How Lenovo Reinvented MarTech with AI Agents and Data Fusion

Background and Traditional MarTech Stack

Lenovo’s legacy MarTech platform consists of six core modules:

Insight : Internet sentiment monitoring, semantic analysis and emotion detection for product, event and social‑matrix insights.

Marketing Activities : Multi‑scenario conference marketing with pre‑configuration, real‑time interaction monitoring and post‑event analytics.

Private‑Domain Marketing : Integration of public and private channels for precise user reach and fine‑grained data monitoring.

Content Management : Centralized multi‑source asset tagging, quantification and evaluation.

BI Analysis : Data‑driven dashboards and visual models for decision support.

Key Limitations of the Traditional Stack

Copywriting relies on individual creativity, causing delays and missing real‑time trends.

Material production, especially image assets, requires manual design; large promotions can demand hundreds of edits per day.

Channel deployment is fragmented across platforms (e.g., Xiaohongshu, JD, Douyin), each with distinct formats and standards, inflating personnel costs.

Data feedback is delayed; analysts must manually aggregate disparate metrics, reducing timeliness and increasing error risk.

AI‑Driven Marketing Chain

Lenovo replaces manual steps with AI assistance across five stages:

Strategy Insight : Real‑time market and sentiment analysis powered by large models.

Audience Selection : AI‑driven clustering and recommendation of precise target groups.

Content Creation : Text‑to‑image, image‑to‑image and text‑to‑video generation, with LoRA fine‑tuning for brand style.

Media Distribution : Automated adaptation and multi‑channel posting (social, email, SMS).

Data Feedback : Real‑time performance metrics feed back into user profiles for continuous optimization.

Self‑Developed Agent Framework

Lenovo built a proprietary AI‑agent platform to satisfy three business requirements:

Business Fit : Highly heterogeneous business units need custom logic that off‑the‑shelf tools cannot provide.

Technical Control : In‑house development enables full control over extensions, security patches and model updates.

Security & Compliance : Enterprise data remains on‑premise, avoiding cloud‑based compliance risks.

The platform offers three capability layers:

Component Tools : Stand‑alone functions (e.g., strategy generation, image creation, content distribution) callable directly or within multi‑turn dialogues.

Workflow : Linear execution paths for fixed processes such as Insight → Audience → Content → Distribution → Analysis.

Multi‑Agent Orchestration : Non‑linear, task‑oriented agents collaborate to solve complex scenarios, each responsible for a sub‑task (copy generation, media posting, etc.).

Core Capabilities

Strategy Insight : Aggregates massive online sentiment data, performs information collection, semantic analysis and emotion judgment, and produces insight reports for product or campaign evaluation.

Audience Selection : Merges Lenovo’s customer‑tag database with AI clustering models to recommend optimized audience packages.

Content Creation : Large‑model generation (text‑to‑image, image‑to‑image, text‑to‑video) accelerates copy and visual asset production; LoRA models adapt style to Lenovo’s brand.

Media Distribution : Unified tools adapt content to major platforms, email and SMS, supporting one‑click or scheduled posting; can be embedded as workflow components.

Data Feedback : Real‑time collection of performance metrics updates user profiles and drives continuous optimization suggestions.

Technical Challenges & Solutions

Data Fusion and Real‑Time Processing

Marketing data is scattered across online services, offline spreadsheets and third‑party sources. Lenovo built an Ultra ID system that unifies cross‑domain identifiers (user, device, cookie) into a single ID, eliminating data silos. Real‑time aggregation uses Flink State and Redis to update Ultra ID profiles with streaming exposure and heartbeat data, enabling instant, cross‑channel audience profiling.

Model Cold‑Start and Continuous Learning

Three‑stage approach:

Pre‑training (Cold Start) : Private deployment of models fine‑tuned on historical business data (labels, campaign results) and expert knowledge.

Continuous Learning (Prediction) : Clustering and predictive models ingest post‑campaign feedback, providing actionable suggestions (e.g., image‑text alignment, optimization tips).

Data Optimization : Secure aggregation of multi‑source experience data from marketers to further refine model performance.

Seamless Integration with Traditional Workflows

The platform provides a low‑code Agent management UI (compatible with MCP, plugins, knowledge bases) allowing users to compose tools, prompts and workflows. AI capabilities are also embedded in classic interfaces, so users can invoke generation or refinement without leaving familiar screens. Both chat‑style interaction and traditional UI coexist, supporting diverse user habits.

Quantified Impact

Efficiency Boost : AI‑assisted strategy, creation and execution increased overall campaign efficiency by ~30%, dramatically reducing manual workload.

Effectiveness Improvement : Data‑driven targeting cut redundant reach by ~35% and significantly raised conversion rates and user engagement.

Cost Control : AI‑generated assets lower design and copy costs; marketers focus on review and performance control rather than full content creation.

Future Directions

Deeper prompt engineering and parallel generation to accelerate text‑to‑image pipelines.

Extending multi‑agent collaboration beyond marketing to sales, service and support for end‑to‑end customer‑journey intelligence.

Continued refinement of short‑term and long‑term memory mechanisms to preserve context across collaborative tasks.