OODA: The Core of Adaptive AI Agents for Advertising Campaigns
Campaign management is less about executing individual tasks than about handling a complex, changing environment. Rule-based automation covers known cases; adaptive agents aim to handle more of the rest. The OODA loop offers a practical structure for designing them, provided their limits and approval points are explicit.
By EpicflarePublished Updated 6 min read
1. Observe
Collect performance data and context: goals, plans, setup, conversions, third-party signals.
2. Orient
Turn observations into hypotheses about what works and why, to be tested.
3. Decide
Choose a plan with sub-goals, success criteria and the limits of what can run unattended.
4. Act
Apply the change through platform APIs, check it, then measure the business result.
Human approval above agreed limits
Managing ad campaigns calls for AI agents that do not merely execute predefined instructions but reason, plan and adapt within a complex system. One useful structure for such agents is the OODA loop: Observe, Orient, Decide and Act. Originally a model of military decision-making, it maps well onto the cycle of campaign management: each action produces new data for the next observation.
OODA Applied Across the Ad Campaign Lifecycle
1. Observe: Ingesting the Campaign Context
An advertising agent’s Observe phase goes beyond campaign performance data. Through APIs or MCP servers, it gathers the context needed to interpret that data:
- Campaign goals: Objectives such as brand awareness, conversions or lifetime value (LTV).
- Planning: Audience planning with data segments, and media planning with inventory targeting and allowlists.
- Execution setup: Open auction, Deal IDs, bid price thresholds and bidding optimization settings.
- Conversion data: Event counts, deduplication and conversion funnel dynamics.
- Complementary signals: Third-party data and metrics such as viewability, invalid traffic (IVT), audio activation and ad clutter.
- Campaign setup: Line item distribution, ad copies per line item, formats and ad copy variants.
The output is a structured, regularly updated view of the campaign and its context. Its quality depends on the freshness, coverage and consistency of the sources: gaps should be flagged, not filled by assumption.
2. Orient: From Raw Data to Working Hypotheses
This is where observations become usable information. The agent identifies patterns and anomalies and forms hypotheses about their causes.
A correlation is not a cause. If one audience group converts better with a given ad copy, delivery, seasonality, budget changes or attribution settings can produce the same pattern. The agent’s explanations are hypotheses to test, for example with a controlled change or a holdout group, before they drive decisions at scale.
Example: audience planning
The agent compares campaign performance across audience groups and their ad copies to suggest what performs well and what underperforms. When a difference appears, it names the factors that could explain it and proposes a test to tell them apart. The same approach applies to creative strategy, bid prices, margin management, delivery pace and capping.
3. Decide: Committing to a Course of Action
With this context, the agent evaluates possible actions and commits to a plan. In a complex system, isolated actions can have unintended consequences, so the decision must account for the interplay between campaign elements.
The plan breaks a global goal into sub-goals, each with a success criterion that shows whether it was achieved. It relies on the agent’s toolbox, the set of capabilities it can use, and states which actions it may take on its own and which need approval.
Example: media planning and bidding
Instead of simply increasing a bid, the agent checks whether the new bid keeps an acceptable margin, whether it keeps the campaign CPA within target and whether it could cause a delivery surge that spends the budget too quickly. It also checks capping and pacing constraints, which may be revised in the process.
4. Act: Executing and Looping Back
The Act phase turns the decision into a change on the platform, through its API or an MCP server. The result of the action becomes new information for the next loop:
- Success: The change is confirmed and meets its technical criteria; the plan is updated and the agent moves to the next task.
- Failure recovery: The change does not meet its criteria; the agent returns to the Decide phase, diagnoses the failure and updates the plan.
- Full reset: After repeated failures, the agent returns to Observe and Orient to reassess the situation, or escalates to a person.
Two kinds of success need to be kept apart. An API confirmation shows that the change was applied; it does not show that the change achieved its business goal. The business result (CPA, delivery, margin) is assessed later, over a defined period, against the criteria set in the Decide phase.
Example: DSP line item update
The agent defines the required changes, such as a new bid, budget, creatives or targeting criteria. It prepares the API request, checks the documentation for the expected keys and values, and sends it to the DSP. The technical success criterion is the API confirmation, followed by reading the line item back to check the values. If the API returns an error, the agent reviews the documentation and retries with a corrected request, up to a set number of attempts, then escalates to a person. Whether the new settings improve results is evaluated separately, once enough data is available.
Limits and Approvals
An adaptive agent works within limits set by people. Before it acts on live campaigns, define what it may change on its own and where a person must approve the action.
Human approval required
- Budget increases or reallocations above an agreed threshold
- Bids or margins outside the agreed range
- New deals, audiences or creatives going live
- Pausing, stopping or restructuring a campaign
- Anything affecting contractual commitments, brand safety or client reporting
Can run within agreed limits
- Small adjustments inside pre-approved ranges
- Changes that are logged and easy to reverse
- Analyses, alerts and recommendations
Stop conditions pause the loop and alert a person: repeated failures, missing or inconsistent data, or results outside the expected range. Every observation, decision and action is logged, so that a person can review why the agent acted.
Conclusion
The OODA loop is a practical way to structure an adaptive agent: it makes each stage explicit, from data collection to action and review. It does not remove the need for clear objectives, reliable data and human oversight.
The quality of each cycle also depends on the business knowledge the agent is given: the context, rules and objectives it works with.
Go further
Agentic AI and media buying: An investment guide
A framework for assessing architecture, interoperability, operating costs and control when investing in AI for media buying.
PDF · 27 pages · Free · one short form unlocks every resource
Related reading
- AI for Ad Operations: From Request to Controlled Execution
Explore an AdOps workflow for deal requests, with explicit business rules, human approval and practical criteria for evaluating AI support.
- Bringing AdTech Expertise into AI Workflows
Turn platform knowledge, business rules and operational playbooks into AI workflow requirements, then test how reliably they are applied.
- AI Agents or Automated Workflows? Choosing the Right Approach
Compare AI agents and automated workflows through task requirements, control, evaluation and operating cost to choose a suitable starting point.
- Building an AI Workflow: Start Small, Evaluate, Expand
Scope an AI workflow around one operational task, define evaluation criteria and add complexity only when testing shows it is needed.