Agencies use AI for programmatic buying by handing agents the repeatable work (finding inventory, sending proposals, checking pacing, drafting reports) while traders review what the agents did, approve anything that commits money and stay answerable to the client. The trading desk does not disappear. Its day reorganises around decision logs, an approvals queue and a shared deal record.

TL;DR: This is one hypothetical day on an agency trading desk running AI buying agents, built from practice agencies described in trade press during 2026. Traders read the overnight logs, clear approvals, update mandates, answer clients, settle exceptions, reconcile delivery and sign off the report. At every hour the agent proposes and records, and a named person decides.

What does a day on an agentic trading desk actually look like?

The desk in this piece is hypothetical: a composite mid-sized agency team running buy-side agents across open web and CTV deals. Every practice attributed to a named company comes from its own statements or from trade press.

IAB Europe's 2026 Impact of AI on Digital Advertising Report found that 86% of respondents' organisations use AI for marketing and 58% expect agentic ad buying to reach operational use or scale within the next year. Only 48% have AI guidelines written specifically for advertising and marketing, so much of the oversight still sits in the desk's daily routine.

For the wider split between AI-assisted optimisation and agent-led negotiation, see How Agencies Are Using AI for Programmatic Buying.

8.30am: What do traders check first in the agents' overnight decision logs?

The first job is reading what the agents did overnight. On the hypothetical desk, the buy-side agent has sent proposals to several sellers, accepted two holds and walked away from one package, and each action sits in a timestamped log next to its brief.

Performance agency Rise, part of Quad, has been testing AI buying agents with a supermarket client using an audit log that shows when agents drift outside pre-set parameters, Digiday reported in September 2026. Group director Klaudia Smykowska described the point of it: "I can make sure everything is recorded and that we can reconstruct what happened and why."

The trader's question for each entry is narrow: was it inside the mandate, and does the stated reason match the brief? Jonathan Whiteside of Dept put the rule plainly in the same report: "Every deliverable has an accountable human."

9.30am: What goes into the approvals queue?

The queue holds every agent action that commits money or changes what runs live. WPP Media's Lauren Wetzel told Digiday in October 2026 that where a decision has "a financial or activation impact", the holding company requires "explicit controls, approval thresholds, accountable human oversight."

In practice the queue is a list of drafts waiting for a name. In Butler/Till's first agentic buying tests, curated inventory from the seller's agent went to agency staff for approval before the agency's agent confirmed the buy, according to Digiday.

One concrete example of a structured queue is Alkimi, where agent deals go through a neutral marketplace and humans keep approval rights at the deal level. Agents there move along an earned autonomy ladder: Observe, Recommend, Draft, Human-approved action, Bounded automatic action. Most items in the hypothetical morning queue sit at Draft or Human-approved action.

11am: How does the desk update an agent's mandate mid-campaign?

Mandate updates are the trader's main lever over agent behaviour. A mandate is the written set of rules an agent buys to (budget, inventory, brand suitability, price boundaries, which terms need a human), so when a client changes a target the desk edits the mandate rather than the bids.

Agencies are codifying those rules into the agents themselves. PMG uses Skills in ChatGPT or Claude to "bake best practices" into an agent's design, consulting and strategy director Dillon Larberg told Digiday.

Standards work points the same way. IAB Tech Lab's AAMP 3.0 extends agentic workflows to the brief stage through the OpenProposal specification, and one request filed in its public comment period, open for public comment until 10 November 2026 after an extension, asks that offered terms be frozen per version, with any change creating a new one.

A deal agreed under mandate version three should be judged against version three, not today's edit. For how desks staff this work, see How Trading Desks Are Restructuring for Agentic Workflows.

1pm: What do clients ask the desk about their agents?

Client questions are mostly about control, not performance. Digiday's Advertising Week coverage found agencies answering client fears of rogue agents with "human hand-holding", and Dentsu's head of platforms Rebekah Shalit described the principle: "Every task and everything we do sits on a sliding scale of low to high human oversight."

The reassurance rests on approval rights and limits. David Dworin, chief product officer at FreeWheel, said in the same coverage: "Most agentic systems are either designed with a human that has to approve everything, or very specific controls that can keep it from getting into trouble."

Agents increasingly draft the reply too. Kovva, a startup profiled by AdExchanger, can pull reports from across platforms and draft answers to client emails, with a human approving each before it is sent. The question underneath every client call is the one IAB Europe's Dimitris Beis put to the industry: "who is accountable when something goes wrong?"

2.30pm: What happens when an agent negotiation hits an exception?

An exception is any negotiation the agent cannot finish inside its mandate. On the hypothetical desk, a CTV seller's agent counters with a make-good clause the mandate does not allow, so the buy-side agent pauses and routes the thread to a trader with both positions attached.

Exceptions matter because completing a deal is not the same as protecting the buyer. In Concourse Bench v1, a controlled benchmark of AI buyers negotiating with fixed-policy sellers, 55 of 96 runs completed a buy, and 95 of the 110 signed contracts promised a viewability floor below 70%. At an exception, the trader checks the protections, not only the price.

The trader can widen the mandate for this deal, counter by hand or let the agent walk away. The reason goes into the log for the next morning's review.

4pm: How does the desk reconcile delivery against the deal record?

Reconciliation is where agentic buying holds up or falls apart. AdExchanger's Allison Schiff wrote in May 2026 that bidding is automated "down to the millisecond", yet day-to-day campaign work, including reconciling reports, "is still largely manual."

On an agentic desk, the afternoon check compares delivery against the terms as agreed, not each side's later version of them. That needs a deal record both buyer and seller can retrieve, written at agreement, so a dispute over impressions or viewability starts from the same document.

WPP Media's Agentic Standards Initiative for video, developed with media owners and standards bodies including IAB Tech Lab, lists measurement and reconciliation among the areas buyer and seller agents need to align on.

5.30pm: What goes into the end-of-day report?

Reporting is where AI is already most at home. In the IAB Europe research, Digiday reported, 22 of 29 respondents named reporting, analysis and dashboards, with programmatic optimisation next.

The agent drafts the day's report and the trader signs it. On the hypothetical desk it covers spend against plan, deals agreed and under which mandate version, approvals, exceptions and open reconciliation gaps.

Efficiency is still how most firms judge the result. IAB Europe found operational efficiency was the most cited measure of AI effectiveness (69%), while respondents rated AI's current performance in advertising operations at an average of 2.76 out of 5.

What does the desk no longer do by hand, and what does it still do?

On the hypothetical desk, the agents now handle:

First proposals and follow-ups to sellers
Post-launch QA checks on geography, pixels and pacing, as AdExchanger describes
First-pass checks on cross-platform reporting discrepancies
First drafts of client updates and the daily report
Logging every action with a timestamp and reason

The traders still own:

Approving actions with financial or activation impact
Writing and versioning the mandate
Deciding negotiation exceptions
Answering to the client
Signing off reconciliation and settling disputes

The rule: the agent may prepare anything, but a named person commits the money. For how the role is shifting at US agencies, see How Agentic Buying Is Changing the Role of the US Agency Trading Desk.

Where do benchmarking, rehearsal and a shared deal record fit?

Alkimi, the enterprise platform for agentic advertising mentioned earlier, runs Concourse, the live benchmark behind the Bench v1 figures, and the Gym, available now, where buyers and sellers build their own agents, set the rules they work to and spar against counterparty agents before any real money moves. Its Marketplace is neutral, works within IAB Tech Lab's AAMP framework today and is being built to be AdCP compatible, with AAMP identity checks and human approval at the deal level.

The DealSheet, part of that Marketplace, is the kind of shared record the 4pm reconciliation relies on: written at agreement, timestamped, held by both parties and retrievable by either side independently. The company also filed comments and change requests during the OpenProposal 3.0 public comment period.

Frequently asked questions

How do agencies use AI for programmatic buying?

Agencies give AI agents repeatable work such as inventory discovery, proposals, pacing checks, QA and first-draft reporting. Traders review the decision logs, approve actions that commit money and stay accountable to the client.

Do AI agents replace the agency trading desk?

No. Agents take over much of the manual execution, while traders still write mandates, approve financial and activation decisions, settle exceptions and sign off reconciliation.

Which agent decisions need human approval?

Actions with financial or activation impact, such as committing budget or accepting a deal, normally go through a human approvals queue. Thresholds let lower-risk actions run inside agreed limits.

Which standards do agentic trading desks work with?

IAB Tech Lab's AAMP (Agentic Advertising Management Protocols) covers agentic buying workflows, including OpenProposal for the brief stage. AdCP, the Ad Context Protocol published by AgenticAdvertising.org, is an open protocol for agents to plan, buy, sell and measure media.

How do agencies reconcile agent-negotiated deals?

They compare delivery against the terms as agreed, ideally from one record both sides can retrieve. Agents do the first pass on discrepancies and people settle them.