Question 1

How are founders making product demos, ads, and social video with AI (Veo, Runway, Kling, HeyGen, ElevenLabs)?

Accepted Answer

The playbook now: write the script yourself, have Gemini or ChatGPT turn it into a shot-by-shot prompt list, generate clips with Veo 3.1 or Gemini Omni Flash inside Google Flow, Runway, or Kling 4.0, then cut it together in Premiere, Descript or DaVinci Resolve with an ElevenLabs voiceover. Omni lets you fix a clip by describing the change instead of generating it again. The reference case is Kalshi's NBA Finals ad, made by one AI filmmaker in three days for about $2,000. For SaaS product demos, founders skip generation and pair real screen captures with AI demo tools like Supademo, since showing the actual product beats synthesizing it.

Question 2

Why can a solo founder now do what needed an agency?

Accepted Answer

Because every line an agency billed for has collapsed into software: video models replace the shoot, avatars and voice clones replace talent, image models replace the studio, and AI editors replace post-production, all for tens of dollars a month. The proof points are real: a national TV ad made solo in three days for $2,000, and creators like Varun Mayya shipping 1,000+ videos a month with a tiny team. The constraint has moved from budget to taste; ideas, scripts, and judgment are now the scarce inputs, which is exactly where a founder has the advantage.

Question 3

What's the realistic quality bar, what works and what still looks fake?

Accepted Answer

The models cleared the famous Will Smith spaghetti test in 2025, but the tells moved rather than vanished: physics drift, dead-eyed emotion, inconsistent lighting, and rubbery hands still betray fully synthetic humans, which is why Coca-Cola's AI Christmas ads got roasted two years running. What reliably works today: b-roll, landscapes, product-in-scene shots, stylised or absurdist comedy (the Kalshi trick), and AI voiceover. What still reads fake: photoreal human close-ups carrying emotional weight. Match the format to the model and keep humans on script and edit.

Question 4

Which AI video tool should I actually use, and for what job?

Accepted Answer

Pick by job, not hype: Veo 3.1 or Gemini Omni Flash in Flow for photoreal product shots with synced sound, Kling 4.0 for long takes of up to 30 seconds, Seedance 2.5 for clips steered by many references (from India, via Runway or Luma), Runway for several models under one subscription, HeyGen or Synthesia for talking heads, ElevenLabs for voice. OpenAI closed Sora in 2026 (the app in April, the API in September), so ignore comparisons that still recommend it. Most founders settle on one generator, one avatar/voice tool and one editor. Rankings move every quarter, so re-test before renewing.

Question 5

How are founders using AI avatars and UGC-style actors instead of hiring creators?

Accepted Answer

Avatar tools (HeyGen, Synthesia, Arcads) turn a script into a talking human on camera in minutes, so founders use them for UGC-style ads, onboarding videos, sales outreach, and multilingual founder updates without booking real creators. Performance teams generate dozens of hook and actor variations for the cost of one influencer shoot and let the ad account pick winners. The trap: avatars carrying emotional brand moments still read synthetic, so keep them for information-dense, high-volume formats.

Question 6

How are founders using AI voice, music, and audio (ElevenLabs, Suno, Gemini Notebook)?

Accepted Answer

Voice is the cheapest quality upgrade in the stack: an ElevenLabs voice gives every demo, ad and explainer professional narration cheaply, and with Eleven v4 one cloned voice speaks 90+ languages. An instant clone works from 10 seconds of audio; a Professional Voice Clone (30+ minutes of clean audio) gets close to indistinguishable from you. Suno v6 makes full songs fast, but Universal and Sony are suing over it and its free plan has no commercial rights, so for an ad use a paid plan or Eleven Music. Gemini Notebook (formerly NotebookLM) turns documents into podcast-style audio, and Sarvam and Murf cover Indian voices.

Question 7

How do founders create product images, brand visuals, and logos with AI?

Accepted Answer

Image models now split by job: Nano Banana 2.1 in Gemini for editing real product photos into lifestyle scenes, ChatGPT Images 2.5 for marketing graphics, mockups and precise edits, Midjourney V8.2 for brand look and campaign imagery, Ideogram for logos and typography, and Adobe Firefly when you need output trained on licensed content. The pattern that works: keep the real product pixels and generate the world around them, rather than the product itself. Pure AI output has no copyright protection, so keep your human edits documented.

Question 8

What does AI video and image actually cost, and what's a sensible starter stack on a founder budget?

Accepted Answer

Rough 2026 numbers: $0.05-0.75 per generated second via APIs, or $5-30 a month for entry app plans, so a 30-second spot costs about $1.50-22 in raw generation, but budget for discarded takes (the Kalshi ad burned 300-400 generations for 15 usable clips). A workable starter stack: Flow on Google AI Pro ($19.99 a month) or Kling, HeyGen Creator ($29) plus ElevenLabs ($6-22), and Descript or the free DaVinci Resolve, all under $100 a month. In India, check whether your Jio plan already includes Google AI Pro. Iteration volume, not price per clip, is the real budget line.

Question 9

How do you write prompts that actually get usable AI video and images?

Accepted Answer

Treat every prompt like a director's brief: subject, action, scene, camera, lighting and audio, one shot per prompt. Google's Veo guides plus JSON prompting (structured fields instead of prose) are the fastest way in, and references beat words for locking a look: Kling 4.0 takes up to 15 assets, Seedance 2.5 up to 30 images. Refine rather than restart: Gemini Omni Flash in Flow edits a clip from a plain instruction, so get the shot close, then describe the fix. Expect roughly a 20:1 generation-to-keeper ratio.

Question 10

How do you keep characters and brand look consistent across AI-generated scenes and assets?

Accepted Answer

Consistency comes from references, not prose: feed the model reference images (Runway's references, Midjourney V8.2's Edit model, Kling 4.0's up to 15 assets, ingredients in Flow) and pin palette, wardrobe and lighting in a reusable template. Midjourney's Edit model replaced Omni and Character Reference in July 2026, so older --oref guides no longer fit. Nano Banana 2.1, ChatGPT Images 2.5 and Gemini Omni Flash hold a character across edits. For a brand, lock one style reference and reuse it everywhere.

Question 11

How do founders dub and localize video into Indian and global languages with AI?

Accepted Answer

AI dubbing turned localization from a lakhs-per-language studio job into a same-day export: HeyGen does lip-synced translation into 175+ languages, ElevenLabs dubs across 90+, and Sarvam dubs one upload into 12 Indian languages keeping each speaker's voice, from Rs 40 a minute. Dubverse and Murf were also built for Indian languages. Informational content is close to ready on the first pass; ads still need a native-speaker review before spend. In India, founders consistently report regional-language versions outperform English on reach and engagement.

Question 12

How do founders turn one long recording into weeks of short clips?

Accepted Answer

The repurposing engine: record once (a podcast, demo, or talk), let OpusClip or Descript find the hooks, auto-caption and reframe to vertical, and post 5-10 clips per hour of footage across Shorts, Reels, and LinkedIn. Text-based editing in Descript makes the cut as fast as editing a doc, and the agentic editing a16z predicted in January is here: Descript's Underlord co-edits with you and Gemini Omni Flash edits a clip from a description. This is the highest-ROI AI video workflow for founders because it needs zero generation, just footage you already have.

Question 13

What are the legal, copyright, and disclosure rules for AI-generated marketing content?

Accepted Answer

Three layers to know: India's amended IT Rules require synthetically generated content to be clearly labelled, with platforms adding 'Made with AI' tags; in the US, the FTC bans AI-generated fake reviews and testimonials and New York requires disclosure of synthetic performers in ads; and pure AI output gets no copyright, only your human contributions do. From 30 October 2026, Meta's terms make you solely responsible for what its AI adds to your ads. The practical rules: label realistic synthetic humans, never fake a testimonial, check what Advantage+ changes, keep records of your creative input, and own your reference assets.

Question 14

Do AI-made ads actually perform, or is it hype?

Accepted Answer

The performance data says yes for cost and testing volume, mixed for absolute quality: Meta reports roughly 22% ROAS lift from its AI-powered campaigns with over 4 million advertisers using its gen-AI tools, and AI UGC cuts cost-per-creative around 10x, letting brands find winners among dozens of cheap variants. But Kantar finds 44% of consumers are bothered by obviously-AI ads, Menlo Ventures found 36% of US consumers less likely to engage with known AI content, and creative quality drives 70-80% of ad performance. The winning pattern: AI for volume, human judgment for the concepts that carry the brand.

How founders use AI for Video, Image & Audio

The questions

Also in Starting Up