From Sketch to Spot: The AI Ad Workflow Behind a $2,000 NBA Finals Commercial

Key Takeaways

• In June 2025 the prediction-market company Kalshi aired a commercial made with AI during Game 3 of the NBA Finals. The filmmaker PJ Accetturo told reporters it cost about $2,000 and took two days (eWeek) or two to three days (other reports).
• The workflow was: a rough script, then Gemini to write a shot list and prompts, then Google's Veo 3 to generate clips, then manual editing. It took 300 to 400 generations to get 15 usable clips.
• That works out to roughly 20 to 27 attempts per keeper, which is the number to plan your own budget around.
• A Kalshi employee said a comparable commercial would typically cost seven figures and take months. That comparison is the company's claim, not an audited figure.
• The workflow generalises: sketch the idea, lock the references, generate in short clips, and edit by hand.

What was actually built

According to eWeek's report of 16 June 2025, the ad was made by Accetturo, a self-described AI filmmaker, for Kalshi. Gemini and ChatGPT helped write the rough script, Gemini turned it into a shot list and prompts, and Veo 3 produced the clips, which Accetturo assembled and edited in a video editor. The article says the ad "reached millions of viewers" but does not name a source for that, so we do not repeat a reach figure here.

Case study: Kalshi, NBA Finals Game 3 (2025). A prediction-market company ran a fully AI-generated commercial during Game 3 of the NBA Finals. The creator reported a cost of about $2,000, two to three days of work and 300 to 400 Veo 3 generations for 15 usable clips. A Kalshi employee compared this with a typical seven-figure, months-long production. The takeaway for a small team is not that AI is free. It is that the cost moved from crew and location to iteration, taste and editing.

The numbers that matter for your budget

Most people read the $2,000 and stop. The more useful number is 300 to 400 generations for 15 clips. If you assume a similar hit rate:

That is arithmetic on a single reported project, not a guarantee. Your hit rate depends on how tight your references and prompts are. The cost of one generation depends on the model you choose, so get a quote for the model and length before you start a batch.

A sketch-to-reference-to-video workflow

Rough clips are easy to generate. A set of clips that look like the same film is the hard part. This order keeps them consistent.

1. One-line idea and one action. The ad must say one thing.
2. Shot list. Ask an assistant to turn the script into a numbered list, one shot per line, with framing, subject, action and mood.
3. Rough visual plan. This can be a pencil sketch, or a simple 3D blockout in Blender with grey boxes for set and product. It fixes camera position and spacing before any model gets involved.
4. Reference images. Turn each sketch into a polished still using an image model, keeping the same character, product and look. These stills become the references for every later clip.
5. Clips from stills. Generate each shot from its reference still, in short lengths, and keep several takes.
6. Edit. Choose takes, trim, add sound and music, and fix continuity in the cut.
7. Review. Watch it on a phone, with and without sound.

In MITPO, Sketch drafts the storyboard for you to approve before any credits are spent, and Creative Studio makes the images and short clips, with the price shown before you generate.

Mistakes that burn generations

• Prompting without references. Every clip invents a new person and a new room.
• Long clips. More seconds mean more chances for something to go wrong. Keep shots short and cut them together.
• Changing the style mid-project. Choose the look first and keep it.
• Skipping the sketch. A rough layout is cheaper to change than a rendered clip.
• Ignoring sound. Sound carries a lot of a commercial's feel. Plan it early.

Rights, faces and disclosure

• Faces. Use faces you have the right to use: your own, a hired model's with a signed release, or a synthetic face with no real person behind it. Do not generate a recognisable real person without consent.
• Brands and music. Clear what is not yours.
• Disclosure. Each ad platform and some countries have their own rules about labelling AI-generated content. Check the current rules for the platforms and regions you buy in before you publish.

FAQ

Did AI really replace a seven-figure production?

For that spot, according to the people involved. The seven-figure comparison is a Kalshi employee's statement, and the $2,000 is the creator's figure. Neither was independently audited, so treat them as reported.

How many generations does a 30-second ad need?

On the reported ratio, plan for roughly 20 to 27 attempts per usable clip, then adjust to your own hit rate after the first batch.

Do I need 3D skills to use a Blender blockout?

No. Grey boxes and a camera are enough to fix framing. You can also use a photographed pencil sketch instead.

Which video model should I use?

Pick the model that fits the shot (motion, length, sound) and your budget, and compare costs before a batch. Our video model comparison is a starting point.