IAB Tech Lab opens agentic framework and on-device AI for publishers

The Interactive Advertising Bureau Technology Laboratory released its Agentic RTB Framework version 1.0 for public comment on November 12, 2025, while simultaneously promoting on-device artificial intelligence as a publisher-focused privacy solution through a technical blog post. The dual announcements position IAB Tech Lab at the intersection of containerized agent deployment and on-device processing technologies that could reshape programmatic advertising infrastructure.

The Agentic RTB Framework specification v1.0 enters a public comment period extending through January 15, 2026. The framework defines standardized requirements for deploying agent-driven containers within OpenRTB environments, addressing how containerized agents can participate in real-time bidding infrastructure with minimal latency impacts. The specification establishes technical protocols for container runtime behavior and defines application programming interfaces for bidstream mutation, enabling delegation of processing tasks to autonomous agents operating within host platforms.

Miguel Morales, Director of Addressability & Privacy Enhancing Technologies at IAB Tech Lab, authored the framework announcement. The Container Project Working Group developed the specification through collaboration with Index Exchange, OpenX, The Trade Desk, and Chalice. Participating organizations include Amazon Ads, Netflix, Yahoo, Paramount, Optable, HUMAN Security, Magnite, PubMatic, WPP Media, and Basis Technologies.

Concurrent with the framework release, IAB Tech Lab published a blog post on November 11 featuring analysis from Anish Aravindakshan of Verve examining on-device cohort modeling. The post argues that on-device artificial intelligence represents an untapped opportunity for publishers seeking privacy-compliant targeting methodologies. On-device cohort modeling processes audience segmentation directly on user devices rather than transmitting individual-level data to external servers, respecting platform privacy norms without relying on third-party cookies.

How the technology works

Understanding these technologies requires examining two distinct approaches to advertising automation. The concepts differ substantially but both address fundamental challenges in digital advertising: speed, privacy, and automation.

Containerized agents function similarly to specialized workers dropped into an assembly line. Traditional programmatic advertising requires companies to build custom connections between their systems and every advertising platform. A fraud detection company wanting to check advertisements across 50 different platforms would need to build 50 separate integrations. Containerized agents eliminate this complexity. The agent arrives as a self-contained package carrying everything needed to perform its job—fraud detection algorithms, brand safety rules, or data enrichment logic. The hosting platform provides the workspace and tools, while the agent performs its specialized task. When an advertisement opportunity arises, the container springs into action, analyzes the situation in milliseconds, and returns its verdict without slowing down the auction. The standardized container format means the same fraud detection agent works identically whether deployed on a supply-side platform, demand-side platform, or publisher server.

On-device cohort modeling operates through an entirely different mechanism focused on privacy preservation. Current advertising typically works like this: websites and applications transmit user behavior data to central servers where algorithms analyze millions of users simultaneously, identifying patterns and assigning advertisements. On-device processing inverts this model. The analysis happens on the smartphone, tablet, or television itself. A news application on a smartphone might notice its owner reads automotive articles frequently, watches home improvement videos, and engages with travel content. Instead of sending this detailed behavioral history to external servers, the device's processor runs a lightweight algorithm determining the owner probably belongs to groups interested in cars, home goods, and vacation packages. Only these broad group memberships—cohorts—leave the device. Advertisers receive a signal indicating "this device owner belongs to the automotive enthusiast cohort" without learning which specific articles were read, when they were accessed, or any identifying details about the person. The phone does the thinking, keeps the sensitive details local, and shares only the conclusions.

The technical elegance lies in matching processing power to privacy requirements. Modern smartphones contain processors more powerful than desktop computers from a decade ago. This computational capability previously sat idle between application launches. On-device processing harnesses that dormant power for privacy-preserving analysis. Machine learning models small enough to fit on consumer devices can identify interests and preferences by examining local application usage patterns. A streaming television device knows which shows its owner watches, which genres get abandoned mid-episode, and when viewing typically occurs. Processing this information locally produces cohort assignments—"evening drama viewer," "sports enthusiast," "family content consumer"—without transmitting the actual viewing history beyond the device. Publishers benefit because they control the algorithm determining cohort assignments. A news publisher could develop proprietary interest detection superior to generic third-party solutions, differentiating their inventory through better audience understanding while maintaining user privacy.

Both technologies address bottlenecks in programmatic advertising infrastructure. Containerized agents solve the integration complexity problem: how do dozens of specialized service providers efficiently connect to hundreds of advertising platforms without building thousands of custom connections? On-device processing solves the privacy paradox: how do publishers monetize through targeted advertising when platforms prohibit tracking individual users across applications and regulators restrict personal data collection? The solutions operate at different points in the advertising workflow but pursue the same goal—maintaining programmatic advertising effectiveness as technical and regulatory constraints eliminate previous methodologies.

Measuring performance in opaque systems

The shift toward autonomous agents and on-device processing creates measurement challenges absent from traditional advertising infrastructure. Advertisers accustomed to granular performance data face a fundamental question: how do you verify effectiveness when artificial intelligence makes decisions autonomously and privacy protections obscure the mechanisms producing results?

Traditional measurement relied on observable inputs and outputs. An advertiser could examine which keywords triggered advertisements, which audience segments received impressions, what creative variations performed best, and how bidding strategies affected costs. Transparency enabled optimization. Campaign managers adjusted based on clear cause-and-effect relationships visible in reporting interfaces. Autonomous agents disrupt this model. When an AI system decides which inventory to purchase, which audiences to target, and what bids to submit—all within milliseconds and across millions of decisions daily—human observation becomes impractical. The question shifts from "what did the system do?" to "did the system achieve desired outcomes?"

Outcome-based measurement focuses on business results rather than process metrics. An advertiser running campaigns through autonomous agents cares whether sales increased, cost per acquisition declined, or return on advertising spend improved. The specific bid adjustments, audience expansions, or inventory selections that produced those results matter less than the aggregate performance. This approach mirrors how advertisers evaluate human media buyers. Nobody demands minute-by-minute documentation of every decision a campaign manager makes. Performance reviews examine results: did the campaigns hit targets? Autonomous agents receive similar evaluation based on outcomes rather than granular decision logs.

On-device processing introduces different measurement constraints. Cohort assignments happen locally without transmitting the underlying behavioral data that informed those assignments. An advertiser knows an impression reached the "automotive enthusiast" cohort but cannot verify how the device determined that classification. Was the cohort assignment accurate? Did the algorithm properly identify genuine interest signals or misinterpret unrelated behaviors? Validation becomes probabilistic rather than deterministic. Advertisers must infer cohort quality from campaign performance. If automotive advertisements delivered to the "automotive enthusiast" cohort generate strong conversion rates, the cohort assignment process probably works correctly. Poor performance suggests either faulty cohort logic or misaligned creative messaging.

Control groups provide one verification methodology. Advertisers can compare campaigns using on-device cohorts against campaigns using traditional targeting methods or no targeting at all. A statistically significant performance improvement demonstrates that on-device cohorts add value regardless of the specific algorithms producing those cohorts. Incrementality testing measures the causal impact of advertising exposure by randomly withholding advertisements from a portion of the target audience, then comparing outcomes between exposed and unexposed groups. The methodology works identically whether cohorts derive from server-side processing or on-device algorithms. Results indicate whether the targeting approach effectively reaches responsive audiences.

Aggregated reporting replaces individual-level tracking in privacy-preserving systems. Instead of detailed user journeys showing exactly which advertisements a specific person saw before purchasing, advertisers receive summary statistics indicating how many people from each cohort converted within various time windows. The aggregation preserves privacy while enabling performance analysis. An advertiser might learn that 3.2 percent of impressions delivered to the "home improvement enthusiast" cohort resulted in conversions within 30 days, compared to 1.8 percent for the "general interest" cohort. This data suffices for budget allocation decisions without requiring individual user tracking.

Attribution modeling becomes more challenging when autonomous agents control campaign execution across multiple platforms simultaneously. Traditional attribution assigns credit for conversions to specific touchpoints in the customer journey. Multi-touch attribution might credit 40 percent to the initial awareness advertisement, 30 percent to mid-funnel consideration content, and 30 percent to the final retargeting impression. Autonomous agents managing campaigns holistically across channels complicate this analysis. The agent might simultaneously adjust search bids, expand social media targeting, and increase video advertisement frequency. Isolating which specific action drove incremental conversions becomes difficult when the agent makes interdependent decisions treating the campaign as an integrated system rather than discrete channels.

Media mix modeling offers a solution for measuring agent-driven campaigns. This statistical approach analyzes historical relationships between advertising investments and business outcomes, accounting for seasonality, competitive activity, and external factors. The models estimate how much each marketing channel contributed to overall results without requiring individual-level attribution. An advertiser using autonomous agents across search, social, display, and video can employ media mix modeling to understand channel effectiveness and optimal budget allocation. The approach treats agent-controlled campaigns as inputs to the model, evaluating aggregate performance rather than dissecting individual decisions.

Benchmark comparisons establish performance expectations for AI-driven systems. An advertiser implementing on-device cohort targeting for the first time lacks reference points for evaluating results. Is a 2.5 percent conversion rate good or poor for that cohort methodology? Industry benchmarks provide context. If similar advertisers achieve 2.8 percent conversion rates using comparable approaches, the 2.5 percent result suggests room for improvement. If the industry average sits at 1.9 percent, the 2.5 percent performance indicates effective implementation. Benchmarking requires careful comparison. Cohort definitions, creative quality, offer competitiveness, and seasonal timing all affect results. Direct comparisons work best when advertisers share detailed implementation approaches, though competitive concerns often limit transparency.

Certification and audit frameworks could address trust deficits in opaque AI systems. IAB Tech Lab's framework development includes requirements for agents to declare their intents and modifications. This transparency enables verification that agents behave as specified. An advertiser deploying a brand safety agent can audit whether the agent actually filters inventory according to stated criteria or allows violations through. Independent auditors might certify that on-device processing implementations genuinely keep data local rather than transmitting detailed user information to external servers despite privacy claims. Verification requires technical capabilities beyond most advertisers' internal resources, creating opportunities for specialized audit services.

The measurement challenge reflects broader tensions between automation efficiency and operational transparency. Autonomous systems promise superior performance through processing capabilities exceeding human analysis. Machine learning algorithms identify patterns in billions of data points, optimizing toward objectives faster than manual campaign management. This capability requires delegating decision authority to the AI system. Excessive constraints or reporting requirements undermine the efficiency gains motivating automation adoption. Advertisers must balance the desire for transparent, explainable decisions against the performance advantages of allowing AI systems substantial autonomy. Different organizations will strike different balances based on risk tolerance, regulatory requirements, and competitive positioning.

The on-device approach contrasts with server-side processing models that have dominated programmatic advertising. Traditional methods aggregate user data on centralized servers where targeting decisions occur before advertisement delivery. On-device processing keeps sensitive user information local to the device while enabling contextual and behavioral targeting through cohort assignments computed directly on smartphones, tablets, or connected television devices.

Technical implementation requires processing power sufficient to execute machine learning models on consumer hardware. Modern mobile devices and streaming devices possess computational capabilities that were previously available only in data center environments. Graphics processing units embedded in smartphones enable real-time inference from neural networks trained to identify user interests and purchasing intent based on application usage patterns, content consumption, and device activity.

The cohort modeling methodology groups users with similar characteristics into segments without revealing individual identities. A device might determine that its owner belongs to cohorts interested in outdoor recreation, home improvement, and premium automotive products based on application interactions and content engagement. These cohort assignments become available to advertising systems without transmitting personally identifiable information beyond the device, maintaining privacy while enabling relevant advertisement delivery.

Platform operators including Apple and Google have implemented privacy frameworks restricting cross-application tracking and limiting data collection practices. Apple's App Tracking Transparency requires explicit user permission for cross-app tracking, resulting in opt-in rates below 25 percent among iOS users. Google announced plans to deprecate third-party cookies in Chrome multiple times before halting implementation in July 2024, creating uncertainty around alternative targeting methodologies.

On-device processing offers publishers control over data processing while complying with platform restrictions. Publishers can implement proprietary algorithms that analyze user behavior within their applications or websites, generating cohort assignments that inform advertisement selection without sharing raw behavioral data with advertisers or intermediaries. This approach preserves the publisher's data asset while enabling programmatic monetization.

The Agentic RTB Framework addresses different technical challenges related to autonomous decision-making within bidstream processing. Agents deployed through the framework can modify bid requests, enhance data signals, apply brand safety rules, or execute fraud detection algorithms. Container-based deployment enables independent software vendors to distribute specialized processing logic that integrates into supply-side platforms and demand-side platforms without requiring custom integration work.

Five critical requirements govern agent deployment under the framework. Agents must participate in the core bidstream, focusing on entities transacting in real-time auctions. Each agent must declare specific intents and any auction changes, allowing orchestrating entities to accept or reject modifications. Container structure follows Open Container Initiative compliance, manageable through Kubernetes, Docker Compose, or cloud-based systems including Amazon's Elastic Container Service. Performance architecture achieves low to sub-millisecond response times through container-based deployment dropped into host networks. Containers communicate through high-performance messaging mechanisms built on gRPC, OpenRTB Patch, and Model Context Protocol compatible with real-time bidding requirements.

IAB Tech Lab's framework development follows the organization's 2025 technical standards roadmap announced January 29, which outlined 31 new specifications or updates addressing privacy regulations, data handling, and streaming media challenges. The organization completed 79 initiatives in 2024, developed with input from over 800 member companies.

The emphasis on publisher benefits in the on-device artificial intelligence blog post reflects broader industry concerns about publisher monetization amid platform changes. IAB Tech Lab launched its Content Monetization Protocols working group on August 20, 2025, responding to evidence that artificial intelligence technologies threaten digital publishing economic sustainability. AI-driven search summaries reduce publisher traffic by 20-60 percent on average, with niche sites experiencing losses up to 90 percent.

Anthony Katsur, Chief Executive Officer at IAB Tech Lab, emphasized the organization's role in technical standards development. The standards body previously challenged Prebid transaction ID implementation changes in August 2025, declaring that modifications materially violated OpenRTB specifications and risked "undermining the integrity and consistency of open technical standards." That enforcement action followed established procedures for technical specification evolution.

The convergence of containerized agent deployment and on-device processing represents parallel approaches to advertising technology challenges. Containerized agents enable centralized processing logic deployment across distributed infrastructure, while on-device processing distributes computation to edge devices. Both methodologies address scalability, latency, and privacy requirements that constrain traditional server-based architectures.

Industry movement toward agentic artificial intelligence has accelerated throughout 2025. McKinsey data indicates $1.1 billion in equity investment flowed into agentic AI during 2024, with job postings related to the technology increasing 985 percent from 2023 to 2024. Six companies launched Ad Context Protocol on October 15, betting that open-source technical standards could enable AI agents to communicate across platforms and execute advertising tasks autonomously.

Amazon launched agentic capabilities across its advertising platform on November 11, transforming tools from question-answering systems into autonomous agents that monitor accounts, optimize inventory, and manage campaigns. The system processes natural language instructions to execute complex workflows including campaign creation, audience targeting, and analytics query generation.

The Agentic RTB Framework specification includes sample code in the IAB Tech Lab GitHub repository. Organizations interested in reviewing technical specifications and providing feedback can access documentation through the IAB Tech Lab website. The public comment period concluding January 15, 2026, provides opportunity for industry participants to contribute to the refinement of deployment standards.

IAB Tech Lab previously released Publisher Advertiser Identity Reconciliation protocol in September 2024 for first-party data matching, Attribution Data Matching Protocol in October 2024 for privacy-preserving conversion measurement, and ID-Less Solutions Guidance in July 2025 for advertising without traditional identifiers. The framework development parallels the organization's Content Monetization Protocols working group addressing AI-driven search impacts on publisher revenue.

The technical specifications matter for marketing professionals managing programmatic advertising operations. Containerized agents could automate brand safety enforcement, fraud detection, and data enrichment tasks currently handled through manual processes or platform-specific implementations. Standardized deployment enables interoperability across supply-side platforms and demand-side platforms, reducing integration complexity for independent software vendors developing specialized processing logic.

On-device processing addresses privacy compliance requirements while maintaining targeting capabilities necessary for programmatic monetization. Publishers implementing on-device cohort modeling can differentiate their inventory through proprietary audience insights unavailable to competitors relying on third-party data providers. The methodology aligns with regulatory frameworks in jurisdictions including the European Union, California, and other privacy-conscious markets restricting personal data collection and cross-context tracking.

Implementation challenges include computational requirements, model distribution, and measurement verification. Processing machine learning models on consumer devices consumes battery power and memory resources that device manufacturers and operating system providers may restrict. Distributing updated models to millions of devices creates bandwidth and version control complexity. Verifying that on-device processing occurs as declared rather than transmitting raw data to servers requires attestation mechanisms and audit capabilities.

The IAB Tech Lab announcements position publishers at the center of privacy-preserving advertising technology evolution. On-device processing gives publishers control over data processing while maintaining user privacy. Containerized agents enable deployment of publisher-specific logic across programmatic infrastructure. Both approaches respond to platform restrictions and regulatory requirements reshaping digital advertising technical architecture.

Timeline

• November 11, 2025: IAB Tech Lab publishes blog post featuring Verve's analysis of on-device cohort modeling for publishers
• November 12, 2025: IAB Tech Lab releases Agentic RTB Framework version 1.0 for public comment with sample code in GitHub repository
• January 15, 2026: Public comment period concludes for Agentic RTB Framework specification
• January 29, 2025: IAB Tech Lab announces 2025 roadmap outlining 31 new specifications addressing privacy and streaming media
• July 17, 2025: IAB Tech Lab releases ID-Less Solutions Guidance for advertising without traditional identifiers
• August 20, 2025: IAB Tech Lab launches Content Monetization Protocols working group addressing AI-driven search impacts
• August 27, 2025: IAB Tech Lab challenges Prebid transaction ID changes declaring OpenRTB specification violations
• September 26, 2024: IAB Tech Lab launches PAIR protocol for first-party data matching
• October 15, 2024: IAB Tech Lab releases Attribution Data Matching Protocol for privacy-preserving measurement

Summary

Who: The Interactive Advertising Bureau Technology Laboratory, a nonprofit consortium developing digital advertising technical standards, released specifications developed by its Container Project Working Group. Anish Aravindakshan from Verve contributed analysis on on-device cohort modeling published in the IAB Tech Lab blog.

What: IAB Tech Lab released Agentic RTB Framework version 1.0 for public comment, establishing standardized specifications for deploying containerized agents within real-time bidding infrastructure. Simultaneously, the organization published analysis promoting on-device artificial intelligence for publisher-focused privacy-compliant audience targeting without third-party cookies.

When: The Agentic RTB Framework specification was released on November 12, 2025, with a public comment period extending through January 15, 2026. The on-device AI blog post was published on November 11, 2025.

Where: The specifications affect global programmatic advertising infrastructure, with participating organizations including Index Exchange, OpenX, The Trade Desk, Amazon Ads, Netflix, Yahoo, Paramount, Optable, HUMAN Security, Magnite, PubMatic, WPP Media, and Basis Technologies. Implementation targets supply-side platforms, demand-side platforms, and publisher applications operating across mobile, connected television, and web environments.

Why: The framework addresses the need for standardized containerized agent deployment as agentic artificial intelligence adoption accelerates across advertising technology, with $1.1 billion in equity investment during 2024. On-device processing responds to platform privacy restrictions and regulatory requirements limiting third-party cookies and cross-application tracking, offering publishers privacy-compliant targeting methodologies while maintaining programmatic monetization capabilities amid traffic declines from AI-driven search.