How to Keep AI-Made Videos On-Brand: Captions, Voice and Intros

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Keep AI-Made Videos On-Brand: Captions, Voice and Intros.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Keep AI-Made Videos On-Brand: Captions, Voice and Intros.

Keep AI-made videos on-brand by controlling the three things AI tools change most: captions (your fonts, colours, spelling and line lengths), voice (one approved voice with a pronunciation list, and consent if it's cloned) and intros and outros (a locked template the AI can't restyle). Put these on a one-page video brand sheet and check every video against it.

AI video drifts off-brand because each tool makes its own defaults look finished: auto-captions in the app's favourite font and spelling, a stock voice that mispronounces your product names, a trendy intro animation. None of these is wrong on its own. Together they make your videos look like everyone else's who uses the same app, and they change whenever the app updates.

Follow me on Instagram@sagnikteaches

The worked example is an illustrative events company that publishes about 12 short videos a month: recaps of weddings and corporate parties, venue walkthroughs and supplier spotlights. It edits with AI tools that cut footage, write captions and generate voiceovers. Examples from a clothing shop, a software reseller, a print shop and a jewellery seller cover problems the events company didn't hit.

Connect on LinkedInSagnik Bhattacharya

What drifts first when AI edits your videos

Before writing rules, look at your last ten videos side by side. The events company did this on a laptop with the sound off, then again with sound, and found the same five elements varying from video to video.

Subscribe on YouTube@codingliquids
ElementWhat the AI tool did by defaultThe on-brand rule it needed
Caption styleBouncy word-by-word captions in the app's default font, a different colour each timeBrand font, white text on a 70% dark box, one or two lines, no animation
Caption spellingThe app's default spelling, and venue and package names mangledThe house spelling set in the tool, a custom vocabulary of names, a human read
VoiceoverThree different stock voices across ten videosOne approved voice, one speed, a pronunciation list
IntroA five-second logo animation from the app's libraryNo intro on social; a one-second logo sting on long videos
MusicWhatever the app suggested, including an AI-generated trackA short list of licensed tracks by mood; AI music only with disclosure

If you make videos with AI tools and haven't yet settled which kinds are worth making, the tutorials on making TikToks and Reels with AI and on whether AI video is good enough for small business marketing cover that first. This one assumes you're making them and want them to look like yours.

Build a one-page video brand sheet

Everything below ends up on one page that whoever edits (a person, or a person steering an AI tool) keeps open. This is the events company's, filled in:

ItemRule
FormatsVertical 9:16 for social; horizontal 16:9 for website and YouTube
Caption font and lookBrand sans-serif, white, on a dark box at 70% opacity; bottom third, clear of platform buttons
Caption lengthOne idea per caption; at most two lines of about 42 characters
SpellingHouse spelling ("colour", "organise", "personalised"); venue, package and supplier names from the name list
VoiceOwner's own recorded voice for long videos; one named stock voice for short ones; speed 1.0
IntroNone on social (open on the best shot); one-second logo sting on long videos
OutroThree-second end card: logo, one line, website address. No "like and subscribe".
ColourBrand warm grade preset; no heavy filters on food or flowers
MusicSix licensed tracks tagged calm, upbeat or formal; no AI-generated music without an AI label
PeopleGuests only with consent; children only with a parent's written permission
DisclosureAny AI-generated footage of a real place, or AI music, gets the platform's AI label

Most editors, including Canva, Adobe Premiere (the renamed Premiere Pro) and DaVinci Resolve, let you save caption styles, colour presets and intro templates. Save them once and select them every time, rather than trusting the AI tool's "auto-brand" to remember.

Captions: your spelling, your line breaks, your safe zones

Captions matter more than most owners assume, because plenty of people scroll social feeds with the sound off, and captions are also how deaf and hard-of-hearing viewers follow along. They're also where AI errors are most visible.

Borrow a professional convention for length

Netflix's English timed text style guide is a useful reference: 42 characters per line, a maximum of two lines, and a reading speed of up to 20 characters per second for adult programmes. You don't need broadcast standards for a 20-second Reel, but those numbers stop the common AI habit of cramming a whole sentence into one caption. The quick sum: a caption on screen for three seconds should hold no more than about 60 characters, which is roughly ten words.

Give the caption tool your vocabulary

Speech-to-text tools mishear names more than anything else. An illustrative example from the events company's autumn recap:

Auto-caption: "Our aurora package includes the festoon canopy and the old dairy courtyard bar, finished by the team from sweet vine catering."

Corrected: "Our Aurora package includes the festoon canopy and the courtyard bar, with dessert by [supplier name]."

Two errors there were names (the package and the supplier) and one was an old venue feature the voiceover had mentioned but the company no longer offers. Many caption tools accept a custom vocabulary or a list of names; add every package, supplier and product name. Then read the captions against the audio anyway, because the tool will still guess at anything new.

Fix spelling at the settings level

Check the caption tool's language setting. An English variant other than the one your brand uses produces spellings that look like typos to your customers (a missing "u" in colour, a "z" where you would write "s"), and correcting them video by video wastes time. Set it once, and add a spot-check to the pre-publish list.

Burned-in or uploaded?

For social cuts, the events company burns captions into the video so its style always shows. For long videos on its website, it uploads a caption file so viewers can switch captions off and so the text can be corrected without re-exporting the video. Either way, keep captions inside the safe area: the bottom corners of vertical video are covered by platform buttons and descriptions.

Voice: one approved voice, a pronunciation list and written consent

Voice is the element viewers notice changing even when they can't say why. There are three sensible options, in the order most small businesses should consider them:

  1. Your own recorded voice. Most trusted, costs nothing but time, and immediately distinctive. A decent USB microphone in a quiet room is enough.
  2. One stock AI voice, always the same one. Consistent and fast. Pick it once, write its name on the brand sheet, and never let an editor choose "whichever sounds nice today".
  3. A cloned voice. Useful when the owner can't record every week. Only clone a voice with the person's written consent, and use a service that verifies it. ElevenLabs, for example, requires a Creator plan or above for its Professional Voice Cloning and checks with a spoken "voice captcha" that the person providing the samples is really speaking; its documentation is clear that responsibility for authorised use rests with the creator.

Whichever you choose, keep a pronunciation list. AI voices guess at names the same way caption tools do. The events company's list, filled in:

WordSay it asWhy it's on the list
Aurora (package)uh-ROAR-uhThe voice stressed the first syllable
CanapésCAN-uh-paysRead as "canapes" with a hard ending
Marqueemar-KEESaid as "MAR-kway" in one test
Our company name[phonetic spelling]Always first on the list

Most AI voice tools let you respell a word phonetically in the script or set a pronunciation dictionary. Put the list in the tool once, and test every new name before a video goes out. For a wider look at what can go wrong with synthetic voices, including scams that use them against businesses, see how to spot deepfake voice and video scams.

Intros, outros and lower thirds as locked templates

AI editors love adding things: a whoosh, a spinning logo, a countdown, a "subscribe" animation. Decide what your intro and outro are, build them once as templates, and treat anything else the tool adds as something to delete.

  • Social videos: no intro. Open on the most striking shot, with the first caption on screen from the first frame. Branding comes from the caption style, the colour grade and a short end card.
  • Long videos: a logo sting of a second or two at most, then straight into content.
  • End card: logo, one line of text and one action. The events company's reads "Planning something? [website]". One action beats three.
  • Lower thirds (the name captions for people on screen): one template, brand font, the person's name and role, on screen for about three seconds.

An illustrative case of why "locked" matters: an AI auto-edit turned footage of a charity gala dinner into a fast-cut montage with an upbeat stock track and a spinning-logo intro. The charity client had asked for something warm and understated. Nothing in the edit was technically wrong, but it looked like a nightclub advert. Because the template and the music list were on the brand sheet, the fix took ten minutes; without them it would have been an argument about taste.

The events company's first month of on-brand AI video

Here is how the rules played out over the first month, with illustrative numbers.

  • Setup (about four hours). Reviewing ten old videos, writing the brand sheet, saving caption styles and colour presets, recording a one-second logo sting, building the name and pronunciation lists, and choosing six licensed music tracks.
  • Per video before the rules: about 50 minutes, of which roughly 20 went on undoing the AI tool's choices (restyling captions, swapping voices, removing intros).
  • Per video after: about 30 minutes, of which 10 is the pre-publish check below. Across 12 videos that's 12 × 20 = 240 minutes saved, four hours a month, which repays the setup in the first month.
  • Errors caught by the check in month one: 7 caption misspellings of names, 2 videos where the tool had switched to a different stock voice after an app update, and 1 AI-generated music track that needed either a disclosure or a swap (it was swapped).

The app update is worth noting. Tool defaults change without warning, which is why "set it once" still needs "check it every time".

Labels and disclosure for AI footage and voices

On-brand also means honest. Platforms and some laws now expect you to label realistic AI content, and getting this wrong damages trust faster than a wrong font.

YouTube's rules are a good model because they are specific. Creators must disclose AI-generated or meaningfully altered content that looks realistic: content that makes a real person appear to say or do something they didn't, alters footage of a real event or place, or generates a realistic scene that didn't happen. You disclose it through the "AI use" setting in the Attributes section of YouTube Studio when uploading. Its disclosure help page lists examples that need a label, including AI-generated music and AI-generated extra footage of a real place, and examples that don't, including caption creation, colour adjustments, production help with scripts or thumbnails, and cloning your own voice for voiceovers.

For the events company, that means: AI-cleaned audio and auto-captions need no label; an AI-generated drone shot of a real venue it never filmed does. If you sell to customers in the EU, the EU AI Act's transparency duties, which have applied since 2 August 2026, include labelling deepfakes, so realistic synthetic video of real people or places needs particular care. The tutorial on labelling AI-generated images and video on social media covers each platform's labels.

One more reason to keep your own files: tools disappear. OpenAI discontinued Sora, closing its web and app experiences in April 2026 and its API in September 2026. Anyone whose videos depended on it had to rebuild. Keep your brand templates, voice scripts, music licences and project files in your own storage, so switching tools is an inconvenience rather than a rebuild.

A pre-publish check for every AI-edited video

About ten minutes per video. Watch once with sound, once without.

  1. Captions in the brand style, inside the safe area, at most two lines each.
  2. Every name spelled correctly, checked against the name list; spelling in your brand's English.
  3. The approved voice, at the approved speed, with every name pronounced correctly.
  4. Intro and outro from the templates; nothing the tool added on its own.
  5. Music from the approved list, or labelled if AI-generated.
  6. Nothing shown or said that you don't offer (old venue features, discontinued packages).
  7. Consent in place for every identifiable guest, and parental permission for any child.
  8. AI label set where the platform requires it.

Where other businesses' videos went off-brand

  • An online clothing shop used an AI voice for product videos that pronounced a fabric name three different ways across a collection. A pronunciation entry fixed it, and the shop added fabric names to its caption vocabulary at the same time.
  • A software reseller tried an AI avatar presenter for a "how to approve your new backup service" video. Customers found it off-putting and several asked whether the email was a scam. It switched to a screen recording with the support lead's real voice, which fitted its brand of plain, trustworthy help. For how-to videos, a real screen and a real voice usually beat a synthetic face.
  • A print shop made process videos of jobs coming off the press. The AI editor's auto-captions labelled a foil-blocking machine as "hot foil blocking" in one video and "foil stamping" in another. Both are used in the trade, but customers read them as two services; the shop picked one term and added it to the vocabulary.
  • A handmade jewellery seller used an AI-generated soundtrack under a making-of video for YouTube. Because YouTube lists AI-generated music as needing disclosure, it either had to set the AI label or swap the track. It chose a licensed track to keep the video about the craft.

If you're weighing up how far to go with AI in video production at all, making marketing videos with AI without being on camera covers the options, and the tutorial on how videographers use AI to edit and caption shows how professionals keep quality up at speed.

AI video branding: follow-up questions

Should captions be burned into the video or uploaded as a file?

Burned-in captions always show and keep your fonts and colours, which suits short social videos watched without sound. Uploaded caption files, such as SRT, can be switched on or off by viewers and edited later without re-exporting, which suits longer videos on your website or YouTube. Many businesses do both: burned-in for social cuts, a file for the long version.

Is it fine to clone my own voice for voiceovers?

It can be, with care. Use a tool that verifies the voice belongs to the person providing it, keep the samples and the voice model under your own account, and decide in advance what the cloned voice may and may not be used for. YouTube lists cloning your own voice for voiceovers as something that doesn't need its AI disclosure.

Do I need a video intro at all?

For short social videos, usually not; a branded caption style and an end card do the job without delaying the first second. For longer videos on your website or YouTube, a short intro of a second or two helps viewers recognise your content. Whatever you choose, make it the same every time.

What if a client wants AI-generated footage of their venue or product?

Agree in writing what will be generated and how it will be labelled. If the footage shows a real place or product doing something that didn't happen, platforms such as YouTube expect a disclosure, and customers who later visit will judge you on the difference between the video and reality.

Further reads

Sources: YouTube Help (Disclosing use of GenAI content); Netflix partner help pages (English timed text style guide: characters per line, lines, reading speed); ElevenLabs documentation (voice cloning, verification and plan requirements); series fact sheet for the Sora discontinuation, EU AI Act Article 50 transparency duties and Adobe Premiere naming (checked September 2026).

Want your AI video workflow set up on-brand?

On a 1:1 call we'll look at the videos you make and the tools you use, build your video brand sheet and templates, and set up a check that fits how often you publish.

Book a 1:1 call with me