01 Facts & context
What did Meta publish during the 2026 advertising week?
2026Year10Month6On the same day, Meta launched a series of AI capacity updates for advertisers and agents during the Advertising Week New York. According to its official press release, these updates cover several directions of creativity, placement, audience and video generation, but the main line is clear: moving AI from ancillary tools to a collaborator that can absorb some of the execution actions.
On the creative side, Ads Creative Studio, which started a small-scale test a few months ago, has entered a broader phase of laying down the insights that can be used to aggregate creative performance. On the side of the project, Meta AI’s business assistant added deeper business memory and began testing so-called “agent” capabilities, including planning and creating captions, generating pictures and files, adjusting audiences or budget changes, and scheduling cyclical tasks and reminders. The Meta Ads MCP server for large advertisers and agents has also expanded its support: creative uploads, advertising previews, effects tracking can be done in existing AI tools, and the work prepared by the agent is entered into the advertising management tool in draft form, subject to permission and explicit approval.
At the audience level, the AI audience discovers that the functionality will enter Detailed Targeting, that advertisers can use natural language to describe the desired client instead of looking in the option, and is scheduled to open later this year; Customer Lifecycle Strategy is open to all advertisers and is used to exclude existing customers in the new caption. On the video side, Advantage+'s creative static material conversion video is fully available, Ads Creative Studio supports the generation and editing of video variants, catalogue advertisers can turn commodity maps into videos in bulk, and recommendations for smart commodities will be expanded in the coming months. The press release cited a positive assessment by an agent executive, which was a statement by a cooperating party and did not represent independent validation.
02 GSP analysis
Where does the platform extend the decision radius to?
What is really interesting about these updates is that instead of a single function, Meta pushed the AI intervention point from "give advice" to "do action." Over the past few years, automation on the platform has occurred mainly at the competitive bidding and budget allocation levels, with creative and audience strategies still largely dominated. Dialogue now creates captions, changes budgets, shifts audiences, and time-bound tasks and reminders mean that the Platform wants to have a place in the strategy-readiness chain. What needs to be seen is that "can do it" and "do it" are two things: the former is a product capacity, and the latter depends on the organization having the means to verify the quality of these actions.
This brings with it a clear choice of organization: which links are given to the platform AI and which must remain inside. An enforceable judgement is to classify each link by two dimensions of "reversibility " and " interpretability " : reversible and easy to explain actions, such as suspension of poor performance material, generation of a draft version of the text, which can be trusted; Irreversible or difficult to explain ex post facto actions, such as pre-positioning of quarterly budget ratios and changing audience exclusion logic, should remain in the human approval chain. The judgement of the GPS is that the value of decision-making lies not in the actions themselves, but in whether the assumptions behind the actions are documented and tested. If the budget adjustment is completed by dialogue and is based only on the words of the assistant, the organization will soon lose its ability to explain why it voted so. On the other hand, if AI is used as an implementation accelerator and targets, budget caps, brand constraints are left in the hands of people, the benefits are focused on efficiency rather than judgement, and the risks are more manageable.
Another option is the difference in access patterns. Small and medium advertisers are better suited to direct use of assistants and audiences in advertising management tools, as their pains are that they do not have enough manpower to do a lot of tests, that the default path in the tool is sufficient, and that additional integration increases the maintenance burden; Large advertisers and agents are more likely to follow the MCP path by embedding the platform ' s capabilities into their existing AI workflows, avoiding back-to-back switching between multiple interfaces, and keeping privileges and clearances in their own systems. The cost structure and governance requirements of these two paths are quite different and are not covered simultaneously by the same set of organizational designs. One scenario assumes that if an agent serves both types of clients, two SOPs will need to be maintained instead of using the same set of processes.
03 GSP analysis
Where's the automated dividend coming from?
The value of introducing AI into advertising operations can be broken down into three mechanisms. The first is to test throughput: Creative Insights structure "what material does well under which type of target" to allow teams to run more effective variables per unit of time, without using memory and cross-mapping to draw conclusions. The second is the speed of response: cycle mandates and reminders remove low-frequency but necessary work such as budget reviews, weekly reports from the to-do list and reduce the waste of forgetting. The third is the cost of preparation: the marginal cost of turning the work of the agent 's output into a draft, and moving the manual "from zero to "review and modify". The three mechanisms correspond to three different types of accounts: testing through throughput is a valid conclusion of the unit ' s human output, responding at a rate of avoided waste, and preparing for cost is the time saved each time.
But all three mechanisms are conditioned. (b) To test throughput is premised on the ability of the team to define what is "effective" or simply to run for more mediocrity; One specific approach would be to prewrite the negative award and minimum sample volume for each cappaign, and then allow the tool to produce a mass output. The speed of response presupposes that the thresholds and rules are clear in advance, otherwise the reminders themselves turn into noise, such as setting the budget review threshold as a consumption deviation plan 15% instead of a daily fixed reminder. Preparation costs are reduced on the assumption that the reviewers have sufficient judgement to identify the key errors in the draft within a few minutes, rather than habitual adoption; A short checklist could be designed, such as whether the audience was in conflict with brand constraints, whether the case touched the red line of compliance and whether the bid was in accordance with the established structure of the account.
Conversely, in the absence of these premises in the organization, automation may bring not dividends but hidden costs: more tests do not have clearer conclusions, faster implementation does not have more robust judgement and more drafts are not reviewed more strongly. The test is not complex: four to six weeks after a certain automation has been implemented, looking back at the record of decision-making at that point, can you still answer "Why the budget was changed" "Why this audience was excluded". If the answer can only be traced to the default recommendations of the tool, it means that the dividend has not been realized and the hidden costs have already occurred. This is why the AI capability is directly equated to efficiency gains and usually reveals problems after two quarters.
04 GSP analysis
Which part of the user's journey was really changed.
The thrust of the update was not evenly balanced from the pull-up to the repurchase. Customer Lifecycle Strategey has a direct impact on newer links: in the past, excluding existing clients by manual spell-outs, easily missing or overexcluding, it is now possible to assign cappit signature as a new feature and to obtain exclusionary advice, and to test an Audience Label to exclude an entire group of customers. This improvement is “whether money is spent on the right people” and is a structural gain rather than a simple efficiency gain.
The audience found that the change was the starting point: the description of the desired client in natural language, rather than the plethora of interest and behaviour labels. For experienced people, this may only save time; For teams that lack audience strategy, it lowers the threshold of entry, while also highlighting the importance of defining “ideal clients” — the more vague the description and the broader the answers the system gives.
Video and creative generation has changed in the medium: the production of materials has moved from project production to closer to the flow line, especially by turning commodity maps in bulk to video, and by adapting hooks and captions to variants. Example of the hypothesis: a brand that makes household goods, originally a quarterly output of twenty videos, may now produce hundreds of variants, but at the same time it needs a mechanism to phase out about ninety per cent of poor performance, otherwise the material bank will quickly lose control. The minimum coverage of the post-trading periods, i.e., retention and repurchase after purchase, also prompted the team not to put all its resources on the side of the guest.
05 GSP analysis
Who's responsible, who's tried, who's replaced.
After the ability to land, the organization of trade unions was forced to redraw the lines. (a) A reasonable division of three levels: target levels are set by marketing directors, with clear budget shares between new and repurchases, acceptable passenger cost zones, and non-touchable brand expressions; The strategy layer is carried out by people with investment and creative judgement who define test assumptions and translate the insights of AI output into action; The executive level is heavily handed over to assistants and agents, and people turn to review and exceptions. The key to this division is not the number of tiers, but the input output of each layer is written.
The duty to review is the most problematic. Meta expresses that each action is subject to the authority given by the advertiser and requires explicit approval before it is published. This is institutional protection, and in practice it becomes a rubber stamp: if there are dozens of drafts waiting to be approved every day, the reviewers will develop a "fast-pointing" habit within two weeks. There is therefore a need to introduce risk classification, to establish drafts affecting large budgets, new products or sensitive populations as mandatory reviews, and to establish regular low-risk adjustments as sample reviews.
What is really being replaced is not posts, but certain tasks. While the work of writing basic texts, being mechanically excluded and organizing weekly reports will be reduced, the responsibility for defining assumptions, judging creative directions, dealing with anomalies and explaining the results externally becomes even greater. The agent ' s value will also shift from implementation to governance and strategy, assuming, for example, that a proxy team has previously charged at the time of commissioning and may in future charge more at the quality of the test design and the output of the conclusion.
06 GSP analysis
How much to invest and what else to explain.
Inputs are calculated on a hierarchical basis. The first tier is the cost of access, including account rights combing, MCP or tool chain interfaces, data and brand regulation imports; The second tier is operating costs, including continuous testing design, review of manpower and material phase-out mechanisms; The third tier is governance costs, including approval rules, log marks and unusual processing processes. Many teams have budgeted only for the first tier, which became the main friction after six months.
There are at least three alternative interpretations of the need for inputs. First, since performance changes may arise mainly from the platform ' s competitive bidding and flow environments, which have limited relevance to these AI functions, it is important that the assessment be conducted with a control group or historical baseline, rather than a simple comparison between the top line and the bottom line. Second, the impact of capacity may be highly concentrated on larger budget advertisers, as they have sufficient data to make insight meaningful and small and medium advertisers may not feel the difference. Third, short-term efficiency gains may be at the expense of long-term capacity: if teams do not make audience and creative decisions in their own hands for a long period of time, internal judgement will deteriorate and migration costs will be high in the event of a change in platform strategy or policy.
A reasonable approach is therefore to test the scale rather than the whole: first, to validate the line of operations with a low budget ratio and a manageable cost of failure, and to explicitly use the time saved for higher value judgements rather than directly translate into more investments.
07 Suggested actions
The order in which the recommendations are carried forward
As a first step, a clear line of operations should be selected for piloting, with priority given to goods or markets with sufficient data and manageable cost of failure. The pilot objective is not to " reduce the cost" but to "comprehensiveize the record of decision-making at a certain point " , which is the basis for any subsequent judgement.
The second step is the ability to pre-empt risk with high certainty. The elimination logic, cycle task and reminder, and creative insight of Customore Lifecycle Strategey, which have a relatively direct impact on the process, fit for the first. Dialogue to create cappaign and budget changes are part of the second batch and require the establishment of rules of competence and approval to determine which actions are automatic and what the thresholds are for manual identification, amount or size of the population. The real difficulty of this step is the determination of thresholds: if set too loose, the review becomes formal; Overstretched, the efficiency of assistants is offset. It would be feasible to first classify them as “reversible” and “reversible” — budget changes and the expansion of the population to be reversible in the short term, with lower approval thresholds; The external release of new materials and the switch of new targets will make it difficult to recover the run-off and will have to be signed manually.
Step three, integrate the draft agent into the existing manual process, instead of creating a separate process. Once the draft has entered the advertising management tool, it should follow the same review path as the manual, avoiding the double standard of “AI version liberal, manual version strict”. Scenario: A brand may require that all drafts be signed by the person responsible for the placement and then entered into the material compliance check before being allowed to be published, and any movement over the signature is considered a process accident and recorded. The fourth step is to establish a phase-out mechanism for materials and audiences, with clear numerical ceilings, retention cycles and sub-standards, to prevent production capacity from turning into material accumulation. Step five, a six months back, focused on two questions: which conclusions AI helped us get faster, and which judgments we stopped doing because we depended on AI. The latter answer determines whether the next step is to continue to expand the mandate or to recover it.
08 Suggested actions
What are the indicators to judge that value is not worth continuing?
The indicators are grouped into the three categories of efficiency, quality and capacity, and a baseline is defined prior to the pilot. Efficiency categories can be seen in the preparation time for a single valid variant, the manual time of a cyclical task, and the error rate excluded by the audience; Quality categories can be seen in the rates of reuse of test findings, success rates of creative start-up, and the share of new clients in the new cappaign; The capability category is the most difficult to quantify but the most important, and it can be seen whether the team can recapitulate the core assumptions and bases of the current placement without recourse to an assistant. The three types of indicators are designed to prevent the masking of each other: efficiency gains may simply shift the cost of the review downstream, and quality improvements may simply result in the selection of a narrow group that is easier to start.
In the experimental design, it is recommended that a "man-led" control module be retained, not of a large scale, but for a sufficient period of time to identify the interference caused by changes in the platform's environment. For each new capacity, the next expected direction and approximate range are written and the reasons for the deviation are recorded against the actual results after the completion of the pilot. This makes it possible to distinguish between effective capacity and environmental change, or whether the expectations themselves are wrong. This could be done by means of a weekly spot-check of recommendations generated by a number of assistants, with an independent determination by those not involved in the project as to whether the same conclusion would be reached in the same context, and by recording the link in the points of disagreement.
Pre-requisites for cessation are required: if the rate of re-entry during the trial period continues to increase, or if the frequency of manual remediation of an unusual action exceeds the agreed threshold, the extension of the mandate should be suspended and the re-engineering of rules and competencies should be returned, rather than absorbed by additional personnel. The significance of the indicators was to help make a judgement as to whether to continue, adjust or withdraw, rather than to justify the project. Scenario: If for two months a line of business requires an individual to correct the exclusion list given by the assistant, that means that the rule itself is not clear and that at this point the added person will only cover up the problem and should go back to amending the rules before talking about expansion.
09 Suggested actions
What are the conditions? When should we turn around?
There are three prerequisites for the establishment of the route: brands and product expressions are relatively stable and can be clearly written in the rules; The business has a certain data density that allows AI to see that it is not all noise; There is a willingness within the organization to assume review and governance responsibilities rather than outsourcing decision-making entirely to the system. In the absence of one of the three articles, the momentum should be slowed. The expression of unstable boundaries implies frequent rewriting of rules and rapid decline in the reusable nature of AI output; Inadequate data densities can stop insight at the common sense level and increase the cost of interpretation; If no one is blamed, the review will degenerate into a formality.
Industry differences are also evident. High frequency, material-driven product classes benefit most from video generation and creative insight; In the low-frequency high-prisoner category, the audience may find it more profitable to exclude existing clients, while bulk video generation has limited significance. Reliance on a strong brandy or compliance-sensitive market should leave more links to mandatory review. An actionable judgement is to ask: is the cost of error a waste of budget on this line of business or a damage to branding and compliance? The former could be more handed over to the system and the latter must be left to others.
Signs that need to be turned around include a substantial change in the platform ' s strategy or tools ' terms of reference leading to an increase in the cost of migration; AI output recommendation systematically deviates from internal judgements in multiple spot checks; Or the team can no longer give an independent reason for critical decision-making. When these signals occur, the right move is to withdraw part of the authorization and re-establish internal judgement, rather than to continue to automate the additions. Capacity is a tool, and organizational ownership of judgement is a long-term asset. Assumptions: When teams are asked why they are adjusting their audience, they only answer "Associative advice so," which is in itself an early warning that should be turned around.
Research sources
- Advertising Week New York 2026: Nuevas capacidades de IA para guiar campañas y llegar a los clientes | Acerca de Meta ↗about.fb.com ·
Prepared from the cited public sources with AI-assisted research. Reported facts and editorial interpretation are presented separately.
CONTINUE THE RESEARCH
Explore your question further.
Request the full research report and discuss your market, brand questions and business context with GSP.
Request the full research report ↗