Marketing floor plans used to travel from Revit to PDF to a graphic designer's resize queue before reaching the campaign platform. This workflow deletes the middle: plans leave Revit already sized, composed and named for the marketing content hub.

THE OLD PIPELINE HAD A HUMAN BOTTLENECK IN THE MIDDLE
Brochure and sales plans were exported to PDF, handed to a graphic designer, manually resized to the campaign format, and only then uploaded to the marketing platform. Nobody in that chain was adding design judgement; they were converting formats. That is machine work.

STAGE ONE: LEAVE REVIT ALREADY CORRECT
A pyRevit tool auto-detects brochure and sales plan views by naming convention, sorts them in floor order, and exports each as a PNG at the exact campaign ratio. The original views are never touched: the tool duplicates the view, expands the crop to the target ratio from centre so nothing is ever clipped, exports, and cleans up after itself. Linework exports at 2x or 3x and downsamples clean. Filenames carry the job, the plan and the target size, so every downstream step can match assets without a human reading them.

STAGE TWO: COMPOSITION WITHOUT A DESIGNER
A companion pipeline composes the marketing set: ground and first floor side by side on a standard canvas, and each plan design option placed location-coordinated on the same setup, matched to its base plan automatically by naming. What a designer used to assemble by eye, per option, is now a batch.

STAGE THREE: AI READS THE PLANS, WITH RULES ABOUT NOT GUESSING
The data the marketing platform needs is not all in filenames. Which side is the driveway on, viewed from the street? How many dwellings? That lives in the drawing itself, so a second MCP server puts Claude's eyes on it: five tools that discover the designs still missing data, fetch the ground-floor plan image, and write back what the model sees, from garage side relative to the entry to a legibility score. The design principle that makes it trustworthy is restraint. If the plan is a duplex, or the model genuinely cannot tell, it does not guess; the design is marked needs_review and left for a person. Brand attribution never touches the AI at all, because a lookup should never be an inference. The whole run is idempotent, and the pipeline is packaged for the marketing team itself: a thirty-minute setup, then five to ten minutes per batch. The BIM team built it; marketing runs it.

THE PRAGMATIC PIVOT
Midway through, reality intervened: much of the estate still lived in the legacy model structure the Revit-side tool assumed away. Rebuilding the exporter for old models meant months of ground work for a temporary payoff. So the pipeline pivoted: for legacy designs it starts from the already-published plan imagery instead of the model, and applies the same composition automation downstream. Same outcome, a fraction of the work. Automation earns its keep by meeting the estate where it is, not where the roadmap says it should be.

WHERE IT LANDS
Composed images plus a per-design JSON of AI-read metadata flow into the marketing content hub, where campaigns pull them directly. Design changes stop leaking stale plans into live campaigns, because the source of the image is the source of truth. It is the third downstream of one chain: the same governed models feed quantities, programmes, and now marketing.