An AI agent can negotiate most of a connected TV (CTV) deal, but twelve terms should wait for a named person's approval before the deal binds. They are the size of the commitment, the delivery type, viewability and completion floors, the measurement currency, frequency caps, content categories, make-goods, cancellation, audience data use, exclusivity, payment terms and seller identity. Each one commits money, defines delivery or exposes the brand.

TL;DR. Trusting an agent with a CTV budget depends less on its capability than on which decisions it may finalise. Reversible terms inside the mandate can be automated, while terms that commit money, loosen protection or bind the brand should stop for approval. None of the sample rules below depends on budget size.

Why do CTV deals need human approval when an agent is buying?

Agentic buying in video has moved past the pilot stage. In IAB's 2026 Digital Video Ad Spend and Strategy Report, two in three US digital video buyers said they were live (21%), testing (20%) or planning to use (25%) agentic AI for digital video campaigns in 2026.

CTV deals also carry more commitment than an auction impression. IAB Tech Lab's explainer on the contract layer for agentic advertising describes direct-sold guaranteed campaigns as "built on commitment", and warns that agent errors in them lead to make-goods, missed flight windows and strained relationships between publisher and buyer.

The standards keep people involved too. IAB Tech Lab describes AAMP, its Agentic Advertising Management Protocols, as bringing agentic AI into advertising "in a responsible and scalable fashion", and AgenticAdvertising.org governs AdCP through "Embedded Human Judgment", the principle "that humans stay in the loop on decisions with real consequences".

Practitioners put it more plainly. "You can trust AI to fly a plane, but I still want a pilot in the cockpit to watch it," Sydney Brower, director of digital media and strategy at InterMedia, told AdExchanger in October 2026.

How does an approval gate work on a CTV deal?

An approval gate is the point where the agent has prepared the deal but a named person must accept it before it binds. Alkimi's Marketplace is one worked example: humans keep approval rights at the deal level, and the agreed terms are written into a DealSheet at agreement, timestamped, held by both parties and retrievable by either side independently.

The sample rules below follow the five stages of earned autonomy: Observe, Recommend, Draft, Human-approved action and Bounded automatic action. Most terms sit at Human-approved action until an agent has a track record, and how to design approval thresholds for agentic campaigns covers setting the limits.

1. Total commitment above a set threshold

Why it needs a person. Booked CTV inventory is reserved and the budget locked, so the commitment cannot be optimised after the fact.

What good looks like. A threshold in the mandate, counted across every deal in the flight so one large commitment cannot be split into several small ones.

Sample approval rule. Any deal that takes the flight total above the threshold needs sign-off from the budget owner.

2. Guaranteed or non-guaranteed delivery

Why it needs a person. The delivery type decides who carries the risk of under-delivery. In IAB Tech Lab's OpenDirect model, a guaranteed line moves through explicit states from Draft to Reserved, Booked, InFlight and Finished, according to the same explainer.

What good looks like. The deal names its type, exact quantity, flight dates and reservation expiry.

Sample approval rule. The agent may reserve inventory, but only a person may move a guaranteed line from reserved to booked.

3. Viewability and completion floors

Why it needs a person. Closing a deal and securing protection are different things. In Concourse Bench v1, a published benchmark in which eight AI buyers negotiated with three software-controlled sellers, 55 of 96 runs completed a buy and 95 of the 110 signed contracts promised a viewability floor below 70% (Concourse).

What good looks like. Written viewability and completion floors, a named measurement method and the share of impressions that must run on a television screen, as covered in CTV viewability floors when the buyer is an AI.

Sample approval rule. Any floor below the mandate minimum, or any deal with no floor at all, goes to a person.

4. Measurement currency

Why it needs a person. The currency decides what an audience guarantee is worth, and TV measurement is in transition. Digiday reported in September 2026 that TV is moving to a multi-currency market, with the Coalition for Innovative Media Measurement forming a task force to create "a provider-neutral currency requirements framework".

What good looks like. One named currency for the guarantee and any make-good trigger, plus an agreed response if the provider changes methodology mid-flight.

Sample approval rule. Any currency not on the advertiser's approved list needs approval.

5. Frequency caps

Why it needs a person. A cap only works if both sides count the same thing. IAB Tech Lab's CTV Programmatic Guide notes that unified reach and frequency reporting is a challenge when one creative carries different IDs across platforms.

What good looks like. A stated unit (device, household or person), scope (deal, seller or whole plan), time window and whether over-cap impressions are billable.

Sample approval rule. The agent may tighten a cap on its own, but loosening one, or accepting a cap with no stated unit, needs approval.

6. Content and brand-safety categories

Why it needs a person. Suitability depends on the brand, the programme and the news cycle. IAB Tech Lab's Content Taxonomy offers a "common language" for describing content and lists brand safety as a typical use, with a CTV genre mapping published alongside version 3.1.

What good looks like. Exclusions written as taxonomy categories the seller's agent can read, with live programming treated as a separate decision.

Sample approval rule. The approved exclusion list applies automatically; any new category, live news or seller that cannot signal genre goes to a person.

7. Make-good terms

Why it needs a person. A make-good decides what the buyer receives when delivery falls short, often weeks after the negotiation ended.

What good looks like. A trigger (which metric, which currency, what shortfall), a form (more impressions, equivalent inventory or credit) and a deadline (see make-goods in agent-negotiated deals).

Sample approval rule. Make-goods matching the mandate template are accepted automatically; offers in different inventory or outside the flight need approval.

8. Cancellation and change terms

Why it needs a person. Cancellation rights are the buyer's exit, and once traded for price they rarely come back. OpenDirect handles post-booking changes through a ChangeRequest, which IAB Tech Lab describes as "a formal post-booking modification proposal with an explicit approve/reject workflow".

What good looks like. A notice period, a cancellation window, a stated outcome for unserved impressions and a log of every post-booking change.

Sample approval rule. Every change to a booked deal, from either side, goes through a change request that a person approves.

9. Audience data use

Why it needs a person. Audience terms decide whose data goes where, and CTV identity is less precise than it looks. Truthset's State of Data Accuracy report, as reported by Digiday, found an 84% error rate for matches between IP addresses and email addresses.

What good looks like. Segments disclosed under IAB Tech Lab's Data Transparency Standard, which sets up to 20 disclosure fields including recency and provenance, and first-party data used for this campaign only.

Sample approval rule. Sharing first-party data with a seller, or using an undisclosed segment, needs approval from the data owner.

10. Exclusivity

Why it needs a person. Exclusivity binds the brand beyond a single deal, and it can clash with other plans the agent cannot see.

What good looks like. A stated scope (category, sponsorship or first look), duration, competitors covered and any competitive separation within an ad break.

Sample approval rule. Any exclusivity, whether given or received, needs approval from the brand owner.

11. Payment terms

Why it needs a person. Payment terms commit the finance team, not only the media team.

What good looks like. A billing basis (booked or delivered, and whose counts), invoice timing, payment currency and a process for disputed impressions.

Sample approval rule. Any departure from the advertiser's standard terms goes to finance.

12. Publisher and seller identity verification

Why it needs a person. Several parties, such as a streaming distributor, a device maker and a programmer, can hold rights to the same CTV ad space, as IAB Tech Lab notes.

What good looks like. The seller is authorised in app-ads.txt and listed in sellers.json, and an agent counterparty is checked against AAMP's Agent Registry, which IAB Tech Lab describes as a framework "for agent identity, verification, and disclosure".

Sample approval rule. A person approves any new counterparty before its first deal.

Which of these terms can an agent eventually agree on its own?

Terms where the agent tightens protection, such as a stricter cap or a higher floor, can move early to Bounded automatic action. Terms that commit money, loosen protection or bind the brand should stay at Human-approved action until the agent's record against the mandate justifies a change.

How should a team put these approval rules into practice?

Write the twelve rules into the mandate before the agent starts, make each approval request show every term against its limit, and keep a record of what was agreed that both sides can retrieve. Alkimi, which publishes the Concourse benchmark cited above, built its Gym so those rules can be rehearsed first: buyers and sellers build their own agents, set the rules they work to and spar against counterparty agents before any real money moves.

Frequently asked questions

Can AI agents be trusted to manage advertising budgets?

Agents can be trusted with budgets inside written limits, provided a person approves any term that commits money, loosens protection or binds the brand. Trust then widens term by term as the agent's record shows it holds to the mandate.

How do you maintain brand safety when AI agents are buying CTV media?

Express exclusions in a shared taxonomy such as IAB Tech Lab's Content Taxonomy, and require approval for any new category or live content. Verify seller identity before the first deal, since exclusions only protect the brand if the inventory is genuine.

Does human approval slow agentic CTV buying down?

Approval adds a step only at the terms that bind, while the agent still handles discovery, comparison and drafting. A pre-approved template for routine terms keeps most deals moving without a wait.

Is this checklist only for large CTV budgets?

No, because each rule is written as a mandate limit rather than an amount. A smaller advertiser sets a lower commitment threshold and keeps the other eleven rules unchanged.

CTV Deal Terms an AI Agent Should Never Agree Without Human Approval