Which steps AI really helps with
Making a video or a post involves many separate jobs, and AI is not equally good at all of them. People who end up disappointed often asked the wrong thing at the wrong step.
| Step | How AI does | Caution |
| Ideas and angles | Very good | It gives general ideas. You must add the local subject yourself |
| Script structure | Very good | It will invent facts. Verify them yourself |
| Bangla writing style | Medium | It smells of translation. It always needs polishing |
| Titles and descriptions | Good | Ask for ten and pick the best one yourself |
| Thumbnail background | Good | Do not let it write Bangla text inside the image |
| Voice over | Good in English, medium in Bangla | In a long video your own voice builds more trust |
| Subtitles | Saves time | More mistakes in Bangla, fix them by hand |
| Making video clips | Good for a few seconds | It cannot keep continuity or realism |
| Editing | Good | Cutting silence and pulling out short clips saves time |
Sabbir's numbers: where the time really goes
Say Sabbir runs a YouTube channel from Dhaka and posts 8 videos a month. He sat down and worked out where the time on one video goes.
| Task | Time before | After using AI | Comment |
| Research and verifying facts | 2 hours | 2 hours | Not less. The need to verify has grown |
| Script structure and draft | 2 hours | 45 minutes | The biggest saving is here |
| Recording | 1 hour | 1 hour | In his own voice, so the same |
| Editing | 3 hours | 2 hours 15 minutes | Time saved on cutting silence and pulling out clips |
| Thumbnail, title, description | 1 hour | 30 minutes | Background by AI, Bangla text by himself |
| Total | 9 hours | 6 hours 30 minutes | 2.5 hours per video |
8 videos a month means 20 hours. Sabbir's real decision was about those 20 hours. He did not increase the number of videos. He spent the time on research and on answering comments.
The reason comes from the numbers. More videos does not surely mean more views. But good information and a relationship with viewers are the only two real assets of his channel, and AI has no role in either.
This decision is the main point of this post. AI will save your time, but what you do with the saved time decides whether you move forward.
Ideas and scripts
The most useful way to use it is to turn your own viewers' questions into raw material. Copy the last month of comments, give them to AI, and say: pull out the 15 real questions of my viewers from these, then give one video title for each.
For scripts the rule fits in one line: AI for the structure, you for the facts. Ask AI for the outline of a 7 minute video on this subject, what to say in the first 15 seconds, where to give an example. But every number, price, date and rule inside must be verified by you. The damage from one wrong price in the comments will not be made up by ten videos.
And write the intro yourself. Viewers decide in the first two lines whether to stay, and that is where AI writing gets caught the most.
A script structure, step by step
Sabbir follows these six steps in every video. He gives AI the structure, then fills every slot with his own facts.
- The first two lines, write them yourself. Say the viewer's problem in their own language. Do not give an introduction. If you start with "let's find out", half the viewers will leave.
- Say what they will get. By the end of this video you will be able to do this thing yourself. With this one line, people stay.
- Three to five main points. More than that and viewers will remember nothing. Ask AI for five, then cut it down to three yourself.
- One real example for each point. Your own experience, your own area, a viewer's question. AI cannot fill this part, and this is where the video becomes yours.
- Say where things can go wrong. What will not work, where money will be wasted. This honesty is what brings viewers back.
- End with one task. Try this one thing today. It works better than asking for a subscribe at the end.
Once you give AI the structure, for every new subject it will draft in the same mould. Your job is then only to fill in and cut.
How to spot the smell of AI in Bangla writing
Viewers cannot explain the word, but they will sense it. A few easy signs:
- Everything is one size. Every paragraph is three lines, every sentence the same length. People do not talk like this.
- Book language. Stiff formal words, such as "the said", "is provided", "is observed", "plays a very important role".
- Cliche openings. "Let's find out", "in today's fast changing world".
- Covering every side. On one hand it is good, on the other hand it is not bad either. No clear opinion.
- Nothing specific. No name, no price, no place, no event.
- List disease. Everything set in bullet points, yet people do not speak in bullets when they talk.
The fix comes out of the same list. Put in the name of your area, the story of one buyer, one real number, one clear opinion. And read the text aloud once. A sentence that feels awkward to say will also grate on the viewer's ear.
Sabbir's rule is even simpler. He always throws away the first paragraph of the draft, because AI's first paragraph is almost always an introduction, and nobody listens to an introduction.
Thumbnails
The most useful thing AI does for thumbnails is backgrounds and concepts. Complex scenes, colour contrast, pictures matched to the subject, you get these in a minute.
But there is one thing you should never let AI do: Bangla text inside the image. The image making Models still break Bangla letters, especially conjuncts and vowel marks. It looks like Bangla, but when you try to read it, it means nothing. You should put the Bangla text in yourself, in any ordinary design tool, choosing the font.
The workflow goes like this. One, make the background with AI, without text. Two, cut out your own face photo separately and place it. Three, type the Bangla text in the design tool, no more than three or four words. Four, look at it small on your phone, because most viewers will see it at that size. If it cannot be read small, being beautiful large is no use.
Another real point: in Bangla content, a real human face, especially your own, will work better than a perfect face made by AI. Viewers recognise a familiar face, and that recognition is the reason for the click.
Voice over and the real state of Bangla voices
In English, AI voices are now so good that they often cannot be caught. In Bangla the state is not the same. There are few voices, and problems remain in several places:
- The tone of question sentences often does not come out right.
- Numbers, dates and money amounts are read in a strange way.
- When there are English words in the middle of Bangla, the pronunciation stumbles, yet this is exactly how we talk.
- Little emotion. It sounds flat, and in a long video it feels tiring.
- Pauses fall in the wrong places, so the meaning can change.
The practical decision: AI voice is fine for a short intro, an ad line or an announcement of a few seconds. But in a video where viewers will trust you and learn something, your own voice is best. A quiet room and a fairly decent microphone, even a phone mic, will build more trust than the most expensive AI voice.
A word for those who feel shy about their own voice. Viewers do not look for a beautiful voice. They look for someone who knows the thing. A regional accent is not a weakness, it is what will set you apart. What you must get right is the sound quality: reduce the echo in the room and keep the mic close to your mouth. These two things do half the work.
Subtitles and translation
Here AI's gain is the clearest. On Facebook most people start watching videos without sound, so having subtitles means more people will watch.
The workflow is simple. Make the auto captions, then read and fix them yourself. Bangla will have more mistakes than English, more still with a regional accent, and the most when you mix Bangla and English as you speak. Even so, correcting mistakes is much faster than typing from zero.
You will be faster if you know where mistakes happen most. People's names, place names, organisation names, numbers and money amounts, and technical English words. If you check these five places first, the rest is usually fine.
Translation is a separate opportunity. If you add English subtitles to a Bangla video, you will reach diaspora viewers and viewers who do not know Bangla. Just check the names, places and technical words yourself once.
Music and sound, which almost everyone forgets
Half of a video is really sound, yet the talk is almost always about pictures.
The biggest trouble comes with background music. A copyright claim on a song on YouTube, the ad money going to someone else, even the video being blocked: these are the most common shocks for new creators. If you use a film or a popular song, it is almost certain.
AI-made music can be one workable answer here, because it is not a published song. But before you use it, check whether the tool's terms allow commercial use, just as I asked you to do for images. And remember that a claim over music comes from the platform's automatic system, so you need to have proof.
Two more things on sound quality. One, keep the background music volume well below the speech, because on a phone's small speaker music covers the speech. Two, AI tools for cleaning sound are useful now, they cut down fan noise or street noise quite a bit. But this is no substitute for recording. A bad recording will never become good, only less bad.
Sabbir's rule is simple: fan off and window shut before recording, and record ten seconds of silence and keep it. Those ten seconds are used later for cleaning the sound.
AI video: what it can do and what it cannot
This is the part where the most exaggerated hype is spread. See the real state in one table.
| What you want to do | State today | Practical decision |
| A few seconds of nature or abstract scenes | Good | Use freely |
| b-roll and backgrounds | Good | Saves both time and money |
| Light movement in a still picture | Good | Wonderful for bringing old photos or graphics to life |
| The same character across several clips | Weak | Clothes and face will change, do not try to tell a story |
| Hands, crowds and text on signboards | Weak | Avoid close shots |
| Bangla lip sync | Weak | Do not use in scenes that show a face |
| A whole video | Almost impossible | Make material, not a video |
Keep the cost side in mind too. Prices are usually counted per second, and getting the clip you want takes ten or twelve tries. Making a full video this way will make the bill grow fast, and the result will not be consistent.
Put plainly, AI video is now a machine for making material, not a machine for making whole videos. Channels that stand entirely on AI video may have views, but they do not build viewer trust.
When a GIF is enough, not a video
Remember the last row of the table: a whole video still cannot be made, material can. And in some places you do not need a video at all. Putting a video in a Facebook comment is a hassle, heavy files do not want to go through WhatsApp, and in email a video often does not even open. In these three places, dropping in a GIF of a few seconds does the job, because it plays by itself and does not wait for anyone to press play.
A GIF has no sound and fewer colours come through. So it is wonderful for small things like a logo appearing, an "open" sign blinking on and off, or a discount badge, and when there is a story or speech, you will need a video.
The question of telling viewers
Whether you must tell viewers can be made easy to answer with one rule. If a viewer could wrongly think this really happened, tell them.
| What you used | Could a viewer think it is real | Need to tell |
| A drawn picture, cartoon, abstract background | No | Not needed |
| A script written with AI, edited by you | No | Not needed, but no harm in saying it |
| An AI voice for an announcement or intro | Maybe | Better to say it, especially in a long video |
| A realistic scene or event | Yes | You must tell |
| Any person's face or voice | Yes | You need permission, and you must tell |
| An AI image as a product photo | Yes | No question of telling, do not do this |
The big platforms now also ask at upload whether the content has realistic altered or synthetic content, and show a label accordingly. Give the true answer in that box. For sponsored work, ask first, because many brands have their own rules.
Many people have one fear: will reach fall if I tell? The honest answer is that the loss from being caught hiding it is much bigger. And a viewer who finds out that you hid it will not come back.
Things you must never do
- Another person's voice or face. Not a leader, an actor, a journalist, a neighbour, anyone. Nobody is unaware of the kind of harm fake audio and video cause in Bangladesh.
- Fake news or pictures of accidents. One AI picture becomes a rumour in a moment, and it cannot be brought back.
- Fake reviews or testimonials. Making a face made by AI praise a product is plain fraud.
- Health and legal advice straight from AI writing. Here the price of a mistake is people's health or money.
- Taking another creator's video and turning it into your own. You will lose copyright for sure, and the platform will also mark you for repetitive, low effort content, which is a risk for monetisation.
- Content made with children's faces or voices. The harm here is the biggest, and there is no way to take it back.
When not to use AI
The list above is ethical, this list is practical. In these places AI will not do harm, but it will make your work worse.
- Where your own experience is the main thing. Personal stories, your failures, replies to a viewer's letter. If AI writes it, it becomes ordinary, and nobody remembers ordinary things.
- Where time is short and you cannot verify. If you use AI's facts in a rush, the mistake will go out, and it cannot be taken back.
- Very small jobs. For a one line caption, opening a tool and writing a Prompt takes more time than writing it yourself.
- A subject you know well. Your own writing will be better, and faster too. Use AI where you get stuck.
- Where the brand's voice matters. In sponsor or client work they have their own language rules, and writing it yourself is easier than explaining that to AI.
- When you are tired. In that state you will not read the draft, you will use it directly. This is how wrong facts get published.
A workflow, start to finish
- Pull out questions from viewer comments, make 15 titles with AI, pick one.
- Gather the facts on the subject yourself, with sources. Do not give this step to AI.
- Have AI make the structure, then put in the facts yourself and verify them.
- Write the intro yourself, polish the rest of the AI draft into your own language, and read it aloud.
- Throw away the first paragraph, because almost always it is a useless introduction.
- Record in your own voice, in a quiet room, with the mic close to your mouth.
- In editing, cut the silence with an auto tool, then fix by hand.
- Make the thumbnail background with AI, put in the Bangla text yourself, and check it small on a phone.
- Make auto subtitles and fix them yourself, checking names and numbers first.
- If needed, add English translation too, for diaspora viewers.
- At upload, give the true answer to the box about AI content.
- Pull three short clips from the long video, and have AI do the descriptions and hashtags.
For sponsors it will be easier if you have a page of your own, with your viewer numbers, past work and contact. There are now tools that let you build such a page by chatting in Bangla without knowing code, for example BanglaCodes.
Common mistakes and fixes
- Mistake: reading the AI draft almost word for word. Fix: read it aloud. Cut any sentence your mouth stumbles on. This one habit will remove half the problems.
- Mistake: publishing without verifying facts. Fix: in the script, write the source next to every number and date. If you cannot find a source, drop the line.
- Mistake: making the thumbnail text with AI. Fix: background by AI, text by you. Bangla letters will break.
- Mistake: AI voice in a long video. Fix: use your own voice. Reducing room echo and keeping the mic close is enough.
- Mistake: publishing auto subtitles without looking. Fix: check names, places and numbers first. The rest is usually fine.
- Mistake: making more videos the goal. Fix: spend the saved time on research and on talking with viewers. This is Sabbir's decision.
- Mistake: testing a new AI tool in every video. Fix: fix one set for three months, and only then think of changing.
- Mistake: making something realistic and staying quiet. Fix: tell them. The cost of being caught is far bigger than the cost of telling.
To make one yourself, open the video making page or the video script making page.
Questions and answers
Can I run a channel by making whole videos with AI? Technically yes, but it will not last. Platforms discourage repetitive, low effort content, and viewers now spot it too. Make AI your assistant, not your presenter.
If I make a Bangla video with an AI voice, will viewers notice? In most cases they will, especially in a long video. In a short clip they may not. The question is not only about being caught, it is about trust, because the voice is a big part of personality in Bangla content.
Why is AI not able to write Bangla text in a thumbnail? Image making Models do not see letters as letters, they see them as shapes, and Bangla conjuncts and vowel marks are complex. So the text looks like Bangla but cannot be read. There is only one fix: you put the text in yourself.
Does YouTube punish you for using AI in content? Using AI is not an offence in itself, and many big channels use it. Trouble comes in two places: posting realistic content without telling, and mass producing the same thing and uploading it. Keep it original and honest and there is no trouble.
If it is a new channel, is it smart to start with AI? At the start, use AI for research and organising, not for voice and personality. The biggest problem of a new channel is that viewers do not know you, and AI does not solve that. Keep your own face and your own voice in the first 20 videos and that recognition will build.
What if a viewer asks in the comments whether I used AI? Tell them straight, which part you did with it and which part you did not. Most viewers do not object to the use of AI, they object to the hiding. An open answer actually builds trust.
Are the rules the same for Facebook Reels and YouTube? The core policy is the same, but every platform has its own rules and label system, and they change. On the platform where you post regularly, read the creator policy page once a year.
How much time will AI really save? Across the script, subtitles, description and thumbnail draft, it usually saves one to three hours per video. But the time for recording, editing and verifying will not shrink, and that is the real work.
What should I do with the saved time? This is the real question, and the least asked. Two things give better results than making more videos: deepening the facts and talking with viewers. Sabbir put those 20 hours into exactly these two, and that is what set his channel apart.
In short
- AI helps at almost every step of content, but the biggest gain is in ideas, script structure, subtitles and translation.
- The smell of AI translation gets caught in Bangla writing, so put in your own area's names, real numbers and a clear opinion to make the writing your own.
- AI images break Bangla letters, so make the thumbnail background with AI and put the Bangla text in yourself.
- A Bangla AI voice is fine for short clips, but in a long video your own voice builds more trust, and auto subtitles must always be fixed by hand.
- AI video is now a machine for making material, not for making whole videos, because problems remain with continuity, hands, text and Bangla lip sync.
- If you make something a viewer could wrongly take as a real event, tell them, and never use anyone's voice or face without permission.