Practice Areas

Agentic Digital Marketing

Run by Agents,
Judged by ROAS

＋60%

Return on Ad Spend

Days, Fully Autonomous

Europe

Multi-Market Media

Agentic

Always-On Operation

The Challenge

A Continental Media Operation Moving Faster Than Any Human Desk Could

Zara runs one of the most demanding paid media operations in European retail: dozens of markets, multiple languages and currencies, a catalogue that turns over in weeks rather than seasons, and demand that swings with weather, drops, and local events. Every one of those variables moves the optimal allocation of spend, and they move daily. A human media desk, however skilled, reconciles them on a weekly cadence at best.

The cost of that lag was structural, not occasional. Budget sat in yesterday's winners while today's intent migrated elsewhere. Bids held flat through demand spikes that deserved aggression and demand troughs that deserved restraint. Creative that had decayed kept spending because no one had reallocated away from it yet. None of this was a failure of talent. It was a failure of clock speed: the market was compounding decisions faster than the operating model could answer them.

Zara did not need another dashboard or another agency retainer. It needed the decision loop itself, bid, budget, audience, and creative, to run continuously at the speed the market actually moves, without surrendering the brand discipline and measurement rigor a retailer of this stature requires.

The story

The stakes

Zara runs one of the most demanding paid media operations in European retail: dozens of markets, a catalogue that turns over in weeks, and demand that moves daily. The optimal allocation of spend changed faster than any human desk could answer it.

The turn

The cost was structural. Budget sat in yesterday's winners, bids held flat through volatility, and decayed creative kept spending, not for lack of talent, but because the market compounded decisions faster than the operating model could.

What we changed

We handed the decision loop to a system of autonomous agents: bidding, budget reallocation, creative rotation, and audience expansion running continuously, governed by human guardrails and a Bayesian attribution layer that fed them causal signal instead of last-click noise.

Where it landed

Over a 60-day autonomous run, return on ad spend rose 60 percent against the baseline, on the same budget. The gain came from faster, better decisions, not more spend.

Figure 01 · Market context

The scale the decisions had to keep up with

€0.0B

Inditex online sales, FY2024

of group revenue, online

0.0B

annual online visits

markets, Europe its largest

Inditex, Zara’s parent, runs one of European retail’s largest digital operations: online sales near €10.2 billion and roughly a quarter of group revenue, across more than 200 markets. At that scale, a media desk reconciling allocation on a weekly cadence leaves return on the table every day it cannot keep up.

Inditex FY2024 results. Market context, not client data.

Diagnostic Findings

Where Return Was Leaking Between Decisions

Reallocation Lag

Budget moved on a weekly review cycle while intent shifted hourly. By the time spend followed demand, the demand had already moved on. The gap between optimal and actual allocation was widest exactly when it mattered most, during drops and seasonal spikes.

Flat Bidding Through Volatility

Bids were set to market averages and rarely adjusted intra-day. High-intent windows were underbid and low-intent windows overbid, so the same budget bought systematically less return than the auction actually allowed.

Creative Decay Left Unpriced

Fatigued creative kept spending because rotation was manual and periodic. Performance was lost not because better assets did not exist, but because the system could not retire and replace them fast enough.

Cross-Market Budget Silos

Each market optimized in isolation. Capital trapped in a saturated market could not flow to an under-served one with higher marginal return, leaving continental ROAS below the sum of its parts.

Attribution Without Causality

Reporting leaned on last-click signals that flattered the channels closest to conversion and starved the ones that created demand. Spend decisions inherited the bias built into the measurement.

24/7

Optimization cadence

Every auction, not every review cycle.

Figure 02 · Diagnostic

Where the recoverable return was leaking

Reallocation lag (weekly cadence vs hourly intent)0%

Spend followed demand only after it had already moved.

Flat bidding through volatility0%

High-intent windows underbid, low-intent windows overbid.

Creative decay left unpriced0%

Fatigued assets kept spending between manual rotations.

Cross-market budget silos0%

Capital trapped in saturated markets couldn't flow.

Attribution without causality0%

Last-click bias baked into the spend decisions.

An illustrative decomposition of the recoverable return gap the diagnostic identified. The leaks were not failures of talent; they were failures of clock speed, the market moving faster than a weekly review could answer.

Stochastic Minds diagnostic. Illustrative decomposition.

The Intervention

A System of Agents, Governed by Strategy

Autonomous Bidding Agents

Per-market agents adjusted bids at auction speed against real-time intent signals, inventory, and pacing. Aggression rose into high-intent windows and pulled back out of low-value ones, capturing the return the weekly cadence had been leaving in the auction.

Continuous Budget Reallocation

A portfolio agent moved spend across markets, channels, and campaigns toward marginal return, hourly. Capital flowed out of saturated pockets into under-served demand, so the continental budget compounded instead of fragmenting.

Creative Rotation and Renewal

Agents monitored fatigue per asset and audience, retired decaying creative the moment its marginal return fell below the alternatives, and promoted replacements automatically. Rotation became continuous rather than calendared.

Audience Expansion Under Constraint

Agents tested and scaled lookalike and intent-based audiences within explicit efficiency floors, expanding reach only where incremental return held and contracting the moment it did not.

Human Guardrails and Bayesian Measurement

Humans set the objective, the brand and bidding guardrails, and the efficiency floors. A Bayesian attribution layer fed the agents causal signal rather than last-click noise, so every autonomous decision optimized for incremental ROAS, not vanity proximity to conversion.

Figure 03 · Operating architecture

One orchestrator, four agents, one causal loop

Humans set the objective and the guardrails. The orchestrator routes capital and intent to specialist agents that act at auction speed; a Bayesian attribution layer feeds causal signal back into the loop, so every autonomous decision optimizes for incremental return rather than proximity to the click.

Operating architecture

The division of labor

Humans owned the strategy.

Agents owned the decisions.

Every bid, budget shift, and creative swap ran inside guardrails a human set, measured by a causal layer that kept the objective honest.

The Results

Sixty Days, Sixty Percent More Return

+60%

Return on ad spend

Across the 60-day autonomous run, ROAS rose 60% against the pre-deployment baseline, on the same budget envelope and the same brand guardrails. The gain came from better decisions, not more spend.

Days fully autonomous

Bid, budget, creative, and audience decisions ran continuously with no human in the loop. Strategists governed the objective and the guardrails; the agents ran the desk.

24/7

Decisions at auction speed

The decision loop moved from a weekly review to every auction. Closing the gap between optimal and actual allocation was where the compounding return came from.

The lift held because the architecture and the operation were one. We did not hand Zara a model and a recommendation deck. We deployed agents that made the decisions, governed by humans who owned the strategy, measured by a causal layer that kept the objective honest. That is the difference between advising on autonomous marketing and operating it.

Figure 04 · The run

ROAS over the 60-day autonomous run

vs baseline+0%

Indexed return on ad spend across the autonomous run, baseline = 100. The slow start is the agents calibrating; the compounding that follows is what a weekly review cadence could never capture. Illustrative shape.

Illustrative, indexed to the pre-deployment baseline.

Figure 05 · In context

Return on ad spend, against the market

0.0x

Fashion paid-media benchmark

0.0x

Zara, pre-deployment baseline

0.0x

Zara, agent-run (60 days)

The fashion ecommerce paid-media benchmark sits near 2.9x. A roughly 3.0x baseline lifted 60 percent lands near 4.8x. The benchmark is real market context; the client figures are illustrative of the engagement.

Benchmark: fashion ecommerce paid media, 2024. Client figures illustrative.

Zara · Europe

Run by agents.

Judged by ROAS.

+0%

return on ad spend · 60-day autonomous run