Guide · 6 min read
How to Build an AI Operating Model for Marketing (Not Just Add AI Tools)
By Fredrika Frenkiel, Head of Studio & AI Creative Operations at Lunar, founder of Code of Alfred. About my work
An AI operating model for marketing is a defined system for how work moves from brief to delivery with AI built into specific stages, not a collection of individual AI tools used inconsistently across a team. Building one means designing the workflow, the decision rights, and the governance layer together, rather than layering tools onto an unchanged process. The model is designed around the needs of the business, so it makes what already works more efficient instead of reinventing the wheel.
1. Understand the difference between using AI tools and having an AI operating model
Most marketing teams are in the first category: individuals using AI tools inconsistently, with no shared workflow, no shared standards, and no compounding benefit across the team. An operating model means the AI usage is structured, the same workflow runs the same way regardless of who's executing it, with clear stages, clear ownership, and shared infrastructure.
2. Design for the business you have, not the one the tooling assumes
The point of an operating model is not novelty. It is throughput. Start from what already works: the planning cadence people trust, the review step that catches real problems, the naming convention everyone already follows. Those are assets, and an AI layer that respects them gets adopted in weeks instead of quarters. Then look for the friction the business actually feels, briefs arriving as loose messages, the same asset rebuilt for five markets, approvals waiting on one calendar, and put AI exactly there. Every candidate change should answer one question: which existing, working step does this make faster, cheaper or more consistent? If the answer is that it replaces a process nobody complained about, it is reinventing the wheel and it will cost you the team's trust before it delivers anything.
3. Map the full brief-to-delivery workflow first
Before designing anything AI-related, document how work actually flows today: where briefs originate, how they're structured (or aren't), how projects get planned, produced, reviewed, and shipped, and across how many markets or formats. This map is the actual object you're redesigning, AI is a layer on top of it, not a replacement for having one.
4. Define what AI owns at each stage
For each stage of the workflow, assign ownership explicitly: AI generates first drafts and variations, structures and tags incoming briefs, scales one approved asset into multiple formats or markets, and analyses performance data for patterns. Humans retain strategic prioritisation, creative judgment, stakeholder relationships, and final approval. Write this as an explicit reference document the team can point to, not an implicit assumption.
5. Build the automation layer to move work between stages, not just generate content
A real operating model connects stages automatically: a brief submitted in one place structures itself, creates the relevant project and folders, and notifies the right people, without a person manually shepherding every handoff. This is often where the largest time savings live, more than in content generation itself.
6. Encode governance into the system, not into a static document
Brand tone, quality standards, and compliance requirements should be built into the tools people use daily, reviewed automatically, not manually checked against a PDF under deadline pressure. This is what allows the model to scale without quality eroding as volume increases.
7. Treat the model as something that evolves, not a one-time build
An AI operating model should improve as you learn what works, new bottlenecks emerge, new regulations appear, new tools become available. Build in a regular review cadence rather than treating the initial build as finished.
This is exactly the kind of system I've designed and built for a multi-market marketing organisation, brief to delivery, with AI built into the operating model itself, not bolted on top of it. See how I approach this in practice.