Instant Studio logo

← All playbooks

Approve the workflow once

· Instant Studio · brand control governance

Where human approval sits in an AI production line A vertical production line: brand rules are encoded and signed off by legal once, then three human gates follow at concepts, first outputs and final cut, with generation, volume production and delivery in between. Encode the brand rules Claims, disclosure, talent and format rules LEGAL SIGNS OFF HERE, ONCE GATE 1 Concepts Direction agreed before anything is made Generate Prompts expand inside the encoded rules GATE 2 First outputs A calibration batch before the volume run Volume run Hundreds of assets, one configuration GATE 3 Final cut The ship decision, plus flagged exceptions Deliver and record Every asset traceable to the workflow
Sign-off happens once on the configuration. Human judgement then sits at three points in the run.

Approval is the first thing that breaks when AI creative production scales.

A production line returning five hundred assets a week hands the brand team ten times the review load it had before, and the reviewers are the same three people. Adobe's 2025 survey of more than 1,600 marketers found 58 per cent of them already spending upwards of 40 per cent of their time managing reviews and approvals, and that was before the volume arrived. Generation that only produces a larger review pile has moved the bottleneck one desk to the left.

The humans stay. The useful question is where to put them.

What happens when every asset goes to review

The default design is to generate freely and route everything through brand review at the end. At fifty assets a week it works. At five hundred it becomes the bottleneck it was meant to prevent. Reviewers rubber-stamp under deadline pressure, inconsistencies ship anyway, and the brand team concludes, reasonably, that the output cannot be trusted. The AI takes the blame for what is really a failure of workflow design.

The pile is also the one place quality cannot be fixed. Review catches a wrong shade of red after the asset exists, one asset at a time, forever. It never stops the next hundred from arriving with the same fault.

Put the rules in front of the model

Brand guidelines, claims language, disclosure requirements, approved talent, format specifications: these are constraints, and constraints belong upstream of the model call. On our platform they are encoded into the Creative Workflow before anything is generated, so a marketer's two-line note gets expanded into full generation prompts inside walls the brand has already agreed.

That changes what arrives at the gate. The reviewer is no longer checking whether the logo lockup is right and whether the claim is legally sayable. Those were settled at generation time. What is left is the judgement a person is genuinely needed for: is this good, does it fit the campaign, would we have shipped it if a crew had shot it?

Legal reviews the workflow, once

The same move solves the harder problem in regulated categories. Counsel vets the rules, templates, claims logic and disclosure behaviour one time, and every asset the workflow then produces stays traceable to that approved configuration.

Approving five thousand assets individually is not an operating model. Approving one controlled workflow, then reviewing exceptions and the final cut, is. Sun Life's campaign produced 27,547 invitations, 22,213 personalised selfie eCards and 5,565 personalised videos across four regulated markets, and every asset was legally approved. A campaign of that shape only exists if the sign-off happens on the configuration.

The practical consequence for a pilot is a scheduling one. Legal needs briefing in the week the workflow is being encoded, not the week the assets land. Teams that book that meeting early get a system their compliance function has already blessed. Teams that skip it get a review queue with a lawyer at the end of it.

Three gates worth keeping

With the rules enforced upstream, human approval concentrates at three points:

  • Concepts, before generation. Creative direction, territory and the shot list. The cheapest place to disagree, because nothing has been made yet.
  • First outputs, before volume. A small batch that proves the workflow is calibrated. The first outputs are usually wrong in instructive ways, and that is the calibration the system learns from.
  • Final assets, before delivery. The ship decision, plus anything the workflow flagged as an exception.

Clients sit at whichever of these they want, and some sit at all three. The gates are deliberate choices about where judgement is worth paying for, which is what stops them from decaying into a rubber stamp.

Make rejection cheap

A gate is only real if saying no is affordable. When a workflow is one indivisible job, rejection means starting again, and the pressure to wave things through is enormous.

A node-level rerun on a Canvas workflow Four connected nodes: product asset, scene, render and output. Brand rules feed the render node. A reviewer changes the scene node, so scene, render and output rerun while the product asset and brand rules stay unchanged. Brand rules Product Scene Render Output UNCHANGED CHANGED RERUN RERUN
The reviewer keeps the model and the lighting, changes the setting, and reruns three nodes instead of the job.

Our workflows run as a Canvas, a graph of nodes, and they rerun node by node. A reviewer who likes the model and the lighting but not the setting changes one node's input and reruns that node. The rest of the asset survives. Corrections become surgical, rejection stops being expensive, and reviewers start using the gate for what it is.

Every one of those decisions is also feedback. Rules get sharper, templates get tighter, and the share of assets approved without rework climbs month on month. That number is the one to watch in a pilot, alongside cost per finished asset and time from brief to live.

Keep the paper trail

Automation that a compliance team can defend needs a record of what governed each run and who cleared the result.

The minimum production record for a finished asset A record card listing eight fields: asset and campaign ID, workflow and template version, source asset references, model settings and route, brand and claims rules applied, talent and rights reference, reviewer decision and timestamp, and delivery format and market. PRODUCTION RECORD Asset and campaign ID Workflow and template version Source asset references Model, settings and route Brand and claims rules applied Talent and rights reference Reviewer, decision and timestamp Delivery format and market
The minimum record for every finished asset. Anything less turns an audit into an archaeology exercise.

Two details matter more than they sound. The record should distinguish generated output from the asset that was cleared to ship, because those are different things and audits ask about the second. And the run should be bound to a specific rule set and template version, so a review months later can reconstruct the conditions that produced the work instead of guessing at them.

Human likenesses need documented consent recorded the same way, showing which assets used which talent under which terms.

Where to start

Pick one campaign, encode its rules, brief legal on the workflow in the first fortnight, and place your reviewers at concepts, first outputs and final cut. Then watch the rework rate for a month.

Our free technical guide, Generative media, production grade, covers the workflow architecture, governance patterns and the numbers a pilot should measure, in more detail than one post can carry.