How a Media Services Company Is Building an AI-Assisted Planning Operation

Introduction

Affiliate media planning is part science, part art, and the art is the hard part. A senior planner can look at a campaign and tell you which publisher will overdeliver, why a homepage placement beats a category page, how a percent-off offer will move a specific audience. That judgment is real and valuable. It is also trapped in people's heads, built over years, and impossible to scale by hiring. On top of it sits genuine uncertainty: forecasts that live in a margin of error, tracking that gets muddy at the placement level, and macro swings (tariffs, consumer panic-buying, shifting discount behavior) that move performance in ways no spreadsheet anticipates.

This established media services company had built a real business on that expertise: deep publisher relationships, sharp rate negotiation, fast deployment. But the operations behind the service were almost entirely manual, and the knowledge that made them good didn't live anywhere a system could use it. Growing demand meant growing cost. Scaling meant hiring. And in a market where AI-first competitors were already moving faster, that math didn't work.

The company needed to scale output without scaling headcount, and to do it without losing the judgment that made the work good in the first place. RapidCanvas helped them build an AI workforce to do both: not to replace the art, but to give it a foundation that learns.
‍

The Problem

The bottlenecks were structural, not behavioral. Even experienced planners were constrained by the mechanics of the work:

Publisher inventory was unstructured. Media kits arrived in every format: PDFs, spreadsheets, free-text emails. There was no single repository of what placements were available, at what cost, and on what terms. Building a plan meant hunting across sources that were never designed to be queried.

Publisher selection was underpowered. Recommending the right publishers for a campaign required synthesizing historical performance data, publisher behavior across advertisers, and placement-level lift, a calculation too complex to do consistently by hand. Most decisions relied on intuition and experience rather than systematic analysis.

Proposal generation was slow. Assembling a media plan (selecting publishers, estimating performance, allocating budget) required hours of manual work per campaign. There was no way to accelerate it without compromising quality.

There was no visibility into operations. Leadership had no structured way to monitor how plans were being built, how long the process took, or which publishers were being used.
‍

The Solution

Placement IQ is a structured planning platform built on top of the company's existing performance data. It has three primary components:

A publisher inventory layer. Media kits are ingested from PDF, spreadsheet, or free-text via an AI-powered extraction pipeline (GPT-4o). The system parses placement names, pricing types, costs, dates, and placement requirements, then routes them to a human review queue where planners validate, correct, and finalize the data before it enters the live inventory. The result is a queryable database of publisher placements with status tracking, tag management, and audit history.

An AI-assisted proposal builder. When a buyer starts a new proposal, the system queries a recommendation engine that scores publishers across three dimensions: historical uplift with this specific advertiser-publisher pair (measured against periods without placements), average uplift across publishers in the same class with similar advertisers, and absolute revenue potential. Weights across the three dimensions are configurable. The buyer reviews the ranked list, builds a shortlist, and moves to budget allocation.

A budget optimizer. Given a total budget and a shortlist of publishers, the system runs a constrained optimization, using linear programming to allocate spend across publishers while respecting per-publisher floor/ceiling percentages, CPA bounds, flat fee constraints, and growth rate assumptions. The result is a forecast: projected revenue, actions, ROAS, and CPA per publisher, with confidence levels based on data availability. Buyers can lock individual publisher allocations and recalculate the remainder. The final proposal exports to PDF or XLSX and is saved to the platform with status tracking (draft → in review → final).

An admin dashboard. Admins have visibility into user activity, proposal creation rates, time-to-review by user, and placement upload volume, giving operations leadership the structured data to understand workflow throughput.
‍

How It Was Built

Phase 1: Publisher intelligence.The system started where the highest-value decisions happen: publisher selection. The team built the recommendation engine and forecasting layer on top of the company's historical performance data. Publishers are scored across three dimensions: measured uplift from actual placements with a specific advertiser, class-level uplift patterns across similar advertiser-publisher pairs, and absolute revenue potential. The budget optimizer launched alongside, taking a shortlist of publishers and a total budget, then running constrained linear programming to allocate spend across publishers with per-publisher floors, ceilings, CPA bounds, flat fee handling, and growth rate assumptions. The full proposal workflow (advertiser selection, recommendations, shortlist, forecast, export) went live here, along with the admin dashboard for operational visibility.

Phase 2: Placement-level intelligence. With publisher-level planning operational, the team added the placement layer. An AI-powered ingestion pipeline (GPT-4o) processes media kits in any format (PDF, spreadsheet, or free-text), extracting placement names, pricing types, costs, dates, and asset requirements. Extracted placements route to a human review queue where planners validate, correct, tag, and finalize records before they enter the live inventory. The result is a queryable placement database with structured taxonomy, status tracking, and audit history. Placement selection was layered into the proposal workflow: buyers can now browse available inventory by publisher, select specific placements, and reconcile placement costs against their publisher-level budget allocations.

Phase 3: Deeper recommendation intelligence (in progress). The current phase builds on both layers. Work underway includes additional recommendation signals tailored to advertiser strategy (going beyond historical pair performance to incorporate advertiser-specific context into ranking) and dynamic narrative generation that explains, in plain language, why a publisher or placement is a fit for a given campaign. The goal is a system that doesn't just rank options but communicates the reasoning behind them.
‍

What Changed

The operational shift is concrete:

Planners spend their time reviewing, not assembling. Media kit ingestion goes from a manual data-entry task to a structured review: the AI extracts, the planner validates. The planner's judgment is applied to decisions that require it.

Buyers get scored recommendations, not blank lists. Rather than starting from scratch on every proposal, buyers receive a ranked list of publishers with explainable scores, weighted toward pair-level performance data when it exists, with class-level proxies filling in where historical coverage is thin.

Budget allocation is systematic and reproducible. The optimizer produces consistent, constraint-aware allocations that buyers can adjust and re-run. The forecast gives buyers a basis for explaining recommendations to clients, not just delivering them.

Operations are visible. The admin dashboard tracks who is building plans, how fast they move from draft to review, and how much inventory is being added to the system, turning an opaque manual workflow into something measurable.
‍

What Comes Next

Two capabilities are in active development. The first is richer recommendation intelligence: adding signals beyond historical pair performance so the system can factor in advertiser strategy, campaign objectives, and category fit when ranking publishers and placements. The second is dynamic narrative generation: rather than presenting buyers with a scored list, the system will produce plain-language explanations of why each publisher and placement belongs in a plan, surfacing the logic that today lives in the recommendation engine but never reaches the buyer directly.

The trajectory is clear. Each phase added a layer of intelligence: first publisher-level, then placement-level, next context-aware reasoning. The system compounds: every reviewed media kit strengthens the inventory, every proposal adds a signal to the forecast model, and every improvement to the recommendation engine makes the next buyer's starting point better than the last. That is the structural advantage RapidCanvas built here: not a tool that automates a workflow, but a platform that gets more useful the more it is used.