TL;DR: A media-buying agent executes within a defined mandate: it submits bids, negotiates deal terms within set parameters, manages frequency, adjusts pacing, and responds to real-time inventory signals. Agents do not define mandates, establish data-processing lawful bases, take legal accountability for commitments, or approve spend above defined thresholds. The IAB Tech Lab's agentic advertising specifications require human approval for material commitments above those thresholds, making threshold design the critical governance variable. A vague mandate does not produce a cautious agent; it produces one that finds every gap in what was left undefined.
Most coverage of AI agents in media buying is prospective. It describes what agents will be able to do once infrastructure is standardised, once mandate tooling matures, once programmatic at scale adopts shared deal records. That coverage is not wrong about direction but it is not useful to practitioners building agent workflows in the near term. What is useful is precision about the current state: which tasks an agent actually executes, which decisions remain outside its scope, and how the handover between agent and human is designed to work.
What specific tasks does a media-buying agent execute today?
A media-buying agent's tasks are determined by its mandate: the instruction set defined by the brand, agency, or operator that authorises it to act. Within that mandate, the agent's core functions include bid submission against defined audience, context, and price parameters; deal-term negotiation with sell-side agents up to the limits the mandate specifies; frequency management across placements to maintain the per-user thresholds the mandate requires; pacing adjustment to distribute delivery across a campaign period; and real-time response to inventory signals including price floors, contextual data, and placement availability.
These are execution tasks. They require speed and pattern recognition at a volume and pace that human buyers cannot sustain across a large campaign portfolio, which is the practical case for using agents for them. They are not judgement tasks. The agent does not assess whether the campaign objective is correct, whether the audience definition serves the brand's underlying goal, or whether the creative being delivered is appropriate for the context it is being placed in.
What are the hard limits on what a media-buying agent can do?
The boundaries on agent scope fall into three categories: governance limits, legal limits, and operational limits.
Governance limits are by design. The IAB Tech Lab's agentic advertising working group specifications require that material commitments above defined spend thresholds trigger human approval before the commitment is finalised. The agent cannot approve a deal that exceeds the threshold in its mandate. It must route the commitment to a human approver and wait for confirmation. This is not a technical limitation; it is a deliberate governance structure, and the threshold itself is the variable that practitioners design.
Legal limits are structural. A media-buying agent is not a legal entity. It cannot take legal accountability for the commitments it makes on a brand or agency's behalf. That accountability rests with the principal who deployed it. For data processing, agents cannot establish the lawful basis required under the UK GDPR and GDPR for using personal data in audience targeting. The data controller's lawful basis must be established before the agent is authorised to use a data asset. An agent that accesses audience data it was not pre-authorised to use is not merely exceeding its mandate; the data controller deploying it may be in breach of data protection obligations.
Operational limits are practical. A media-buying agent has no visibility into the creative asset it is placing. It cannot assess whether a video advertisement is appropriate for the context it wins, whether the creative carries the correct format specifications for the placement, or whether it conflicts with any editorial adjacency rules that were not encoded in the mandate. Creative review remains a human task. An agent that wins a contextual placement without creative-format constraints in its mandate may deliver a mismatch.
Where are the designed handover points between agent and human?
Well-designed agent workflows define explicit handover points rather than leaving the boundary to emerge from agent failures.
The primary handover is the spend threshold. When a single deal commitment, or a cumulative spend total over a defined period, reaches the threshold specified in the mandate, the agent routes the decision to a human approver and holds the commitment until confirmation is received or the window expires. The threshold is the practitioner's main lever for governing agent scope.
Secondary handovers include: new data asset access requests, where the agent identifies an audience segment not pre-authorised in the mandate and must request authorisation before using it; mandate renewal, where a campaign period ends and the agent's authority lapses until a new mandate is issued; and any commitment the agent cannot classify as within mandate scope, which defaults to human review rather than agent refusal.
The IAB Tech Lab's published agentic advertising working group summary identifies the design of the human-agent handover, not the capability of the agent itself, as the primary determinant of whether agentic buying creates accountability risk or manages it. The agent executes what the mandate allows. The mandate is the practitioner's responsibility.
What does this mean for how media teams need to work?
Agent deployment does not reduce the work required from media teams; it changes which work they do. Bid management, pacing, and frequency optimisation shift to the agent. Mandate design, threshold calibration, and approval-queue management become the human tasks.
This is a meaningful shift in required skills. Mandate design is an analytical and governance function: translating a campaign brief into precise, unambiguous agent instructions with defined constraints on data use, spend authority, creative requirements, and brand-safety parameters. A practitioner who can write a clean mandate that produces the intended agent behaviour without exploitable gaps is a different kind of operator than one who was optimising bid multipliers manually.
The organisations building that competency now, while agentic volumes are low and the cost of a vague mandate is a $3,000-a-day lesson rather than a multi-million-pound one, have a material advantage over those who wait until the infrastructure is mandatory. Andrew Mole of pubX confirmed the $3,000-a-day volume figure in an ADOTAT investigation published in August 2026.
*This article references the IAB Tech Lab's published agentic advertising working group specifications and working group summary; and volume data from Andrew Mole of pubX, published in an ADOTAT investigation in August 2026.*