Marketing ops · Production

Spotify builds a multi-agent architecture for AI-powered advertising media planning

The problem

Spotify's ads business had the same core decisions re-implemented per buying channel and surface, causing behavior to drift over time with no unified intent layer to understand advertiser goals and orchestrate Ads APIs consistently. Advertisers also had to manually configure all campaign dimensions without optimization guidance from historical performance data.

First attempt

The standard playbook of a new service with a state machine and REST endpoints did not fit because workflows are combinatorial and decisions needed to appear consistently everywhere. The previous manual campaign configuration had complex UI flows, no optimization guidance, slow iteration, and no access to historical performance data.

Workflow diagram · grounded in source
1
Advertiser submits natural language brief
Trigger
Advertisers generate optimized media plans through natural language interaction.
▾ source quote
“enables advertisers to generate optimized media plans through natural language interaction”
2
RouterAgent classifies message
Routing
The RouterAgent analyzes incoming user messages and determines what information is present.
▾ source quote
“The RouterAgent analyzes incoming user messages and determines what information is present: This fast routing step prevents unnecessary LLM calls and enables conditional agent execution.”
3
Parallel resolution agents extract parameters
Ai action
Specialized resolution agents each extract a focused dimension of the campaign: goals, audience targeting, budget, and schedule.
▾ source quote
“GoalResolverAgent: Maps user intent to campaign objectives (REACH, CLICKS, APP_INSTALLS, etc.) and searches for appropriate ad categories AudienceResolverAgent: Extracts targeting criteria including interests (from a predefined taxonomy), geographic targets, age ranges, and gender BudgetAgent: Parses various budget formats ($5000, 5k, …”
4
MediaPlannerAgent generates recommendations
Ai action
The MediaPlannerAgent takes all resolved information and generates optimized ad set recommendations using a heuristics-based engine backed by historical performance data.
▾ source quote
“The MediaPlannerAgent is where the magic happens. It takes all resolved information and generates optimized ad set recommendations using a heuristics-based engine backed by historical performance data.”
5
Optimized media plan delivered
Output
The system generates optimized, data-driven media plans in seconds.
▾ source quote
“generate optimized, data-driven media plans in seconds”
Reported outcome

The Ads AI multi-agent system reduced media plan creation time from 15-30 minutes to 5-10 seconds and cut required user inputs from 20+ form fields to 1-3 natural language messages, with every recommendation backed by historical performance data from thousands of campaigns.

Reported metrics
Media plan creation time (agentic)5-10 seconds
Media plan creation time (manual baseline)15-30 minutes
Required user inputs (agentic)1-3 natural language messages
Required user inputs (manual baseline)20+
Show all 5 reported metrics
media plan creation time (agentic)5-10 seconds
media plan creation time (manual baseline)15-30 minutes
required user inputs (agentic)1-3 natural language messages
required user inputs (manual baseline)20+
agent response latency with parallel execution~3-5s
Reported stack
Vertex AIGemini 2.5 ProgRPCPostgreSQLGoogle CloudApollo
◆ Does this fit your context?

Compare to your context

Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.

Compare to your context →
~30 seconds · free
Source
https://engineering.atspotify.com/2026/02/our-multi-agent-architecture-for-smarter-advertising
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The Ads AI multi-agent system reduced media plan creation time from 15-30 minutes to 5-10 seconds and cut required user inputs from 20+ form fields to 1-3 natural language messages, with every recommendation backed by…

What tools did this team use?

Vertex AI, Gemini 2.5 Pro, gRPC, PostgreSQL, Google Cloud, Apollo.

What results were reported?

Media plan creation time (agentic): 5-10 seconds; Media plan creation time (manual baseline): 15-30 minutes; Required user inputs (agentic): 1-3 natural language messages; Required user inputs (manual baseline): 20+ (source-reported, not independently verified).

What failed first in this deployment?

The standard playbook of a new service with a state machine and REST endpoints did not fit because workflows are combinatorial and decisions needed to appear consistently everywhere.

How is this marketing ops AI workflow structured?

Advertiser submits natural language brief → RouterAgent classifies message → Parallel resolution agents extract parameters → MediaPlannerAgent generates recommendations → Optimized media plan delivered.

Related marketing ops cases

WHAT TO DO WITH THIS

Now compare it to your context

This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.

Browse the Marketing ops cases →