How AI Agents Transform International Advertising Platforms

Large‑Model Characteristics

Since GPT‑3.5, large models provide deep semantic understanding, native multimodality, and chain‑of‑thought reasoning that can decompose tasks, call external APIs and execute decisions.

Commercial Platform Adoption

Platforms evolve along three axes:

Interaction redesign : How to retrieve and bill ads when users query an AI directly.

Creative productivity : AI‑generated assets replace manual creation.

Intelligent ad operations : AI assists or automates bidding, budgeting and optimization.

Amazon’s Rufus emphasizes precise transaction intent, Google protects its search‑ad core while adding AI, and TikTok’s Tako targets non‑standard goods.

Lazada’s Focus

Lazada concentrates on “intelligent ad operations” because interaction redesign impacts C‑side revenue and creative production is constrained by ad‑risk controls. The goal is to blend empowerment (Google, Amazon) and delegation (Taobao, ByteDance) models.

Overall Solution

Three fundamental gaps are addressed:

Execution gap : Converting natural‑language intent (e.g., “reduce cost”) into deterministic API calls without hallucination.

Supply‑demand asymmetry : Merchants lack awareness of abundant platform optimization strategies.

Lack of domain expertise : General LLMs miss advertising‑specific knowledge such as compensation rules and benchmark cases.

The solution is a knowledge‑driven, proactive‑reactive multi‑agent network that acts as an on‑demand expert and a background watchdog.

Design Principles

Bidirectional isomorphism : Both explicit user queries and implicit system signals are mapped to a unified “intent object” for shared reasoning brains, RAG knowledge bases and atomic tool libraries.

Retrieval‑augmented cognitive reasoning : Before any generation, relevant business knowledge (competition context, successful cases) is retrieved to ground responses and suppress hallucinations.

Cognition‑execution decoupling : The AI layer handles probabilistic tasks (intent understanding, strategy generation) while a deterministic engineering layer executes atomic tools.

Fractal collaboration & protocol first : Multi‑agent collaboration replaces a monolithic model; a Model‑Context‑Protocol (MCP) standard connects heterogeneous ad services with AI components.

Closed‑Loop Process (P‑R‑C‑D‑A)

Perception : Multimodal entry (chat or background monitor) captures user input or system signals.

Retrieval : RAG fetches knowledge from vector databases (FAQs, case libraries) and platform state (account balance, plan status).

Cognition : The reasoning brain routes tasks—simple queries to atomic tools, standard flows to composite skills, complex diagnostics to domain sub‑agents.

Decision : Value arbitration shows AI suggestions only when estimated benefit exceeds current operation; financial actions trigger a Human‑in‑the‑Loop safeguard.

Actuation : Commands are dispatched through a capability center; results update context and memory.

Architecture

A bus‑style design separates a central orchestrator from capabilities:

Unified interaction gateway with a Chat Adaptor streams dialogue.

Smart Slot Renderer injects AI suggestion cards into UI slots.

Context Manager maintains short‑term session memory and long‑term merchant profiles.

Knowledge Router performs intent analysis, queries vector stores and assembles prompts.

Planner decomposes tasks and selects tool‑invocation strategies.

Ability registration center exposes three tool modes:

Mode A – Atomic tools (e.g., getReport(), updateCampaign()).

Mode B – Composite skills (e.g., “new‑merchant rapid ramp‑up” DAG workflow).

Mode C – Domain sub‑agents for complex diagnostics.

Underlying data services include a Vector Database for embeddings, a Knowledge Graph for entity relations, and real‑time metrics from Data Services.

MCP standardizes interactions, allowing any future model to plug in without breaking existing services.

Deployment Scenarios

ChatLSS – Interaction Paradigm Reconstruction

Core challenges

Intent alignment: Natural‑language commands such as “close the under‑performing plan” must be translated into precise API parameters.

Zero‑tolerance execution errors: Mis‑interpreted budgets or accidental plan deletions cause irreversible financial loss.

Design principles

Pre‑confirm intent before execution (Human‑in‑the‑Loop).

Expose business logic as atomic tools for function calling.

Log every action and provide rollback capability, treating each operation as a database transaction.

The system combines a Workflow engine with ReAct to handle high‑frequency standard queries and deeper reasoning for non‑standard cases.

Paimon – Scenario‑Based Predictive Intervention

Core challenges

“Tunnel effect”: Merchants habitually use familiar ad formats (e.g., search ads) and miss new opportunities.

Low‑signal static suggestions: Rule‑based alerts are noisy and rarely clicked.

Design principles

Non‑intrusive embedding: AI hints appear as “co‑pilot” bubbles at critical touchpoints.

Full‑stack spatiotemporal awareness: The system knows the current page, recent behavior and historical preferences.

Probabilistic value ranking: Only the highest‑CTR/CVR suggestion is displayed.

The workflow follows the same P‑R‑C‑D‑A loop. Offline merchant‑profile agents periodically synthesize long‑term behavior into high‑level features; online intent‑understanding agents use RAG‑enhanced reasoning with real‑time context to generate strategy cards via a ReAct‑based planner.

Results and Outlook

In two fiscal years Lazada launched ChatLSS and Paimon, achieving measurable upgrades in operation modes and decision‑system completeness. Key architectural lessons include:

Bidirectional isomorphism for unified intent handling.

Knowledge‑first reasoning to curb hallucinations.

Cognition‑execution decoupling for safety and scalability.

Multi‑agent collaboration with MCP for seamless integration across heterogeneous services.

Future work will focus on:

Generative strategy recall : Incrementally surfacing high‑quality algorithmic recommendations during peak campaigns.

Interaction upgrade : Embedding assistant‑style bubbles throughout the platform to deliver timely, personalized nudges.

These directions aim to evolve the ad platform from a collection of tools into an AI‑driven solution that proactively assists merchants.