Lead with the real test: does the ad survive a living‑room screen?
In June 2025 a 30‑second ad for Kalshi ran during Game 3 of the NBA Finals on YouTube TV — and no one filmed a frame of it. NPR reported the AI prompting that replaced studios, directors and actors cost under $2,000. The catch: the creator still ran hundreds of generations to find 15 usable clips.
If you want an AI spot that can run beside traditionally produced ads, you can get there without a crew. But you cannot skip the same production discipline: a clear brief, shot‑by‑shot prompts, continuity checks and a simple, repeatable test plan before you spend on media.
Why this matters now
Generative tools are becoming part of advertisers’ toolkits. The IAB’s 2026 video report found more buyers using generative AI year over year and that one‑third of digital video ads now include some AI work. Smaller advertisers are also buying connected TV more often as self‑serve platforms lower the media entry barrier.
At the same time, consumer sentiment is mixed. A 2026 Canva and The Harris Poll study found 70% of people say AI‑generated ads can feel like they’re missing something, though 68% said they don’t mind AI in ads when it’s relevant or useful. Viewers judge the finished spot, not the tool that made it — so judge your output the same way they will.
A practical five‑step production workflow
Step 1 — One‑line brief and locked assets
Start with one sentence that answers: who is this for, what you sell, and what the viewer should feel or do. (“Busy parents — a 10‑minute meal kit — order this week’s box.”) Anything vaguer is a draft, not a brief.
Collect assets the model can’t invent: product page link, high‑res logo, brand colors and fonts, one verifiable proof point (real review count, a backed guarantee or a stat), and the offer. If the tool accepts product URLs, give it the page so the script pulls accurate details rather than guesses.
Step 2 — Script to a 30‑second beat sheet
A 30‑second spot fits a problem, the product, proof and a clear ask. Many marketers prefer this length: a 2026 Wyzowl survey found a majority rate 30 seconds to two minutes as most effective for video.
Assign every interval a job:
- 0–3s: Hook — must read on mute
- 3–12s: The world or the customer problem
- 12–22s: Product + one proof point (hero shot)
- 22–27s: Offer or next step (price, deal, CTA)
- 27–30s: End card — logo, URL, held frame
Keep spoken copy tight (about 70 words). Read it aloud against a timer and cut words, not beats.
Step 3 — Generate shot‑by‑shot, not with one giant prompt
Treat each beat as a mini brief: subject, setting, camera move, lighting and mood. Concrete prompts beat vague directives; tell the model the material facts you need to remain consistent (product dimensions, label copy, reference image).
Expect discard rates. High‑profile creators report hundreds of generations to assemble a handful of usable clips; budget time and credits for that. Prefer tools that let you regenerate a single shot without rebuilding the entire spot.
Step 4 — Finish audio and typography for how people actually watch
Choose one consistent voice for the VO and keep it across versions. Mix music under dialogue and automate or manually duck music during speech so the VO reads clearly on phone speakers. Add captions because many viewers watch muted.
Export every aspect ratio you plan to run (9:16, 16:9, 1:1 or 4:5) and check that key frames and text survive each crop — don’t let important copy sit at the edges where it will be cut off.
Step 5 — Run the six‑point TV‑quality checklist, fix the failing frame, then test variants
Before buying media, verify the exported ad against six practical checks viewers will notice:
- Hook reads on mute (first 3 seconds).
- Hero frame survives a freeze — no warped hands, melted text, drifting labels.
- Continuity holds — product shape, color and lighting remain consistent across cuts.
- Voice sits above music and is clear on phone speakers.
- Type is readable on a phone and the end card stays long enough to read twice.
- It reads like your brand even with the logo covered (colors, fonts, tone).
Fail one check and treat the spot as a draft. Fix only the offending shot where possible; then create three variants that change a single element (hook, offer, or voice) and run a side‑by‑side test on a modest budget to let performance pick the winner.
Budget and vendor considerations
Traditional small productions often start in the low thousands; the Kalshi example shows prompting costs can be much lower, though that figure excluded creator time and the generation effort behind the final 30‑second cut. Many platforms use subscription or credit models — check whether they support per‑shot regeneration and whether credits are consumed by small edits. Those features materially change the economics.
What to watch next — and the practical next step
Start with one controlled experiment: produce a single 30‑second spot, run the six‑point quality check and launch three single‑change variants on a small buy. Invest in a locked brief and product assets first — they reduce wasted generations and continuity problems more than any prompt tweak.
Watch two trends closely: the rise of connected TV self‑serve among small advertisers, and shifting consumer tolerance for AI content when it delivers clear relevance or utility. Tool features that matter most in practice are single‑shot regeneration, brand asset kits and caption exports — they distinguish a prototype from production‑ready creative.
Practical next step: draft your one‑line brief now and gather the product page, high‑res logo and one verifiable proof point. With that in place you can run a low‑cost test production and learn where the tool needs guardrails and where it genuinely speeds you up.