An agentic media buy runs in seven stages: a human writes the brief, the brief becomes a mandate, a buyer agent discovers what sellers offer, buyer and seller agents negotiate, anything above the agent's authority goes to a person, the agreed terms are written to a record both sides hold, and delivery is checked against that record. The walkthrough below follows one hypothetical campaign through all seven. The brand, people, publishers and every number in it are invented for illustration.

TL;DR. In an agentic media buy, the human work moves to the start and to the exceptions: writing the mandate, approving commitments above a threshold and judging anything unfamiliar. The agents do the comparison and the rounds of offer and counter-offer. A standard product description, such as IAB Tech Lab's OpenProposal specification, lets a buyer agent compare offers from different sellers in one format. The deal only holds up later if both parties keep the same record of what was agreed, so the delivery check compares like with like.

What does the brief and mandate look like?

The hypothetical brand is Fernlow, a fictional UK outdoor clothing company launching a waterproof jacket. Its head of marketing, also fictional, writes the first prompt: reach UK adults aged 25 to 54 who are interested in walking and hiking, across six weeks of autumn, with a budget of £50,000.

A prompt is not a mandate. Before the buyer agent spends anything, the brief is turned into explicit limits, and the head of marketing approves them. In this example the mandate sets:

Price: a ceiling of £14 CPM.
Quality: a viewability floor of 70%, measured to the MRC's 2014 viewable impression guidelines.
Frequency: a cap of three impressions per person per week.
Deal types: private marketplace or programmatic guaranteed only.
Authority: any single deal above £15,000 needs human sign-off.
Unfamiliar items: a new measurement method or a seller outside the approved list goes to a person.

How does the buyer agent find what sellers offer?

The buyer agent sends the brief to sellers and receives product descriptions back. IAB Tech Lab's OpenProposal specification, part of its AAMP framework, standardises how publishers define products and packages and represent them in RFP responses. In this hypothetical, three sellers respond in that style, each stating inventory, format, audience, price, flight and measurement basis.

Publisher A, an outdoor activity site, offers a video and display package of 2,000,000 impressions at £13.50 CPM with a 65% viewability floor. Publisher B, the travel section of a national news site, offers 750,000 display impressions at £12 CPM with a 70% floor. Publisher C, a weather app, offers display inventory at £9 CPM, with viewability measured by its own in-house method.

Because the three offers share one structure, the agent can compare them field by field. Publisher B fits the mandate as offered. Publisher A is within price but below the viewability floor. Publisher C uses a measurement method the mandate does not recognise.

How do the negotiation rounds run?

The buyer agent negotiates with Publisher A's selling agent, which works to its own mandate set by the publisher. In this hypothetical exchange:

Round 1: the buyer agent asks for £11.50 CPM at a 70% floor.
Round 2: the selling agent counters at £13.20 CPM, accepting the 70% floor.
Round 3: the buyer agent offers £12.50 CPM and adds a makegood clause: if weekly viewability falls below 70%, the seller adds impressions to cover the shortfall.
Round 4: the selling agent accepts the makegood clause at £12.80 CPM.

The buyer agent never trades the viewability floor for a lower price. At 2,000,000 impressions and £12.80 CPM, the deal is worth £25,600.

Publisher B's offer needs no negotiation. At £9,000 it is under the £15,000 limit and inside every other term, so the agent executes it and logs the decision.

When does the agent hand a decision to a person?

Two items in this walkthrough go to the head of marketing. The Publisher A deal goes up because £25,600 exceeds the £15,000 authority limit. The agent sends a short summary: the terms, the rounds of negotiation, what it conceded, and the alternatives it considered. The head of marketing approves with one condition: cut the volume to 1,600,000 impressions to keep headroom for later optimisation. Publisher A's selling agent accepts the reduced volume at the same price, bringing the deal to £20,480.

Publisher C goes up because of the unfamiliar measurement method, not the price. The agent does not guess whether an in-house metric is equivalent to the agreed standard. The head of marketing declines the offer.

This is where human judgement belongs in an agentic buy: on size, on novelty and on conditions, rather than on every routine step. Fernlow ends the round with £29,480 committed across two deals and £20,520 held back.

What goes into the shared deal record?

Once terms are agreed and approved, they are written to a deal record. For the Publisher A deal, that record holds the parties and the agents acting for them, the inventory and format, 1,600,000 impressions at £12.80 CPM, the 70% viewability floor and its measurement standard, the frequency cap, the flight dates, the makegood clause, who approved the deal and when, and a timestamp.

What matters most is that buyer and seller hold the same record, not two versions of it. In the Agree. Transact. Verify. simulation, buyer and seller records kept separately disagreed in 95.3% of 90,202 simulated deals, against 0.19% when both sides worked from one shared record. A delivery check is only as good as the record it is checked against.

How is delivery checked against the deal?

At the end of week two in this hypothetical, Publisher A reports 540,000 impressions with measured viewability of 64%. The buyer agent compares that against the record. At the agreed 70% floor, 540,000 impressions should have produced 378,000 viewable impressions; 64% produced 345,600. The shortfall is 32,400 viewable impressions.

The selling agent reads the same record, sees the same clause and schedules the makegood. Nobody argues about what was agreed, because there is only one version of it.

What did the humans actually decide?

Across the whole buy, people made three decisions: approving the mandate, approving the large deal with a condition and declining the unfamiliar offer. The agents did the reading, comparing, negotiating, recording and checking. Fernlow stays accountable for every commitment, because an agent is not a legal entity.

Alkimi runs a neutral agentic advertising marketplace in which this pattern is the default: its agent is listed in the IAB Tech Lab Agent Registry, it checks that connecting agents are authorised by the buyer or publisher they represent, and agreed terms are written to a timestamped DealSheet that either party can retrieve independently.

Frequently asked questions

Is this worked example a real campaign?

No. Fernlow, its staff, the three publishers and every figure in the walkthrough are hypothetical, chosen to show how each stage works.

What is a mandate in agentic advertising?

A mandate is the set of limits a person approves before an agent acts: price, quality, frequency, deal types, an authority threshold and rules for anything unfamiliar.

When does a buying agent escalate to a human?

When a decision exceeds its authority, for example a commitment above a set value, or when it meets something the mandate does not cover, such as an unrecognised measurement method.

What is OpenProposal?

OpenProposal is an IAB Tech Lab specification, part of AAMP, that standardises how publishers define advertising products and packages and represent them in RFP responses.

Why does a shared deal record matter?

Because a delivery check compares what was delivered with what was agreed. If buyer and seller hold different versions of the terms, the dispute starts with which version is right.

Agentic Advertising Worked Through: How Does One Brief Go From First Prompt to Approved Deal?