Agentic media buying: The data and integration challenges for digital agencies

AI is moving beyond creative production and recommendations into the operational side of advertising. At Advertising Week New York (AWNY) 2026, this shift was a recurring theme, with discussions highlighting the role of structured data, connected systems and human oversight in making AI more useful across advertising workflows.

For digital agencies, agentic media buying could change how campaigns are planned, optimised and managed. Instead of navigating multiple advertising dashboards to analyse performance and adjust bids, media buyers could increasingly rely on AI agents to interpret data and execute tasks across platforms.

But this requires more than connecting a Large Language Model (LLM) to advertising tools. Campaign data needs to be reliable, systems must work together, and automated decisions must stay within clearly defined limits.

As we explored in How digital agencies can stay competitive in the age of AI and constant change, adopting AI is rarely just about the tools themselves. It also requires the right technical foundations.

Here are three challenges digital agencies need to address before they can rely on AI agents to execute media buying tasks.

1. Connecting fragmented platforms and ensuring data reliability

The agency problem: How do you bring data from different advertising platforms together and ensure it is reliable enough for AI agents to make informed decisions?

Media buyers work across a fragmented ecosystem. Campaign data, audience signals and performance metrics are spread across Google, Meta, demand-side platforms (DSPs) and publisher systems. Each platform may use different reporting structures, naming conventions and attribution methods.

Without consistent data, an AI agent could misinterpret campaign performance, compare incompatible metrics or recommend changes based on incomplete information. Automating these decisions would make existing data quality issues harder to spot and potentially more costly.

Before introducing agentic media buying, agencies need reliable data ingestion pipelines that collect information from advertising platforms, validate it and standardise it for analysis. Depending on the organisation's needs, this may involve a centralised data warehouse or data lake, alongside consistent definitions for campaigns, audiences, conversions and performance metrics.

The goal is to establish a dependable source of campaign data that AI agents can access across systems. When metrics are consistent and data quality is monitored, agents have a stronger foundation for identifying inefficiencies and recommending budget adjustments.

2. Enabling cross-platform execution without losing budget control

The agency problem: How can an AI agent act across multiple advertising platforms without exceeding its permissions or making costly changes?

One of the biggest concerns with agentic media buying is giving AI the ability to change live campaigns. An agent that can adjust bids, reallocate budgets or modify targeting needs clearly defined boundaries. A convincing recommendation is not enough to justify unrestricted access to advertising accounts.

Agencies should approach this transition gradually, moving from AI-assisted analysis to controlled execution as the system's reliability is established.

Initially, AI agents can analyse cross-platform performance, identify opportunities and recommend changes for a media buyer to review. Once the workflow has been tested, selected actions can be automated within predefined limits.

This requires more than a well-written prompt or a capable AI model. The integration layer must enforce permissions and validate actions before they reach the relevant advertising platform.

For example, an agency might allow an agent to reallocate up to 15% of a campaign budget within a defined period, while requiring human approval for larger changes. This is an illustrative policy rather than a universal threshold; the appropriate limits depend on the client's requirements and risk tolerance.

The Model Context Protocol (MCP) can help AI applications interact with external tools through a standardised interface. Where appropriate, MCP connectors can expose selected advertising capabilities to an agent.

However, MCP does not provide budget governance by itself. Access controls, financial limits and validation rules must be enforced by the underlying systems.

Human-in-the-Loop (HITL) workflows remain important throughout this process. They allow media buyers to approve higher-risk actions, review exceptions and retain responsibility for decisions that should not be fully automated.

3. Automating workflows while keeping reporting transparent

The agency problem: How do you automate campaign management while remaining accountable for decisions made by AI agents?

When a media buyer adjusts a campaign manually, they can explain why a budget was moved, which performance indicators informed the decision and what they expected to achieve. When an AI agent executes the same action through an API, that context can easily disappear unless the workflow is designed to preserve it.

This matters because clients still expect agencies to explain campaign performance and justify how their budgets are being used. There is a very real risk that handing execution entirely over to the native AI of major ad platforms will turn media buying into a black box. Without independent oversight, these algorithms will naturally optimise for the walled garden's own ad revenue, pushing budgets into formats that offer little visibility and diminishing returns for the client.

Automation cannot become an excuse for decisions that nobody can adequately review. To prevent this, agencies need reporting and audit mechanisms that record the actions agents take and the context behind them. These records should capture relevant inputs, the rules and permissions applied, the changes made, approval steps where applicable, and the resulting campaign performance.

For example, a reporting system could show that an agent reduced spend on a campaign that consistently exceeded its target cost per conversion and reallocated the budget to a campaign performing within the agreed target. The agency could then compare the outcome with the original objective and determine whether the intervention delivered the expected result.

The aim is not necessarily to expose every internal step of an LLM's reasoning. It is to make automated actions traceable, explainable in operational terms and open to review.

This also requires clear responsibilities. Agencies need to establish which decisions an agent can make independently, which require approval and who is accountable for monitoring the results.
When reporting is built into the automation architecture from the beginning, AI agents can help teams manage campaigns more efficiently without undermining client trust.

Building the foundations for agentic media buying

Agentic media buying is not simply a matter of adding an AI model to an existing advertising workflow. It requires the right data architecture, reliable integrations and controls that allow agents to act within defined boundaries.

For digital agencies, preparing for this shift means addressing three connected priorities:

• Data engineering: Building and maintaining reliable data pipelines that bring information from advertising platforms into a consistent structure.
• API integrations and MCP connectors: Connecting AI agents to advertising tools and internal systems through controlled, secure interfaces.
• Workflow automation and governance: Implementing permissions, financial guardrails, approval processes and reporting mechanisms that keep automated decisions accountable.

These foundations can also support other AI initiatives across an agency, from performance analysis and reporting to wider workflow automation.

Vega IT works with digital agencies on custom software development, data engineering and system integrations that connect existing platforms and support new AI-enabled workflows.

If you're exploring agentic media buying, we can help you assess your current infrastructure, identify integration gaps and define a practical path towards controlled AI execution.