Spending authority requires rules about evidence

An advertising agent can gain permission to move millions of dollars before a team defines what evidence should justify its next budget increase. Agents already use corporate cards, distribute budgets and make purchases with fewer transaction-by-transaction human approvals. In agentic buying, software takes responsibility for purchasing and optimization. Giving it authority to commit more money makes the quality of its evidence a question of budget governance.

The infrastructure supporting that authority is developing. IAB Tech Lab’s Agentic Advertising Management Protocols, or AAMP, address buyer and seller agents, registries, interoperability, the ability of systems to work together, and trust. IAB’s 2026 Outlook found that, among respondents familiar with agentic AI for ad buying, 37% expected a hybrid deployment model and 10% planned entirely in-house development. These deployment choices require buying teams and technology providers to divide responsibility for the agent’s decisions.

That division must include deciding which evidence deserves confidence. AAMP can support exchanges between buyer and seller agents, but exchanging information does not establish that a participant or decision deserves trust. Experienced buyers make judgments that may never have become written policy, and software cannot be assumed to inherit them. Teams need to make those judgments explicit before an agent’s spending authority grows beyond their ability to assess its purchases.

Early performance can arrive before quality evidence

The evidence problem begins with the information used in programmatic buying, where software purchases advertising opportunities. Each decision combines information about an impression, an ad display, with information about the user and the opportunity’s origin or route through intermediaries. Some information is measured directly, some is inferred, and some passes through several companies before reaching the buyer. A spending policy needs to account for these differences in reliability.

These differences have affected buying decisions for years, but automation leaves less time to examine them. An agent can receive a promising result and raise spending before a person investigates how the result arose. Experienced buyers often compensate by recognizing providers whose measurements overstate performance, weighting metrics differently and questioning unusually clean or consistent numbers. Their apparent intuition is accumulated pattern recognition, which must become explicit requirements for an agent to use it.

Consider a hypothetical sequence in which that judgment is absent. A traffic source reports a favorable cost per acquisition, or CPA, a strong click-through rate, or CTR, and completed conversions. Because the agent’s objective is to improve acquisition performance, those results support giving the source more budget. The optimization logic can therefore work as intended while relying on an incomplete picture of quality.

That picture remains incomplete because some relevant observations arrive later. Quality checks may examine whether device characteristics fit together, whether behavior is consistent, what happens after a click and whether conversions remain valid after review. Some checks need activity that has yet to occur when the initial acquisition result appears. An experienced buyer may hold the budget steady while those observations develop, whereas an agent allowed to expand on early performance may already have increased spending.

This is premature expansion: more money is committed while traffic authenticity and conversion validity remain unresolved. A capable detector still needs observations to judge. Some checks can inform a purchase quickly, while others require a longer observation period, so the timing that matters is the gap between the expansion trigger and the evidence needed to justify it. Better eventual detection cannot protect an earlier decision unless the buying system also controls when more budget becomes available.

The financial importance of measurement is already visible in ANA’s Q1 2026 Programmatic Transparency Benchmark. Higher-performing advertisers directed 54% of programmatic spending to qualified impressions, compared with 32.1% among lower-performing advertisers. ANA attributed part of that difference to stronger measurement and more concentrated supply chains. The benchmark compares the share of spending reaching qualified impressions; its findings concern the buying conditions into which agents are being introduced.

Under those conditions, early performance can release more budget while quality remains unresolved, increasing the agent’s financial commitment before the team can assess whether the result warrants that confidence. Keeping expansion conditional gives the required observations time to arrive. But waiting helps only if the later evidence adds information beyond what the agent already knows.

Okoone experts
LET'S TALK!

A project in mind?
Schedule a 30-minute meeting with us.

Senior experts helping you move faster across product, engineering, cloud & AI.

Please enter a valid business email address.

Corroboration depends on independent measurement

Later evidence can repeat an earlier observation even when it comes through a different provider. Adding fraud scores, quality services and reporting interfaces gives an agent more information, but several providers may depend on the same upstream dataset. Their agreement can therefore reflect a shared input. A system that counts each provider as separate corroboration will place too much confidence in that agreement.

Imagine three providers judging a traffic source acceptable while all three use the same underlying dataset. The reports deliver the same evidence through several channels, so treating them as independent confirmations adds confidence without adding materially different knowledge. Procurement and implementation teams therefore need to establish how each conclusion was reached. Provider count alone cannot answer that question.

To answer it, teams need provenance: information about where data came from and how it was produced. W3C describes provenance in terms of the entities and activities responsible for producing data, which allows its trustworthiness to be assessed. An agent needs that context alongside an input to decide how much weight to give it. A score separated from its origin and method leaves the system unable to evaluate the basis for the conclusion.

The label “Viewable,” meaning an ad met a standard for visibility on screen, illustrates what this context must explain. Before the label affects spending, the team needs to know who measured viewability, which method they used and which impression path the measurement covers. These details establish what the label means for the purchase. When an agent can act without seeking approval, information about the measurement becomes part of the evidence authorizing the spend.

Even with that information, participants can disagree about methodology. In 2025, research firm Adalytics and verification vendors DoubleVerify, IAS and HUMAN publicly disputed methods for classifying General Invalid Traffic, a category used to identify invalid traffic. The verification vendors have a commercial interest in buyers trusting their measurements, so their positions carry that stake. DoubleVerify filed a defamation suit, which a US federal court allowed to proceed in May 2026; that procedural decision did not determine whose measurement method was correct.

The dispute concerns measurement methods, whereas ANA’s benchmark concerns the economic outcomes associated with measurement and supply-chain practices. The reason to control an agent’s expansion follows from the buying sequence itself: spending can increase before the evidence needed to justify it becomes available. When teams specify that evidence, they also need to distinguish performance measures from quality checks. CPA, CTR and conversion rate, or CR, describe related aspects of performance; corroboration requires a relevant observation produced through a different measurement path.

Turn evidence requirements into spending controls

Once teams have defined useful evidence, they need to specify how it changes the agent’s authority. Before granting more autonomy, buying teams must decide which inputs are usable, which claims require corroboration before spending and which details must be retained for later review. These decisions set the limits within which the agent may optimize. If they remain implicit, the model must choose its own standard of proof while pursuing its spending objective.

Enforcing those limits requires agreement across buying, implementation, procurement and budget governance. Procurement must select providers whose data can meet the requirements, and implementation teams must make those data available when decisions occur. Budget governance must then ensure that the restrictions operate as rules the agent cannot waive. Four controls turn those shared responsibilities into operating requirements; they govern authority without promising higher returns.

Cap expansion until required checks resolve

The first control addresses the delay between an early result and the quality evidence needed to assess it. Keep each source within a spending limit until the required checks have had time to resolve. Strong initial performance should leave the cap in force while those conditions remain unmet. The earliest permitted expansion should follow the slowest relevant quality input, because that input determines when the full requirement can be evaluated.

Within the cap, the agent can continue buying, optimizing and negotiating at machine speed. The team must tie the waiting period to the measurements needed for that source, so an unrelated favorable result cannot release more money. This separates ongoing execution from permission to enlarge the allocation. The agent can improve performance within its current authority while the evidence for greater authority develops.

Require an independently measured quality gate

A spending cap becomes useful only when the team defines what can release it. Require at least one independently measured quality gate: a condition that must pass before the budget can increase. The gate should examine a relevant quality question through a process outside the performance-measurement path. Possible checks include traffic-quality assessment, device consistency, behavioral consistency or post-conversion validation.

Independence must be established when the gate is designed, using the questions about shared datasets discussed earlier. The implementation needs to identify which evidence permits expansion and explain how it adds information beyond reported performance. A second interface using the same upstream data cannot meet the requirement merely by carrying a different provider name. Likewise, the performance metrics in the earlier buying sequence cannot substitute for the required quality observation.

Make quality and security flags binding

Passing a gate must also leave the system subject to quality and security restrictions. When a designated flag appears, the system must enforce a stop or a mandatory budget reduction. The agent must have no discretion to trade that intervention for an expected performance gain. Because its optimization objective gives it a continuing reason to spend, the trust policy needs authority over that objective.

A binding flag establishes which actions remain available after a problem is detected. An advisory warning instead leaves the agent to weigh the warning against the attractive results that encouraged expansion. Once the restriction takes effect, the system can continue optimizing among permitted actions. Compliance with the restriction must be a condition of operation.

Retain the evidence authorizing every increase

Enforced restrictions also need records that let a team examine how they operated. For every expansion, retain the inputs available at decision time, their origins, the gates already passed and the reason the spending change was permitted. Simply recording that the agent took the action gives reviewers too little information. They need to reconstruct the relationship between the evidence, the applicable policy and the resulting decision.

Suppose, hypothetically, that an agent increases a source’s budget on Tuesday and the team reviews it on Friday. Additional validation may arrive in between and change the team’s understanding of the source. The record must distinguish what the agent could know on Tuesday from what became known afterward, so reviewers can determine whether it stayed within its authority. Preserving that distinction makes accountability possible without requiring a person to approve every transaction.

In conclusion

The decision for business leaders is how much financial authority an advertising agent should receive before the quality of its purchases is known. An agent can optimize effectively against early results while committing more money than the available evidence justifies. Permission to execute purchases should therefore remain separate from permission to expand a budget.

Before increasing that authority, leaders need agreement on what evidence permits expansion, who is responsible for assessing its independence and which restrictions the agent cannot override. Buying teams, procurement and technology leaders share that responsibility. A control is useful only if the required data arrive in time and the buying system can enforce the resulting limit.

This approach preserves speed within defined boundaries. Agents can continue buying and optimizing while unresolved quality questions limit further exposure. As evidence develops, spending authority can grow with it, supported by records that explain each increase. The standard for greater autonomy should be whether the organization can justify the agent’s next commitment, not simply whether the agent can make it.

Alexander Procter

November 17, 2026

10 Min

Okoone experts
LET'S TALK!

A project in mind?
Schedule a 30-minute meeting with us.

Senior experts helping you move faster across product, engineering, cloud & AI.

Please enter a valid business email address.