What Does an Approved AI Product Ad Really Cost?

Define the accepted ad before you count anything

Start with the buying question, not the tool that will answer it. For our fictional mug campaign, the brief concerns one matte-blue, handleless, 450 ml mug with a removable screw lid; the package holds the mug and the lid, nothing else.

The requested batch is six distinct master ads, each delivered in three required formats — six creative treatments and eighteen final exports, not twenty-four separate ideas. A wider crop of an existing concept is an adaptation, not a seventh concept, and six background colours on one layout is not six different buying arguments.

Write down what makes a master "accepted" before anyone generates anything: the correct product variant, the intended idea communicated clearly, approved copy, and a pass through both product and creative review. A delivery package only counts once its required exports have been checked in the context they'll actually run in — otherwise the denominator moves every time someone wants the cost to look better.

Give production an authoritative reference pack

Production needs just enough material to tell creative freedom apart from a product change. Assemble a compact pack: the correct variant's photos, approved specifications, included components, permitted claims, the current offer, and brand assets the business is actually entitled to use.

Separate product references from mood references. A competitor's mug might have gorgeous lighting, but it's a poor authority for the shape of your own lid — label each reference by purpose so a pleasing detail doesn't quietly migrate into the product. For the mug example, the fixed facts are the body shape, the matte-blue finish, the lid geometry, the missing handle and the actual contents of the box. If a scene needs an angle the pack doesn't cover, go and shoot that angle; asking a generator to imagine the underside isn't the same as photographing it.

Write the fixed requirements next to the creative choices that stay open. Background, lighting direction, text placement and composition can be explored freely. Product identity and approved claims cannot move, which makes a correction request specific instead of a vague ask to make something "look more realistic." Give the pack a version number and an owner — a folder of near-identical reference files is not an approval system.

Run a bounded revision loop, then check what it implies

Treat the first candidates as proposals, not finished creative. For the mug, one version might show the lid removed beside the body; another might go close on the finish; a third might place the product in a workday scene. Each has to preserve the fixed facts and read clearly on its own.

Give every candidate an identifier and keep the prompt, references and generation record attached to it, along with whether it was accepted, corrected or rejected — and why. A critic model can flag an obvious problem such as the wrong colour, invented text or a missing part, but a high score from it never proves product accuracy by itself. Use separate pass/fail conditions for facts and for polish: an attractive composition does not excuse an invented handle, and averaging "looks good" with "is correct" into one score hides exactly the failure review exists to catch.

An attractive composition cannot compensate for an invented handle.

Set a bounded loop before you start: who can request another attempt, what budget remains, and when the team switches to a photograph instead of another generation round. Then check what the finished image implies, not just whether it looks plausible — a mug photographed lying open in a bag can suggest a property the copy never claims; a hand rendered too large shrinks the perceived product. Run the final export, not a preview, through a short checklist:

Check	Acceptance question
Exact variant	Are colour, finish, shape and distinguishing details correct?
Components	Does the image show only what's actually included in the package?
Scale and use	Could staging or perspective imply unsupported size or capability?
Claims	Does every factual promise match the approved product record?
Cropping and text	Does the intended meaning survive the required formats?
Final file	Does the delivered export match the version that was approved?

Give product approval to someone who can compare the asset against the real item and its records; give creative approval to whoever owns message and brand fit. One person can hold both roles on a small team, but both decisions still have to happen separately.

Build the cost ledger that survives a challenge

For a defined batch, the useful number is simple: cost per approved master package equals total attributable batch cost divided by the number of accepted masters delivered in every required format. The numerator covers generation (including rejected attempts), relevant software allocation, labour and a fair share of setup and maintenance. The denominator excludes anything unfinished.

Running the mug example at an illustrative €40-per-hour labour rate, both methods deliver the same six accepted masters and eighteen exports. In the AI-assisted version, generation plus rejected attempts cost around €36, production and correction run €200 instead of €420, and review runs slightly higher at €80 against €60 — a completed-batch cost near €540 versus €660 for the existing method. That's €90 per approved master against €110: roughly 18% less for the same deliverable, with the biggest swing sitting in production and correction time, not the generation line.

Those figures describe one hypothetical batch under stated assumptions, not market prices. Dividing the same €540 by eighteen gives €30 per export, which fairly describes a delivery package — not the cost of eighteen independent concepts, and resizes should never be counted as new ideas.

Watch two traps before calling anything a saving. An incomplete batch is not a finished one: four approved masters at €540 still need the other two built before the comparison is fair. And a shared setup allocation — say €1,200 spread across twenty expected batches — only holds up while that basis stays credible; if only eight batches ever use it, the per-batch share roughly doubles.

A cheap generation that creates expensive uncertainty is not a finished production result.

Keep the ledger simple enough that someone else can rebuild it without taking your word for it:

• Define the accepted package first — product, buying question, distinct masters, formats and approval conditions.
• Prepare one authoritative reference pack, versioned, with a named owner.
• Run a bounded revision loop and log why each candidate was accepted, corrected or rejected.
• Check what the image implies, not just whether it looks right, in the final export and real crop.
• Count the whole batch — rejected attempts, labour and allocations — not just the generation charge.
• Compare finished output to finished output before calling anything a saving.

The eventual answer is usually a hybrid: AI for some production tasks, the existing method for others. That's an easy call to make once the records show where each one actually earns its cost.