at $18.26 against the $2.00 the dashboard showed; three operations had an understated rate in the interface, and a voice design cost $3.00 where the button promised $1.50.
The lesson that cost us sixteen dollars of difference: count from the invoice, not from the price list. How we measure the cost of a single call, and why it does not match a token estimate, is in its own article.
The money in this niche does not go into one beautiful model. It goes into repeatability: a hundred clips with the same face and the same voice in each. Which makes cost per second more important than cost per clip, and makes the choice of model something other than a matter of taste.
What breaks first
We walked into every one of these building the same pipelines inside Sndwich. Not one of them is about generation quality.
The face driftsevery clip is a slightly different person
The voice shiftsthe dub does not match the last one
The shape is offthe clip cannot be promoted
The account stopsaction blocked
Face and voiceSeparate generations give you four different people. Keeping one takes deliberate work, and the how is in our piece on face and voice.
Aspect ratioInstagram only promotes an exact 9:16. One pixel off and the boost button is unavailable; checking takes ten seconds.
Action limitsA fresh account that starts posting and replying at full speed gets limited on the pattern it shows, not on a counter it crossed.
Provider moderationPublic video services reject some frames as suggestive, including a swimsuit in a fitness clip. That is solved with your own hardware, not with argument.
The AI labelOf the four profiles above, three carry the AI-generated label; the million-follower one does not, and it has no verification badge either. Platform rules here move faster than the articles about them.
AI labelVerifiedlilmiquelayesyesnaina_avtryesyeskyraonigyesyesmother.satorinono
Same sample, two columns straight from the profiles. The account selling a physical product to a million people runs without the AI label and without verification. Not advice to copy, an observation about how unevenly the rule lands.
What the process actually looks like
Strip the romance and a working AI influencer is four repeating steps. As dull as any production line.
Voice first, face second.A description of the character plus two or three dozen sample lines buys more consistency than the best single image: every later post is written from them, in the same manner.
Fix the look once.One reference image of the character that every later frame derives from. Not "generate it again", but "take the same one".
Plan a week, do not post a day.Posts are written in batches and placed on a calendar. Otherwise the volume you saw in the numbers above never accumulates.
Replies are a separate thing.Comments and replies are a different voice and a different risk. In our product they switch on with a separate approval of the persona, and until then the replier answers in a neutral voice.
The fourth step is the one most often underrated: followers arrive for a conversation, not for a picture. How replies work as a growth channel in their own right we covered using X as the example.
Short checklist
Test whether the face repeats before judging how pretty it is
Count cost per second of video, not per clip
Reconcile the provider invoice against what the button said
Export exactly 1080x1920 or there will be no promotion
Accumulate volume with a plan, not with inspiration
Turn replies on separately and deliberately
This pipeline is the product
Sndwich was built around consistency: a persona with real examples, one look across every frame, a publishing plan and a separate track for replies. The numbers in this article come from our own ledger of calls.
The persona holds the voice. Sample lines, anti-examples and per-channel styles, not one prompt for everything.
Video with the same face. A reference frame of the character, wardrobe changes and a dub from the same voice.
Every call's cost is recorded. Task, model and price at the moment of the call, instead of guessing from a rate card.
11 minHow to create an AI model for Instagram and TikTok: face swap video, voice clone, first clip step by step
6 minAspect ratio is not valid for Reels ads: what Instagram actually measures
5 minLLM cost per request: what 204 production calls actually cost