Media buying has relied on delegated decisions for years: traders define objectives, platforms bid at machine speed, and agencies manage execution. AI agents extend that delegation into planning and supplier selection. The business question is which choices were made, what evidence explains them, and who is accountable for their quality.
An advertising agent may interpret a brief, compare proposed inventory, configure a deal, route spend and adjust a campaign. It can follow its permissions perfectly and still make choices the buyer would want examined. A paid invoice proves a purchase occurred. A compliant campaign report can prove that selected rules were followed. Neither necessarily explains why one economically relevant route was preferred over another.
That distinction matters more as advertising adds agent-to-agent workflows. It is also a reason to respect, rather than dismiss, the auditing and verification systems buyers already use.
Agent-led media buying is already underway
The World Federation of Advertisers' September 2026 survey of 54 senior respondents at 46 major brand owners found that 96% now use generative or agentic AI somewhere in marketing. The headline is broad: it covers content creation, research and workflow tools as well as buying. Within the same survey, 54% reported using AI for media buying or optimization and 46% reported workflow automation or autonomous-agent uses, while 75% required human oversight of AI-generated outputs. The survey is not evidence that most media spending has been delegated to independent buying agents.[1]
There are narrower examples of execution. In December 2025, PubMatic and independent agency Butler/Till announced an agent-led connected-TV campaign for Geloso Beverage Group's Clubtails. The agency supplied a natural-language brief and PubMatic's system interpreted it and carried out planning and activation within defined parameters.[2] A subsequent vendor-hosted case study reports lower costs and additional impressions using Butler/Till internal December 2025 data. It is a vendor-reported case study, not an independent audit.[3]
PubMatic separately said in April 2026 that 30 live end-to-end campaigns were operating on AgenticOS; an August 2026 company release described a configurable governance architecture and identified Rise, a Quad agency, among its early pilot partners.[4] [5] These are company-reported deployment counts and a named governance pilot, not third-party verification of buyers' decision independence. The Trade Desk's August 2026 Koa Agents update, by contrast, calls its conversational experience a closed beta; campaign creation and troubleshooting assistance should not be described as full autonomous media purchasing.[6]
Infrastructure is advancing too. In September 2026 IAB Tech Lab introduced AAMP 3.0 and OpenProposal to standardize the path from request for proposal to buy commitment. A proposed technical framework can make interactions easier to record and govern, but it is not proof of broad adoption, correct commercial choices, or certification of independent judgment.[7]
Five questions hiding inside "was this a valid buy?"
Consider a fictional advertiser authorizing a six-week campaign. The agency may use automation, provided it stays within a budget, excludes certain content, favors approved sellers and meets a measurement requirement. The agent selects one eligible supply path repeatedly. Spend, impressions and suitability checks all reconcile.
That is a successful outcome for several kinds of control. It still leaves five distinct questions.
Authority: Who approved the objective, budget, markets, inventory restrictions and degree of autonomy? Who may change them? Authorization evidence includes the operative brief, delegated permissions, human approvals and versioned exceptions. It establishes permission, not quality of judgment.
Execution and compliance: Which instructions were in force when the agent acted, and what events show it obeyed them? Versioned configuration, change logs, action records and exception reports can establish rule compliance within their recorded scope. An agent's after-the-fact prose explanation is not a substitute for contemporaneous evidence.
Transaction and financial reconciliation: What was requested, bid, won, delivered, billed and paid? Agency, buying-platform, seller and finance records can be reconciled, subject to matching rights and missing data. This tests what happened and where money went; it does not necessarily test all options the agent could have selected.
Supply-path and media-quality verification: Was the seller authorized? Which intermediaries were disclosed? Were placements measurable, viewable, legitimate and suitable for the brand? Ads.txt, sellers.json and supply-chain disclosures serve different provenance purposes.[8] Independent measurement providers check delivered-media quality.[9] They are valuable controls, but a verified seller is not automatically the most efficient seller.
Commercial judgment and independence: Did the process consider materially different available routes, or did common defaults, incentives, data or optimization services narrow the choice? Were apparent alternatives actually eligible and exposed to the buyer? This is the question existing controls are least designed to answer, because invoices, logs and quality checks record what was chosen, not what was never offered. Similar choices do not establish a shared controller or coordination, and separate corporate identities do not establish distinct economic decision policies. The records may stop short of an answer.
The last question is conditional. Some buying systems will generate sufficiently complete, reproducible decision evidence for the specific buyer question. Others will not. There is no universal "agent independence" certificate that can be inferred from an invoice or a signed activity log.
A clean audit trail can still tell an incomplete story
Imagine that a buyer's approved seller list contains three exchanges. An agent's log shows it considered three routes, then selected the cheapest quoted option. A reconciliation confirms the price and delivered impressions. The advertiser might reasonably sign off.
But suppose the "three" routes were actually three reseller representations of the same underlying inventory, or one route was excluded by an upstream filter the agent never saw. Alternatively, a vendor's private incentive or a common optimizer could have shaped the offered menu. These are hypothetical failure modes, not allegations against a named provider.
Even a good decision log can establish only what its recording system observed. A candidate absent from that system's opportunity set cannot be treated as a deliberately rejected candidate. Another limitation runs the opposite way: two independently controlled tools may pick the same placement because its price, quality and expected value genuinely are better. Uniform behavior is a prompt for a question, not a finding of misconduct.
This is why a buyer should separate the decision actually taken, the menu of genuinely available choices, and the reason those choices were available. Completing that chain may require permissioned records from more than one organization. If they are unavailable, the defensible answer may be "not determinable from these records," not "independent" or "compromised."
Seller-identity disclosures, supply-chain records and supply-path optimization can help buyers identify overlapping routes to inventory. They cannot establish whether an upstream filter removed an alternative before the agent saw it, or whether a shared optimizer shaped the choices presented. That is the gap Exchange Optics is researching: how to assess whether the available choices were meaningfully distinct, and whether the accessible evidence can support that conclusion.
The industry already has meaningful answers to other questions
Advertisers are not starting from zero. The Association of National Advertisers' Q1 2026 programmatic benchmark reported that higher-performing participants converted 54.0% of spend into qualified impressions, versus 32.1% for the lower-performing cohort. These are benchmark-defined measures of fraud-free, measurable, viewable and made-for-advertising-free delivery—not measures of the independence of automated buying decisions. The gap nonetheless demonstrates the practical value of established quality and supply management.[10]
ISBA's 2024 Programmatic Audit Permission Letter Toolkit deals with a different constraint: obtaining permissions across the supply chain to audit financial transactions. Without contractual access, even a well-framed independent question may be unanswerable.[11] Independent media-quality providers likewise verify specific delivered-ad properties. DoubleVerify's May 2026 announcement of post-bid LinkedIn Audience Network measurement, for example, covers invalid traffic, viewability, brand suitability and geography—not a full causal account of why a buyer selected a given route.[9]
Agency stewardship, spend and invoice reconciliation, supply-path optimization, verification partners, contractual rights, internal controls and specialist forensic audits can already be sufficient. Buying another review makes sense only if the unresolved question is material, observable, legally accessible and not covered by existing assurance.
What should an advertiser actually require?
Before expanding an agent's authority, the advertiser, agency and platforms should agree on a modest evidence contract. The exact controls will vary by channel and commercial agreement, but the ownership should be intelligible:
- Advertiser: Define the campaign's economic objective, permissible autonomy, required protections, acceptable trade-offs and who may approve changes. Decide which outcomes justify investigation.
- Agency or buying principal: Translate the mandate into operating limits; disclose material automation and conflicts; preserve versions, approvals, exceptions and the basis for client reporting. Escalate changes that exceed delegated authority.
- Agent and platform operator: Enforce permissions, record consequential actions and failures, identify material upstream selection boundaries, and provide the permitted exports and reconciliations. Explain what the logs do not cover.
- Independent verification or audit provider, when warranted: Test the specifically contracted claim against admissible evidence, name contradictory explanations, report limitations and retain the right to return an unfavorable or unresolved result. That is easier to trust when the provider's fee does not depend on the outcome; see Research 03. Independence must be assessed, not presumed from an outside logo.
For a buyer reviewing a proposed agentic campaign, useful questions are practical: Which human approves the initial mandate? Can an optimizer alter the seller list, objective or fees without renewed authority? Are agents acting for the advertiser, the agency, the marketplace—or several parties with competing incentives? Which record shows an instruction changed? Who reconciles spend against delivery? What evidence could disprove the vendor's explanation? Who receives a negative review?
WFA's 2026 survey reported that 69% of respondents require agencies to disclose AI use and 59% have AI-related contractual clauses. But only 13% reported a great deal of visibility into agency AI use. Those are self-reported governance figures for a defined survey sample—not proof that the contracts work or that autonomous buying is pervasive.[1]
When independent decision verification changes a buying decision
A buyer may decide that its agency's controls and independent quality verification are enough. That can be an entirely sound decision. Some media buys are small, highly constrained or reversible. Human traders still supervise many workflows. An exacting independent study may cost more, expose sensitive vendor data or create privacy and contractual problems without improving a commercial decision.
The stronger case arises when an automated strategy controls consequential spend, material supplier choices are opaque, paths repeatedly converge for unexplained reasons, or a buyer must decide whether to extend authority to a system it cannot effectively challenge. No amount of attractive reporting can reconstruct choices that were never recorded or never available.
The buyer's standard should therefore be proportionate: verify authorization; reconcile execution; check media quality and supply provenance; investigate decision selection only where the incremental question matters and the evidence can support it. An honest inability to determine the answer is itself useful governance information.
The person accountable for a media investment does not disappear when an agent places the order. What changes is the evidence that person should be entitled to inspect—and the discipline required to distinguish a documented transaction from an independently defensible decision.
Exchange Optics is seeking Founding Design Partners to investigate what automated-buying records can—and cannot—establish. If you are extending an agent's buying authority and want to explore a bounded source-readiness assessment, get in touch.
References
[1] World Federation of Advertisers, "96% of major brands now use AI in marketing, WFA research finds," September 30, 2026. https://wfanet.org/knowledge/item/2026/09/30/96-of-major-brands-now-use-ai-in-marketing-wfa-research-finds
[2] PubMatic and Butler/Till, "PubMatic and Butler/Till Launch AdCP-Enabled Agentic AI Campaign Across CTV," December 15, 2025. https://pubmatic.com/news/pubmatic-and-butler-till-launch-adcp-enabled-agentic-ai-campaign-across-ctv/
[3] PubMatic, "Butler/Till Unlocks Greater Efficiency and Performance with PubMatic AgenticOS," company case study, source data labeled December 2025, webpage undated. https://pubmatic.com/case-studies/butler-till-unlocks-greater-efficiency-and-performance-with-pubmatic-agenticos/
[4] PubMatic, "PubMatic's AgenticOS Accelerates Globally as Agentic Campaigns Unlock Efficiency and Performance," April 27, 2026. https://pubmatic.com/news/pubmatics-agenticos-accelerates-globally-as-agentic-campaigns-unlock-efficiency-and-performance/
[5] PubMatic, "PubMatic Launches Governance Architecture for Agentic Advertising on AgenticOS," August 5, 2026. https://pubmatic.com/news/pubmatic-launches-governance-architecture-for-agentic-advertising-on-agenticos/
[6] The Trade Desk, "Experience the next generation of AI assistance with Koa Agents," August 21, 2026. https://www.thetradedesk.com/resources/experience-the-next-generation-of-ai-assistance-with-koa-agents
[7] IAB Tech Lab, "IAB Tech Lab Introduces AAMP 3.0 to Standardize the RFP-to-Buy Process for Agentic Advertising," September 22, 2026. https://iabtechlab.com/press-releases/iab-tech-lab-introduces-aamp-3-0-with-openproposal/
[8] IAB Tech Lab, "Authorized Digital Sellers with ads.txt and app-ads.txt," updated September 14, 2026. https://iabtechlab.com/ads-txt/
[9] DoubleVerify, "DoubleVerify Delivers Global Media Quality Measurement for LinkedIn Audience Network," May 21, 2026. https://doubleverify.com/company/newsroom/doubleverify-delivers-industry-first-global-media-quality-measurement-for-linkedin-audience-network-elevating-transparency-for-b2b-advertisers
[10] Association of National Advertisers, "ANA Programmatic Transparency Benchmark Reveals Widening Gap Between Top and Bottom Advertisers," May 27, 2026. https://www.ana.net/content/show/id/pr-2026-05-programmatic
[11] ISBA, "Financial Audit Toolkit – programmatic advertising," September 11, 2024. https://www.isba.org.uk/knowledge/financial-audit-toolkit-programmatic-advertising