Imagine giving one system your ad accounts, brand kit, product imagery and historical campaigns. On Monday it generates a new batch of ads. After approval, it publishes variants to Meta, Google and TikTok. By Wednesday it has read the performance signals and is already producing the next round around the work that earned spend.
At that point, you are no longer buying creative output alone. The system is participating in production, account operations, budget movement, performance interpretation and the next creative decision. A contract that specifies only the number of videos and images leaves the important questions unanswered: what is the system learning for, who can intervene, and what does the brand retain when the relationship ends?
What changed—and what has not been proved
Runway introduced Runway Ads on September 30, 2026. The company says the product connects an ad account and brand kit, generates video and image creative, publishes approved variants directly to Meta, Google and TikTok, reads performance back, and produces another round based on what earned spend. Automated brand checks feed an approval queue; human approval is on by default, although teams can enable automated publishing by campaign or variant type. Runway also describes controls for daily budget changes, net-new audience share and retargeting caps.
Availability matters. Runway Ads is still being piloted with selected enterprise partners. The launch post also reports that Runway's own program increased weekly ad volume from 77 to roughly 900 while doubling return on ad spend. Those are first-party results from one advertiser, not a benchmark or forecast for customers.
There are adjacent signals from the major platforms. On October 5, TikTok announced Smart+ Creative Selection, which brings advertiser assets, Symphony-generated creative and creator content into one approval workflow. Google Ads' AI Max reporting can connect a search term with the headlines and landing page used in the same ad journey.
These announcements establish what three companies have built. They do not establish that the entire advertising industry has completed the same transition. The procurement problem, however, is already real for any team considering one of these systems.
“Earned spend” is not a business objective
An integrated system can observe platform-native signals: impressions, clicks, conversions, spend allocation and other events available through an ad account. The business may care about something else: contribution margin, qualified new customers, return rates, lead quality, inventory pressure or long-term brand effects.
A sensational hook may earn budget while creating the wrong product expectation. A discount-led variant may convert efficiently while eroding margin or training customers to wait for promotions. Even when the platform data is accurate, what the system can observe and what the company should optimize are not the same thing.
That is why the first deliverable in an autonomous creative engagement should be a learning mandate, not a volume commitment. The structured product facts, rights and editable assets in an Agent-Ready Brand Pack can serve as inputs. The activation layer needs six additional decision rights.
The ONCE Creative Learning Rights Schedule
“Rights” here is a commercial governance label, not a claim about a universal legal category. These are six control areas to specify across the brief, contract, account permissions and acceptance checklist.
1. Objective rights: define what the loop is allowed to optimize
Name the primary business objective, secondary indicators and non-negotiable guardrails. “Improve performance” is not sufficient. Is the priority contribution-margin-adjusted acquisition cost, first-time customers, qualified leads, movement of a specific SKU, or efficient reach?
Document which outcomes the system can actually read. If it sees spend and platform conversions but not returns or margin, it should not silently translate “earned more budget” into “deserves more production.”
2. Input and provenance rights: define what may become training material for the next variant
Brand guidelines, historical ads, creator content, product feeds, music and human performances do not carry identical permissions. For each source, state whether it may be used for generation, localization, modification, cross-market derivatives or improvement beyond the brand's own account. Record territory, term, owner and verification responsibility.
An uploaded asset is not automatically cleared for every future use. Creator and talent agreements are especially likely to distinguish the original placement from AI-altered derivatives or new markets.
3. Approval and activation rights: define who can make a creative real
Human approval by default is a useful starting point, but approval should be risk-tiered. Aspect-ratio adaptations, caption layout and non-claim backgrounds may qualify for batch approval. Price, efficacy, human performance, promotional terms and new-language claims should retain a named human gate.
Account permissions also need separation. Who may connect an account, upload inputs, approve creative, change budgets or switch on automated publishing? A shared administrator login is not a governance model.
4. Observation and export rights: make the learning legible outside the tool
Each cycle should export stable variant IDs, source assets, changed variables, platform and placement, market, observation window, spend, native metric definitions, approver and the stated reason for the next generation.
A folder containing only “winning” files is inadequate. Without the control, rejected variants and delivery context, the brand cannot tell whether the creative idea, audience, timing or budget distribution caused the difference. It also cannot transport the learning to another channel or supplier.
5. Intervention and rollback rights: stop a bad lesson before it branches
Predefine daily budget-change limits, maximum spend per unproven variant, anomaly thresholds, excluded URLs, forbidden claims, complaint triggers and emergency contacts. Specify whether pausing a parent variant also stops its descendants, who removes incorrect derivatives and how the action is logged.
In a regenerative system, one error can become a family tree. Rollback must address the lineage, not just the latest ad.
6. Portability and exit rights: retain the work that makes the next team smarter
The exit package should include editable masters, approved and retired variants, a variable map, native metrics and definitions, test conclusions, unresolved hypotheses, rights expirations, credential revocation and any required deletion confirmation.
The brand does not necessarily own a vendor's model or software. It should own enough operational memory for the next team to understand what was tested, why it stopped and what remains worth pursuing.
Choose the operating model before choosing the tool
A project-based asset package remains appropriate for a launch, a short campaign, a small number of markets or highly bespoke storytelling. The cost of deep account integration may exceed the value of the learning.
A supervised learning loop fits stable products with ongoing media, clear objectives and a team prepared to review results every one or two weeks. The system generates and organizes; people approve, interpret and choose the next variable.
Partially delegated publishing and budget control belongs with mature teams that have reliable product truth, clean rights, measurable outcomes, active monitoring and revocable access. Delegation should expand by risk class, not through a single all-or-nothing switch.
A practical test is simple: if the team cannot consistently answer what changed, why the variant was allowed to launch and which next decision the result informed, it is not ready for an unsupervised loop.
Price the approved learning cycle—not the asset count
When no defensible market average exists, use the buyer's own cost structure:
Cost per approved learning cycle = (system and account setup + product truth and master assets + generation and human review + media operations + measurement and interpretation + rights and exit management) / completed approved learning cycles
A completed cycle requires a declared hypothesis, comparable variants, compliant approval, a defined observation window, an interpretation and a decision to continue, stop or test again. Five hundred generated ads without one completed decision do not create a denominator of 500.
This metric does not replace margin, return on ad spend or sales. It answers a different management question: how much reusable judgment did the creative and operating budget actually buy?
What changes for production partners
This is an ONCE inference, not a declaration that the production market has already changed. If generation, activation and performance feedback continue to converge, pure file-output pricing will face pressure. The more defensible role for a production partner is to capture reliable product truth, design comparable variables, manage rights, configure approval gates, interpret channel differences and turn the result into the next shot list.
That model has conditions. It is most valuable for always-on programs with sufficient spend, stable products and usable feedback. Seasonal campaigns, singular narratives and businesses that cannot reliably connect media outcomes to operating results may still be better served by excellent project-based production.
Runway Ads demonstrates that one vendor has connected steps that previously lived across multiple tools. It does not require a brand to replace its agency tomorrow. It does require a better procurement document today.
Ask more than how many ads the system can produce. Ask who defines success, who approves activation, what the brand can observe, how mistakes roll back, and what leaves with the brand at the end.
The scarce asset is not the next ad. It is ownership of the next decision.
