Notes from the lab · What it costs
Why your cost per asset never falls, and the one thing that fixes it
Model prices have dropped every quarter for 2 years. If your cost per asset has not, something is eating the saving. This note names it.
The arithmetic nobody runs
Cost per approved asset is the only number that matters, and it has two parts: what each attempt costs, and how many attempts a keeper takes.
| Scenario | Cost per attempt | Approval rate | Cost per approved asset |
|---|---|---|---|
| Today, typical | ₹40 compute + ₹210 human handling | 30% | ₹833 |
| Compute price cut to zero | ₹0 compute + ₹210 human handling | 30% | ₹700 |
| Approval rate fixed instead | ₹40 compute + ₹210 human handling | 90% | ₹278 |
Every attempt drags human time with it: the brief, the review, the rejection note, the re-brief. That handling cost dwarfs compute, so cheaper models barely move the total. The approval rate moves it 3x. The lever is not the model price. It is how often the first attempt is right.
Why the approval rate never improves
Because the learning keeps leaving. What makes attempt one land is accumulated judgment: this brand never shoots from below, the founder hates blue backgrounds, the Diwali campaign needs the product in the first 2 seconds. In the standard setup that judgment lives in people and decks. The agency team rotates, the deck goes stale, the vendor changes at the annual review, and campaign 12 starts as ignorant as campaign 1.
You are not paying for production 12 times. You are paying for the same education 12 times.
The one thing that fixes it
Make the judgment machine-readable and keep it. Every approval and every rejection in our pipeline becomes a training signal attached to the brand, not to a person. Reject an asset and the reason is captured; the system stops proposing that mistake. Approve one and the pattern reinforces. The brand file gets smarter with every shipped asset, and it cannot resign.
On one account we measured it cleanly: first-pass approval went from 56% to 92% over 90 days of the quality loop running. Compute prices were flat that quarter. Cost per approved asset fell to roughly a third, entirely from the approval rate.
What to do with this
- Measure cost per approved asset, not cost per generation. Include the human handling time.
- Track first-pass approval rate monthly. If it is flat, you are re-buying the learning curve.
- Ask where the judgment lives. If the answer is a person or a deck, it leaves when they do.
Want your approval rate on a curve like that?
We will show you the 56 to 92 loop on the call.