← All services
01 / 06 · Function brief

Marketing Operations.

The whole campaign loop is agentic now. Own your side of it.

Position

Agents now plan, buy, and optimise your media. What every platform asks in return is signals about your customers. We automate the first, and ration the second.

The business case

Automate the whole loop. Leak nothing.

The campaign is now a loop that agents run, and every ad channel has grown an agent socket and a conversion API to feed. Winning it takes three moves:

01
Agents run the loop

Strategy, planning, creative, activation, wired through the official MCP servers of Meta, TikTok, Google, and Amazon, behind a gateway you log, with a human on strategy.

02
Channels get evidence, never identities

Hashed, consented conversion events to every CAPI, Netflix included. On iOS, SKAdNetwork and AdAttributionKit attribute with no ATT prompt, no intrusive ad SDKs.

03
The brain stays home

The agentic CDP as a no-copy layer on your own warehouse; decisioning on sovereign EU inference. No foreign jurisdiction ever holds your customers. Europe has priced the alternative: €1.2B against Meta, €530M against TikTok.

The operating model

One loop above the line.
Only evidence below it.

Your perimeter · EU-resident
Agents run this
01
Strategy
02
Plan
03
Create
04
Activate
05
Optimise
Customer 360 · no-copy, on your warehouseAgentic CDP context layerDecisioning on sovereign EU inference
The sovereignty line
↓ Campaigns out

Agents operate every channel through official MCP sockets, logged, budgeted, on-brand.

↑ Signals back

Hashed, consented, aggregated conversion events. iOS via SKAdNetwork / AdAttributionKit, no ATT prompt.

The channels
InstagramMCPTikTokMCPGoogleMCPAmazonMCPNetflixCAPIEvery CAPI channel+

That is the whole model. The detail, sources, and the 90-day plan are in the field study below, or bring us one campaign and we will map it live.

Inside the capability

PrivacyCloud, the data clean room that never leaves your building.

Every clean room on the market asks both parties to upload their customers to a third place and trust the walls. Ours moves the compute to the data instead, with open-weight models and differential privacy running inside a perimeter you own.

The engagement, documented

Fourteen weeks inside a sovereign marketing operation.

Five chapters, exactly as a client lives them: diagnosis, encoding, the daily view, the engine, and outcomes in two columns. Modelled client, real method.

Read the engagement →
Generative video, in production

Film-grade ads, written,
voiced, and synced in code.

Marketing teams now need more finished video than any production budget can carry, in more formats and more languages. We build the pipeline that makes that possible, and we built it on ourselves first: a campaign of nine spots for a consumer product of ours, in vertical and horizontal cuts, across five languages, rendered from code rather than from a shoot.

Spot 01

Two characters and a landscape nobody booked, carried by one continuous voice track.

Spot 04

Interiors and performance. The voice is generated first, then the faces are synced to it.

Spot 10

The shot that cannot be filmed. Same pipeline, same afternoon, same cost as the others.

Spots produced for firmas.io, a consumer product of ours. Watch the full set →

01
Script

One file per spot: narration, shot prompts, durations, and every locale variant, versioned like code.

02
Voice

Speech generated per scene and per language, cast to a named voice and timed before a single frame exists.

03
Shots

Text-to-video and image-to-video against character references, so the same face survives from scene to scene.

04
Sync

A lip-sync pass marries the generated performance to the generated voice, separately in every language.

05
Render

Composited in code and transcoded once, out in vertical and horizontal cuts, ready to post.

Full disclosure on the stack. The generation models here are rented (Kling for motion, ElevenLabs for voice) and called from a server we operate. What stays ours is the part that compounds: the scripts, the character references, the timing, the render, and every finished asset. That is the same line we draw everywhere else, rent the capability, own the harness. For an agency it is the difference between a tool subscription and a production line, and as open-weight video and voice models mature, the same pipeline runs on hardware you own without changing shape.

The signal

What's actually
shipping.

CustomerLake, Databricks' agentic CDP, runs continuous “infinity campaigns” straight from the warehouse.

Databricks

Hightouch's agents: 10 billion marketing decisions made, ~25% average lift, a $2.75B valuation.

Hightouch

Meta's official MCP server writes full campaigns from natural language, open to any developer.

Meta

TikTok's MCP Server connects your AI agents “directly to the TikTok Ads platform” to plan, launch, optimise.

TikTok for Business

Even TV wants server-side evidence: Netflix launched its own Conversion API in March 2026.

Netflix

Apple: AdAttributionKit attributes campaigns “so you don't need to use the AppTrackingTransparency prompt.”

Apple
Engagement scope

Rebuild marketing operations for the agentic era: strategy, planning, creative, and activation run by agents through the MCP sockets every channel now exposes, TikTok, Instagram, Google, Amazon, even Netflix, on a first-party, consent-clean foundation. We design the automation and the signal engineering, so every platform gets the conversion evidence its algorithms need without ever touching an identity it shouldn't hold.

  • 01Agentic campaign operations: strategy → planning → activation through platform MCP servers, with guardrails and audit
  • 02Cookieless, consent-first, server-side collection, first-party taxonomies designed for agent consumption
  • 03The agentic CDP, sovereign: Databricks / Hightouch as a no-copy context layer, decisioning on EU inference
  • 04Privacy-centric mobile measurement: SKAdNetwork / AdAttributionKit on iOS, no ad-SDK forcing the ATT prompt
  • 05Conversion APIs across every channel, Meta, TikTok, Netflix, Amazon, hashed, minimal, consent-gated
  • 06Identity resolution and clean-room activation that respect data minimization
  • 07Generative video at campaign scale: scripted, voiced, lip-synced, and rendered from code, in every format and language you sell in
Working session

Bring us one
marketing operations
workload.

In one 90-minute working session we map it across the frontier-sovereign axis, cost, risk, and strategic value. Then we tell you what we'd do.

Request this research

Working-session request