An approval threshold is a limit, usually expressed as an amount of money, above which an automated agent or a person may not act alone and must obtain sign-off from someone with more authority. Below the line, the action executes; above it, the action pauses until a named human approves, rejects or amends it. Finance departments and public buyers have used the device for decades. It has gained a new meaning because agents built on large language models (LLMs) now book advertising, and the threshold decides how much of that buying happens without a person looking at it.

How an approval threshold works

Every threshold has three parts: a measure, a limit and an escalation path. The measure is usually the value of a single transaction, but can be a cumulative total, a change against a plan, or a category of decision. The limit is set by whoever delegates the authority. The escalation path names who receives the request, how they authenticate, and what happens if nobody answers.

The Ad Context Protocol (AdCP), an open protocol for agent-to-agent media buying, gives the clearest public specification of the mechanics in advertising. A buyer first registers a campaign plan with a separate governance agent through a sync_plans call. The plan carries a total budget, and since version 3.0, released on April 22, 2026, a field called reallocation_threshold. Version 3.0 replaced an earlier authority_level setting with two separate controls: reallocation_threshold, a number of zero or more (or a reallocation_unlimited flag) that governs how much budget the agent may move on its own, and human_review_required, a boolean that forces review of every decision, according to the AdCP changelog.

Before any purchase, the buying agent calls check_governance with the full request it intends to send. The governance agent answers approved, denied or conditions. When a buy exceeds the agent's authority, the check goes asynchronous while a person reviews it. Sellers must run their own execution check, using credentials exchanged through sync_governance, so a buyer cannot skip the step.

AdCP's documentation works through an example. A $50,000 campaign plan sets a reallocation threshold of $5,000, meaning the orchestrating agent may shift up to that sum between sellers unaided. The agent is authorised for $20,000 per transaction. It then proposes a $25,000 connected television buy. The governance agent flags the excess, a human approver reviews it, and approves with a condition that delivery be reported weekly rather than at the end of the flight. The approval, the approver and the condition are written to an audit record. "This boundary is a human decision, not a technical default. The agent cannot change it," the documentation says.

AdCP also describes audit, advisory and enforce modes, so an organisation can see what its thresholds would block before they block anything. Version 3.0 made the review flag hard to switch off: plans that touch housing, lending, employment or pharmaceutical advertising cannot set human_review_required to false, and downgrading it requires a human_override record with a written reason of at least 20 characters, an approver's email and an approval no older than 24 hours.

IAB Tech Lab's open-source buyer agent, part of its Agentic Advertising Management Protocols (AAMP), uses simpler defaults. Bookings are held in an "awaiting_approval" state unless an auto_approve flag is set, and that flag defaults to false. Its pacing engine caps any single budget reallocation proposal at 30% of the total budget, according to the code in IAB Tech Lab's repository as of October 2026.

Who sets them in ad buying

Buy-side thresholds are set by the advertiser or agency delegating authority. Platforms impose their own. PubMatic's guardrail architecture, launched on August 5, 2026, uses dual-tier permissions in which platform-wide budget thresholds sit above buyer settings; an agent exceeding its authority stops and notifies an approver, who authenticates with a one-time passcode.

Yahoo DSP's agents, announced at CES on January 6, 2026, can execute changes in the platform only with human user approval, and advertisers can audit and override them. Fluency, which says it manages $3 billion in annual media spend, denies LLMs execution rights altogether and leaves spend to deterministic agents following advertiser rules, with campaigns launching only after the advertiser's team signs off. The threshold, in effect, is zero for the model.

Origin in finance and procurement

Corporate delegation of authority policies long predate software agents: each role gets a ceiling, and anything larger goes up the hierarchy. Public procurement turned such ceilings into law.

In the United States, the Federal Acquisition Streamlining Act of 1994 created a micro-purchase threshold of $2,500, below which an official could buy without competitive quotes, and a simplified acquisition threshold of $100,000. Both are adjusted for inflation. A final rule published in the Federal Register on August 27, 2025 raised them to $15,000 and $350,000 respectively from October 1, 2025, up from $10,000 and $250,000. The Federal Acquisition Regulation also forbids breaking a requirement into smaller purchases merely to stay below the simplified acquisition threshold, an early acknowledgement that every limit invites splitting.

The European Union applies the same logic to disclosure: under Delegated Regulation (EU) 2025/2152, contracts above €5,404,000 for works and €140,000 for central government supplies and services must be advertised EU-wide in 2026 and 2027, down from €5,538,000 and €143,000, according to law firm Arthur Cox.

The move to software agents

Agent research inherited the idea early. In April 1996 Anthony Chavez and Pattie Maes of the Massachusetts Institute of Technology's Media Lab presented Kasbah, whose buying agents held a highest acceptable price and left final approval to their owners. The modern version arrived with LLM agents that operate browsers. OpenAI's Operator, released on January 23, 2025, was designed so that "before finalizing any significant action, such as submitting an order or sending an email, Operator should ask for approval," according to OpenAI. Payment protocols then wrote ceilings into credentials: a delegated token under the Agentic Commerce Protocol of September 29, 2025 carries a maximum amount, and Google's Agent Payments Protocol (AP2) uses signed mandates that can contain a price limit.

Regulation reinforces the pattern without setting numbers. Article 14 of the EU AI Act requires that people overseeing high-risk systems be able "to decide, in any particular situation, not to use the high-risk AI system or to otherwise disregard, override or reverse the output," and to stop it. Those duties for stand-alone high-risk systems were pushed back to December 2, 2027 under the 2026 Omnibus deal. Article 22 of the General Data Protection Regulation (GDPR), on solely automated decisions, underpins AdCP's forced review.

Why it matters for marketers

The threshold decides how much autonomy an agent actually has, whatever its marketing says. Shailley Singh, chief operating officer of IAB Tech Lab, listed human approval before spend among the 13 functions on which AAMP and AdCP overlap, which makes the threshold one of the few concepts both competing standards treat as essential.

Caution in the market is explicit. In January 2026 Nishant Khatri, executive vice president of product management at PubMatic, told Digiday: "We're intentionally being cautious on what we're directly and entirely attributing to agentic systems at this stage." Low thresholds keep humans in most decisions and erode the efficiency agents promise; high ones shift liability onto whoever set the number. The figure in a governance plan is also evidence of who authorised a disputed buy.

Limitations and disputes

Splitting. A per-transaction limit can be defeated by several transactions just under it. Cumulative and reallocation limits exist to catch this, but only if the plan aggregates across sellers and time; AdCP added experimental date-based budget periods in a 3.3 beta.

Rubber-stamping. Approvers flooded with requests may stop reading them. The AI Act names the risk in Article 14(4)(b), which asks that overseers stay aware of "automatically relying or over-relying on the output." A threshold that is too low can produce less scrutiny, not more.

Choosing the number. No standard says what a sensible threshold is. AdCP's figures are illustrations, and PubMatic, Yahoo DSP and Fluency have published none.

Verification. Most guardrails are vendor-asserted. PPC Land's report on PubMatic noted that there was no third-party attestation or exportable log format, so a buyer cannot independently confirm that a threshold fired.

Advance consent. Approval can be granted so broadly that it means little. A checkbox in the Google Ads support form authorises a Google specialist to make changes directly without a confirmation step, which PPC Land reported in February 2026; Ginny Marvin, Google's Ads Product Liaison, said the wording "is not new."

Spending limit or budget cap. A cap stops spend outright when reached; nobody can approve past it within the system. A threshold triggers a request and lets a human authorise more.

Human in the loop. The broader design principle that a person participates in an automated process. A threshold is one way of deciding when that participation happens.

Multi-party approval. Google Ads requires a second administrator to accept sensitive account changes, such as adding users or changing access levels, with requests expiring after 20 days, according to Google's API documentation. The trigger is the type of action, not its value.

Ad approval. Platform policy review of an ad before it serves. It concerns content compliance, not spending authority.

Recent developments

Governance is moving from vendor features into specifications. IAB Tech Lab extended the comment period on AAMP 3.0, whose OpenProposal draft defines a lifecycle for buyer and seller agents, to November 10, 2026. Regulators are asking the same questions: the UK Information Commissioner's Office has invited evidence on agentic AI, including guardrails and approvals, with responses due by November 20, 2026. Agents already operate at volume: 17.5% of 36,289 Meta ad batches uploaded through one tool in September 2026 came from AI agents.

Timeline

  • October 13, 1994 - Federal Acquisition Streamlining Act creates a $2,500 micro-purchase threshold and a $100,000 simplified acquisition threshold in the US
  • April 1996 - MIT Media Lab presents Kasbah, whose buying agents leave final approval to their owners
  • January 23, 2025 - OpenAI releases Operator, which asks for approval before significant actions
  • September 16, 2025 - Google announces the Agent Payments Protocol with signed spending mandates
  • September 29, 2025 - OpenAI and Stripe release the Agentic Commerce Protocol, with amount-limited payment tokens
  • October 1, 2025 - US micro-purchase and simplified acquisition thresholds rise to $15,000 and $350,000
  • October 22, 2025 - European Commission adopts 2026-2027 procurement thresholds, including €5,404,000 for works
  • January 6, 2026 - Yahoo DSP launches agents that execute changes with human approval
  • April 22, 2026 - AdCP 3.0 introduces reallocation_threshold and human_review_required in campaign governance
  • May 2026 - EU provisional Omnibus deal moves AI Act Annex III high-risk duties to December 2, 2027
  • August 5, 2026 - PubMatic launches guardrails with budget thresholds and authenticated approvals
  • August 11, 2026 - Fluency sets out deterministic-only execution of live spend
  • November 10, 2026 - Extended close of comments on AAMP 3.0
  • November 20, 2026 - Deadline for evidence to the UK ICO on agentic AI

Summary

Who. Advertisers, agencies and finance teams set approval thresholds; platforms such as PubMatic, Yahoo DSP and Fluency enforce their own; standards bodies including IAB Tech Lab and the AdCP community specify them; procurement law and the EU AI Act frame them.

What. A limit, usually monetary, above which an agent or employee must obtain human sign-off before acting, with a defined measure, a set figure and an escalation path.

When. Rooted in corporate delegation of authority and statutory procurement limits such as the US thresholds of 1994; applied to software agents from 2025 and written into ad buying protocols in 2026, notably AdCP 3.0 on April 22, 2026.

Where. In governance agents and plan registries, buyer-agent booking flows, DSP and sell-side platform permissions, payment credentials and public procurement rules.

Why. Agents can commit money faster than people can check it. The threshold decides which decisions stay with humans, creates a record of who authorised what, and gives regulators and auditors a point of control.