Generative AI and creator media are making it faster to produce ad creative—and dramatically increasing the number of versions available. For marketing leaders, the challenge has shifted from “not enough creative” to “with so many versions, where should we invest the budget?” If a model can score every ad, should higher-scoring creative receive more budget? Could enterprises even allow the system to adjust campaign spending automatically?
We believe enterprises can incorporate AI scoring into advertising investment decisions to screen creative earlier, but scores should not directly trigger budget changes. Before implementation, enterprises must translate the scoring criteria into a shared set of rules across departments, require managerial approval, and retain records of both decisions and execution. AI can help narrow the options, but the final budget decision—and accountability for it—must remain with an authorized decision-maker.
An 81% Accuracy Rate Is Informative, but It Does Not Guarantee That Increasing the Budget Will Work
According to TechOrange, Kroger partnered with Vidmob and MMA Global to analyze 1,934 video and image ads that Kroger had published on Meta and Google Display & Video 360 (DV360). The research team examined each ad’s messaging, narrative structure, brand presentation, and character interactions, then cross-referenced those creative attributes with actual e-commerce conversion results. The model was trained using a full year of marketing campaign data and subsequently validated against data from a later quarter that it had not previously encountered.
The research report stated that predictive creative scoring predicted e-commerce conversion rates with 81% accuracy. Creative that followed the study’s recommendations achieved an average conversion rate four times that of creative that did not. The model also indicated that reallocating the same budget to higher-scoring creative could generate up to 2.2 times as many conversions.
These three figures describe different things: 81% is the prediction accuracy, four times is the difference in average conversion rates between two groups of creative, and 2.2 times is the model’s estimate after budget reallocation. They should not be combined and presented as a single performance guarantee. The original report’s headline says “three times higher,” while the body says “four times as high”—two different ways of expressing the result. When preparing internal presentations, enterprises should clearly identify the original wording rather than highlighting only the most eye-catching multiplier.
When evaluating similar tools or proposals, executives can begin by asking several specific questions: What time period did the model’s data cover? Was the model separately validated against subsequent data? How was a conversion defined? Which platforms and campaigns do the results apply to? Rather than focusing immediately on more granular figures, enterprises should first establish which business outcome the score represents. Otherwise, they risk treating research findings as their own company’s expected results.
Establish Shared Rules for Interpreting Scores Before Bringing Them into Budget Meetings
Even after a model assigns a high score, the marketing, brand, and data teams may interpret it differently. Brand teams need to verify that trademarks and language comply with established guidelines. Media teams focus on audiences, placements, and campaign periods. Data teams must confirm the definition of conversion events, the scope of the sample, and whether any data is missing. Unless the enterprise establishes shared rules in advance, different departments may reach different conclusions when interpreting the same score in separate meetings.
The three teams can begin by jointly defining three categories of criteria: requirements, exclusions, and exceptions. Requirements should cover whether the brand name, product information, and campaign restrictions are complete. Exclusions may include expired pricing, creative with unclear licensing rights, and language that does not comply with brand guidelines. Exception criteria should specify which items must be reviewed by a person when dealing with new products, seasonal campaigns, or insufficient data. These criteria are not intended to restrict creativity. Their purpose is to ensure that every review begins with the same baseline requirements.
An enterprise can start with a pilot campaign, recording both AI scores and human judgments in the same review record, and then compare the actual conversion performance of high- and low-scoring creative. Scores should not be allowed to change the entire budget at the outset. Once the team has confirmed that the rules are workable and the operating process has become stable, it can formalize the established practices as an AI Skill in EgentWrX, enabling the AI Agent to follow the same rules every time. Before scaling the process across the organization, the enterprise must still review and approve this standardized approach. During the pilot stage, the relevant teams should jointly define the requirements, exclusions, and exceptions from the beginning rather than simply handing each department a score to interpret independently.
Require Managerial Approval Before Adjusting Budgets—and Preserve a Clear Audit Trail
Enterprises can divide the process into four stages: creative organization, AI evaluation, marketing manager review, and budget adjustment. The information provided to the manager must include more than a score. It should also include the creative version, the scope of the data used for the evaluation, the recommended budget adjustment, and any applicable exception criteria. The person responsible may adjust the budget only after receiving managerial approval. If the manager rejects or modifies the recommendation, the reason should also be documented in the enterprise’s own review form or process documentation.
EgentWrX can connect multiple tasks into a workflow and incorporate human approval before the process advances to the next stage. After the AI Agent completes the organization and evaluation of the creative, the workflow pauses for managerial approval rather than proceeding automatically to the budget-adjustment stage. Enterprises can require the responsible employee to document the creative version, evaluation date, recommendation, approver, and final action in their own review form or process documentation. Managers can then use the EgentWrX audit log to review the actions performed by the AI Agent and use a hash chain to verify the integrity of the records. Enterprises can use these mechanisms to govern the process and preserve evidence for subsequent review, but they do not guarantee improved advertising performance.
During the pilot period, enterprises can routinely sample two types of cases: those in which the manager accepted the AI recommendation and those in which the manager rejected or modified it. The team should compare the data used at the time, the manager’s rationale, and the actual campaign results to determine whether the rules overlooked any brand risks or important business considerations. Before deciding whether to expand the approach to other campaigns, the enterprise should document who is responsible for approval, which items must be reviewed, and how cases will be audited afterward. No matter how precise a score appears, an enterprise must not use it as a reason to bypass the manager responsible for the budget.
FAQ
Can we introduce AI scoring without a large volume of historical data?
You can begin by organizing the data and conducting a pilot, but the scores should not immediately influence a large budget. The enterprise should first verify that creative versions, campaign conditions, and actual conversion results can be matched accurately. It can then select one campaign to build a comparable record and determine whether the available data is sufficient to support the next stage of decision-making.
If AI scoring reaches 81% accuracy, can we shift the budget to higher-scoring creative?
No. The 81% figure is the accuracy achieved in that study using specific data and validation methods. It does not mean the model will make the correct judgment every time. Enterprises must still define the applicable conditions, the permitted range of budget adjustments, and the managerial approval process. They must also continue testing the reliability of the scores against the actual results of their own campaigns.
What information should a manager review at a minimum?
At a minimum, the manager should see the creative version, evaluation date, data scope, AI recommendation, proposed budget adjustment, and exception criteria. If the interface displays only an overall score, the manager cannot determine whether the recommendation complies with brand guidelines. It will also be difficult to trace the cause if performance falls short of expectations.
Can we adjust the existing budget without producing new creative?
Yes. An enterprise can first group its existing creative and test reallocating part of the budget to higher-scoring ads, but it should define the maximum adjustment and observation period in advance. The 2.2-times figure cited in the Kroger study is the model’s estimated maximum outcome. It should not be treated as a result that other enterprises can expect to achieve.
References
- No More Blind Ad Spending: Kroger Study Finds Ads Following AI Recommendations Deliver Conversion Rates Three Times Higher — TechOrange’s summary of the predictive creative scoring study conducted by Kroger, Vidmob, and MMA Global.
