AI UGC Unit Economics: From Model Cost to Cost per Approved Creative

"This model costs 17 cents per second" tells you almost nothing about AI UGC economics.

It ignores:

• failed attempts;
• rejected source frames;
• reviewer time;
• editing;
• storage;
• provider minimums;
• unused concepts;
• client revisions;
• distribution;
• attribution.

The number that matters is not cost per generated second.

It is cost per approved, deployed creative—and eventually cost per incremental business outcome.

Define the production units

Track:

• concept;
• script;
• source frame;
• generation attempt;
• approved asset;
• platform cut;
• deployment;
• measured outcome.

If one approved asset creates three crops and runs on four placements, report:

1 approved creative → 3 exports → 4 deployments

Do not report 12 creatives.

Calculate source-frame cost

Let:

• Cf = cost per frame attempt;
• Af = average attempts per approved frame;
• Rf = reviewer minutes per attempt;
• Hr = loaded reviewer cost per minute.

Example:

The reviewer dominates the API.

That is common. Optimizing a ten-cent model call while the creative lead spends six minutes hunting through near-duplicates is fake efficiency.

Calculate approved-video cost

Let:

• Cv = cost per video attempt;
• Av = attempts per approved video;
• Rv = review minutes per attempt;
• E = editing and assembly cost;
• F = approved source-frame cost;
• O = allocated orchestration/storage overhead.

Use observed attempt rates by shot class:

• static reaction;
• dialogue;
• product interaction;
• multi-person scene;
• hand action;
• camera movement.

A global average will underprice the difficult work clients ask for most.

Allocate failed concepts

Suppose ten concepts enter production and two become approved assets.

The eight failed concepts are research cost. They belong in the economics.

Example:

Calling each final render "$8 of generation" would be technically true and economically useless.

Include opportunity cost

Creative lead time is scarce.

Track:

• minutes to brief;
• minutes to choose a frame;
• minutes to review attempts;
• minutes to resolve claims;
• minutes to package and report.

Then ask:

If automation saves generation cost but increases senior review, is it actually cheaper?

Sometimes the answer is no.

Separate fixed and variable cost

Fixed or reusable:

• creator design;
• rights agreement;
• persona document;
• reference set;
• prompt/tool development;
• templates;
• integration;
• benchmark set.

Variable:

• new concept research;
• generation attempts;
• review;
• editing;
• client revision;
• export;
• distribution;
• measurement.

Amortize reusable assets over their actual useful life.

Do not assume an avatar will be reused for a year before the brand has approved its second video.

Model three approval rates

Technical approval

The file is valid and free of blocking generation defects.

Editorial approval

The creative is good enough to represent the brand.

Client approval

The customer accepts this version for use.

If each is 80%:

A pipeline can look healthy at every stage and still approve half the work.

Route models by expected approved cost

For each model and shot class:

Then add reviewer time and latency.

Do not fill this table from provider marketing. Run your own benchmark.

Price retries explicitly

Set:

• included attempt count;
• client revision allowance;
• what counts as a defect;
• what counts as a changed brief;
• model/provider substitution rights;
• rush pricing;
• cancellation terms.

A generation defect is not the same as:

We changed the product positioning after approving the script.

Without this distinction, the agency absorbs strategy churn as "AI retries."

Add distribution cost

Per-deployment cost can include:

• adaptation;
• caption;
• disclosure;
• authorized upload;
• scheduling;
• community management;
• moderation;
• reporting.

If one approved asset produces ten legitimate deployments, production cost can be amortized:

But business value is not multiplied by the same arithmetic. Ten placements can reach overlapping people or fatigue the audience.

Measure the outcome ladder

Move from cheap metrics to valuable ones:

Calculate:

"Attributable" is difficult. Use holdouts, geo tests, platform experiments, or incrementality methods where the spend justifies them.

Last-click screenshots are not a full causal model.

Example decision

Route A:

• $20 generation;
• 25% final approval;
• 12 reviewer minutes per candidate.

Route B:

• $60 generation;
• 70% final approval;
• 5 reviewer minutes per candidate.

At $1 per reviewer minute:

The expensive model is cheaper.

Change the shot class and approval rates, and the decision may reverse.

The dashboard I want

• cost per concept;
• cost per approved source frame;
• cost per approved asset;
• reviewer minutes per approval;
• attempts by defect reason;
• approval rate by model and shot class;
• client revision rate;
• time from evidence to approval;
• deployments per approved asset;
• qualified outcomes per deployment;
• marginal production CAC;
• asset-family fatigue.

That dashboard tells you whether the operation is learning.

The uncomfortable conclusion

AI UGC is not a money-printing machine.

It can make creative exploration and adaptation dramatically cheaper. That increases the number of useful tests a small team can run.

Cheap tests only create value when:

• the product is good;
• the audience is real;
• the claim is true;
• the creative is approved;
• distribution is legitimate;
• measurement closes the loop.

One video. Hundreds of posts. The margin comes from reusing approved assets and learning across deployments—not pretending rejected generations were free.

For stage-based model decisions, read AI UGC model routing. For production capacity, use How to scale AI UGC production.