Bringing AdTech Expertise into AI Workflows

A general-purpose model is a capable student. It does not know, however, which bid settings and inventory sources suit a given campaign goal for a given advertiser. That knowledge comes from experience, and it is often what separates a generic output from one a team can actually use.

The Knowledge Gap

To be useful in day-to-day operations, a workflow needs more than generic knowledge. It needs to account for:

• Platform specifics: The rules for buying on Google Ads are different from Meta, which are different from The Trade Desk, for instance. The workflow needs these differences spelled out.
• Campaign goals: A campaign focused on sales (ROAS) is managed differently from a branding campaign. The metrics, such as viewability, VTR or click rate, the corresponding bidding strategies and the creative approaches all differ.
• Operational nuances: From setting the right CPM limit on a specific site to knowing which audience and inventory combinations lead to poor delivery, the workflow needs the tricks of the trade.
• Team requirements: It should also follow the team's internal standards, such as naming conventions and how to structure line items and creative sets.
• Escalation criteria: It needs to know which situations fall outside its scope and who should review them.

Example: A Playbook for Video Branding Campaigns

One practical way to formalize this knowledge is a playbook.

Here is a simplified example for managing branding campaigns with video:

The playbook also tells you how to test the workflow. For example, give it a campaign with a low completion rate and check that it considers the preroll duration and the skip policy before recommending a change to bids or targeting.

The Problem with One Big Playbook

It is tempting to put all your AdTech knowledge into a single, massive playbook. But this creates a document that is too large for a workflow (or a person, for that matter) to navigate effectively.

Instead, build a structured library of shorter guides. This is similar to how a company trains a new employee: with a set of practical guides, not a single, confusing encyclopedia.

Start with the situations your team encounters most often. Test exceptions explicitly, record gaps and define when the workflow should ask for help.

Connecting the Library to the Workflow

A workflow that makes decisions follows a cycle similar to the one a specialist uses. One useful model is the Observe, Orient, Decide, Act (OODA) loop: the system analyzes the situation, forms an understanding, plans a course of action and executes it. The relevant knowledge should be available at each stage of this loop, not only at the end.

Retrieval can be organized as a small pipeline around the main workflow. It typically relies on a few components:

Evaluating Whether the Knowledge Is Applied

Access to the right material does not mean it is used correctly. Evaluation closes the loop:

Treat the Workflow Like a New Team Member

Think of an AI workflow not as a black box but as a capable new team member. To do the job well, it needs what you would give a new hire:

The aim is not to replace the team's judgment, but to make its knowledge available in a form a workflow can use, and to check that it is used correctly.