The Dubai business owner we're onboarding right now asked me exactly one question before anything else: "What do I actually have to do?" Fair question. The honest answer — one recording session and a weekly 20-minute approval — sounded so thin he made me walk him through the whole process, week by week. This post is that walkthrough.
Key Takeaways
- You film once. A single 60–90 minute capture session trains the face, expressions, lipsync and voice. Then you're done filming.
- Four weeks to first delivery. Capture → build → script system → publish-ready videos. Anyone promising 48 hours is skipping the review loop.
- The review loop is where realism lives. Test takes, your feedback, refinement — until strangers can't tell in-feed.
- Your ongoing time: ~20 min/week. Script approvals. Production, editing and formatting all happen without you.
- Judge providers by a live clone, not a showreel. Ours is public: @amanrai.official.
Week 1: the capture session — what you record, once
Everything downstream depends on this session, so it's worth doing properly. You'll record 60–90 minutes of footage, guided by a recording sheet we send in advance. It has two halves:
- Scripted reading — you read provided lines straight to camera. This gives the model clean phoneme-to-mouth data, which is what makes lipsync frame-accurate later.
- Free talking — you answer casual prompts about your business, off the cuff. This is where the model learns your real gestures, your pauses, the way your eyebrows move when you disagree with something.
The technical bar is lower than people expect. Clean 1080p works; 4K is better. What actually matters: even light on your face (a window or one softbox), a quiet room, and no autofocus hunting. Phone on a tripod beats a shaky DSLR. We've rejected footage shot on ₹3 lakh cameras and approved footage shot on an iPhone against a bedroom wall — light and stability decide it, not the gear.
The voice sample
Inside the same session you'll record 2–3 minutes of clean speech for voice training. Read naturally, at your normal pace, with your normal accent. Don't perform. The model needs how you actually sound on a call, because that's what your audience will compare it against. Once trained, that voice can deliver scripts in languages you don't even speak — Hindi and English from one Indian founder's build, or English and Arabic for a Gulf audience.
What you'll need to provide — the full checklist
This is everything we ask a client for, with the honest time cost of each item:
| You provide | What it's for | Your time |
|---|---|---|
| Capture session footage | Trains face, expressions, lipsync | 60–90 min, once |
| Voice sample (2–3 min, clean) | Trains your voice model | Inside the same session |
| Brand kit — logo, colors, fonts | On-brand scenes, captions, end cards | 10 min to send |
| 5–10 topics you have opinions on | Seeds the script bank in your voice | 15 min voice note |
| One voice-mining call | We learn your phrasing and takes | 30 min, once |
| Channel access (optional) | Only if we publish for you | 10 min |
| Weekly script approvals | Nothing publishes without your yes | ~20 min/week |
Total hands-on time in month one: about three hours. Here's how those minutes actually distribute — note what's missing from the chart: any production, editing or publishing time at all.
Week 2: clone build and the review loop
Now the work moves to our side. The footage trains the visual model, the sample trains the voice model, and we assemble a scene kit — the branded environments your clone will appear in, so every video doesn't look like the same grey booth.
Then comes the part cheap providers skip: the review loop. We generate test takes and put them through the only test that matters — does it survive a phone screen, in-feed, viewed by someone who doesn't know? You review the takes too. Skin texture, gaze, the hands, the way the head turns. We refine and regenerate until it passes. On our own founder's clone this loop ran multiple rounds before anything went public, and it's why @amanrai.official can run daily without anyone calling it out.
A clone is built in the review loop, not in the render button. Skip the loop and you get the dead-eyed junk flooding LinkedIn right now.
Week 3: the script system — your voice on paper first
A perfect clone reading generic AI filler is still a failure. So before production starts, we build the script system: a 30-minute voice-mining call where we pull out your phrasing, your strong takes, the phrases you'd never say. That becomes a script bank — hooks, stories, opinions — written the way you talk, not the way a language model talks.
Then the approval gate. Every script goes to you before production. You'll kill some. Good — that's the system working. My rule for our own account: if a script couldn't start a small argument in the comments, it doesn't get produced. Agreeable content is invisible content.
Week 4: production, editing and publish-ready delivery
Approved scripts go into batch production — a week or a month of videos generated in one pass through your clone. Then editing turns raw generations into finished posts: cut to length, captioned (most viewers watch muted), music bedded, b-roll where it earns its place, and exported in 9:16, 1:1 and 16:9 so each platform gets its native format.
What lands in your folder is genuinely upload-ready: named files, thumbnails, caption text, posting schedule. Or — on plans where we publish — it just appears on your channels and you find out from the comments. Pricing for both routes is on the AI clone service page, and the deeper cost math is in how much an AI clone costs.
After week 4: the weekly cadence
Once the clone is live, the rhythm is boringly reliable. Monday: scripts land in your inbox. You spend 20 minutes approving, killing, or tweaking. Friday: edited videos are delivered or published. Once a quarter we suggest a 30-minute refresh session so the clone keeps up with your haircut. That's the entire operating model — and the speed difference against a traditional shoot pipeline is the reason it works:
Red flags of a bad clone provider
The clone market in 2026 is a mess, and I say that as someone selling into it. Before you sign with anyone — us included — run these checks:
| Red flag | What it actually tells you |
|---|---|
| No live public clone of their own | They can't clear their own quality bar. Ask for a URL, not a demo reel — reels are cherry-picked. |
| "Ready in 48 hours" | No review loop. You'll get take one, dead eyes included. |
| No revision rounds in the price | The realism work happens in revisions. If they're not priced in, they won't happen. |
| "Just send us old footage" | A clone trained on scraps looks like scraps. A real provider insists on a proper capture session. |
| Generic text-to-speech voice | The sound gives it away before the face does. Your voice must be trained on your voice. |
| Won't say who writes scripts | You'll get AI filler in your mouth — the fastest way to torch a personal brand. |
One more filter that settles it fast: ask to see their clone next to their real footage and pick which is which. If they hesitate, walk. We've made this test easy on ourselves by publishing daily — the comparison between Aman Rai on a real call and the clone on Instagram is public and permanent. And if you're weighing doing all this yourself with HeyGen or Synthesia instead, the honest side-by-side is in AI clone service vs DIY tools — short version: the tools are fine, the 15 hours a week you'd spend running them isn't.
WANT THE PROCESS RUN FOR YOU?
One capture session, four weeks, then your face posts weekly while you run the business. We'll scope cadence, languages and price on a single call — ₹, AED, $ and £ all fine.
Start on WhatsApp See the Full ServiceFAQ: AI clone creation
How long does it take to create an AI clone?
About four weeks from capture to first publish-ready videos: capture and voice training in week 1, build and review loop in week 2, script system in week 3, first edited batch by week 4. Providers promising 48 hours are skipping the review loop — which is where the realism comes from.
What do I need to record for an AI clone?
One 60–90 minute session: scripted reading plus free talking, in even light and a quiet room, at clean 1080p or better, plus a 2–3 minute voice sample. A phone on a tripod near a window genuinely works. That's the only filming you do.
Can my clone speak languages I don't speak?
Yes. The trained voice model can deliver scripts in multiple languages with your tone — Hindi and English, or English and Arabic, from one build. Lipsync regenerates per language so the mouth matches the words in each version. We deliver in up to 8 languages.
Do I have to keep filming after the clone is built?
No. After the capture session, every video generates through the clone. Your ongoing role is ~20 minutes a week approving scripts, plus an optional 30-minute refresh once a quarter for haircuts and wardrobe drift.
How do I know the clone won't look obviously fake?
Judge the provider's own live clone before you pay — ours runs daily at @amanrai.official. During your build, the review loop iterates on lipsync, gaze and gesture until strangers scrolling on a phone can't tell. No live clone plus no revision rounds equals fake-looking output, every time.
What's the difference between an AI clone and a stock avatar?
A stock avatar is a licensed pre-made face — instant, but nobody builds trust with a rented face. An AI clone is trained on you: your face, voice, gestures and scenes. That's why clones need a capture session and a build loop, and why they're the only option that grows a personal brand.