JD's Agentic Advertising Paradigm: AI Agents Replace Human Decisions Across Ad Chain

At the JDD 2026 Advertising Technology Forum, JD Advertising presented "Agentic Advertising Marketing New Paradigm." Speaker Zhang Zehua explained that generative LLMs bring not marginal efficiency gains but a transfer of decision authority: the advertising chain's decision-makers are shifting from "human" to "human + AI Agent" across three links — advertisers (placement), consumers, and platforms.

01 Decision-Makers Changed, Entire Ad Chain Reconstructed

Previously, humans made all decisions: advertisers adjusted plans, consumers browsed lists, platforms handled distribution. Now agents create a hybrid "human + AI Agent" mode. On the placement side, humans no longer monitor plans line-by-line; they delegate placement decisions to agents, and traditional product forms and bidding mechanisms cannot handle 7×24 decision density. On the consumer side, users no longer just search or browse shelves; they chat with human + AI to express needs, turning consumption decisions from "scrolling lists" into "selecting from dialogue results." On the platform side, general LLMs chat well but struggle to plug directly into ad-specific decisions — they must understand business goals, call real tools, and act within tens of milliseconds.

The three parties' goals diverge: advertiser agents want ROI, consumer agents want value, platforms want global efficiency. The question is not whether to adopt agents but who first solves the three links' reconstruction.

02 Merchants Define "What," Agent Handles "How"

JD's "Jing Xiaotong" (京小通) targets the third AI capability layer — autonomous operation. Three layers exist: (1) automation — human sets strategy, machine executes; (2) generative assistance — human asks, AI answers (copywriting, reports); (3) autonomous operation — human defines goals, agent continuously creates results. Jing Xiaotong is an Advertising Agent at layer three.

Mode: merchant defines WHAT (business goal), not HOW. Example: "This new product completes cold start in 30 days, maintains ROI, finds scaling room." Agent builds a business world model (product, store, audience, account, industry), planner decomposes goal into hypotheses, dynamic tasks, success/exit conditions, then calls tools across strategy, audience, creative, placement, analysis via ad APIs. Online decisions combine rules, LLMs, prediction models, Bandit, causal experiments, and Memory — not pure LLM.

Efficiency gains: previously dozens of daily clicks across product lines, 5+ steps to download creative data, then Excel for daily reports. Now "show all plan spend progress" → 5 seconds; "check last 7 days creatives" → 5 seconds, ~100% accuracy; "generate yesterday's report" → 2 seconds. Frees humans from data-moving to strategy.

Case Study: Shan Ze (山泽)

During big promotion, Shan Ze used Jing Xiaotong full-managed service targeting the 2–5 AM human blind spot. Agent captured traffic windows, auto-optimized, low-cost volume acquisition; night placement conversion efficiency rose 30%. Daytime too: competitor budget shifts, traffic cost changes, search term spikes — windows often minutes — agent 7×24 captures these fragmented opportunities.

Safety: unrestricted automation scares advertisers. Seven guardrails implemented: permission isolation, risk grading, execution simulation, key approvals, anomaly circuit-breaking, full audit, one-click rollback. Merchant hands goal to agent but keeps red lines on steering wheel.

03 Placement Brain: From Probability Prediction to Logical Reasoning

Core is the Advertiser Decision Reasoning Large Model ("Placement Brain"). Traditional ad algorithms excel at probability prediction (click-through, conversion) but business decisions need explanation: "add budget or pull back, why, how to revert if wrong." Requires shift from probability prediction to logical reasoning.

Three-layer model: (1) Base layer fuses platform ad data, merchant placement behavior effects, external industry hotspots. (2) Model layer: general LLM base with vertical ad knowledge fusion, learns placement decisions. (3) Capability layer outputs pre-placement strategy suggestions, intelligent diagnosis, full intelligent hosting. Three steps: dynamic perception — real-time traffic quality, competitive landscape, goal tracking, woven into dynamic competitive relationship graph; reasoning decision — identifies traffic depressions, gives executable personalized budget/bid suggestions; feedback loop — decision trajectories, reasoning chains, reflection, attribution fed back to training, model grows more industry-aware.

Results: advertiser adoption rate +15.4%, advertiser ROI +8.7%.

04 Interactive Shopping Guide: Capturing 70% Vague Intent

Consumer-side change felt earlier. ~70% of user visits are vague intent — not "iPhone 16 Pro 256G" but "good and cheap" or "gift for elder." Short phrases, scenarios, emotions — traditional keyword ads fear these: poor matching, high bounce, low conversion.

Need multi-turn conversational shopping agent to translate vague needs into product profiles. Commercialization under experience constraint: shopping guide can monetize but only if UX prioritized; if chat feels bad, any monetization damages. For advertisers: used to buy keywords; now must appear in the few dialogue rounds where user clarifies needs. JD APP now runs hybrid dual-drive: "search-recommend traditional model + AI model." Both shelf and AI shopping domains need stronger intent reasoning to translate vague needs into clear product profiles.

Think Then Recommend (TTR, WWW 2026)

Core: "think clearly first." User says "want a pair of shoes." Traditional system grabs "shoes" and pushes bestsellers. But user browsed rackets and sportswear last week — shoes likely for court. TTR splits into three agents: basic intent extraction — judges current need from multi-turn dialogue + current query; fine-grained intent generation — historical behavior completion, cold-start popularity fallback, grows "want a pair of shoes" into structured intent with brand, category, price; preference-aligned recommendation — vector recall, then semantic alignment, re-rank, outputs results with reasons.

RIGER (SIGIR 2026)

Solves "after understanding, how to productionize." Understanding alone insufficient; must inject into ad recall pipeline — map expressions like "Spring Festival, gift, specialty" to product-system-recognized Intent SID. Online: intent SID + behavior SID fusion, exploring latent needs while preserving existing preferences. Results: user understanding capability +227%, online click +1.6%, spend +1.3%, strict cold-start new items +4.08%, serving 190M users. Direct ad impact: previously invisible needs in behavior now recallable; cold-start products get more chances.

05 Foundation: Open Architecture & Super-Node Compute, Determining If Agents Reach Production

Two infrastructure pillars:

Open Architecture

Not just another API set; provides complete intelligent agent runtime environment enabling external agents and platform agents to communicate and collaborate — Agent-to-Agent (A2A). Three-layer open design: (1) Client access layer — A2A gateway handles auth, rate-limiting, routing; external agents plug in. (2) Dedicated runtime — large clients get isolated dedicated runtime with version management and rollback. (3) Capability open layer — platform core ad capabilities encapsulated as Skills and A2A toolsets, strict authorized boundary access. Key technical direction for next-gen ad platforms.

Compute Foundation

Many agents need persistent memory and fast multi-round reasoning; core pain points: ultra-low latency and massive KVCache. Past year: introduced unified compute + intelligent compute super-node system. Cross-node communication latency -90%, cache pool expanded to 300TB+, supports thousand-node, hundred-TB KVCache shared caching. Three key technologies: URMA cross-node high-speed transfer — solves large KVCache object cross-machine transfer bottleneck; OS heterogeneous pooled memory management — connects HBM, local memory, SSD multi-level cache; super-node communication — million-level RPC, zero-serialization copy. In plain terms: fast reaction, big brain, high throughput. Without this foundation, large-scale ad agents stay at demo stage.

Speed Guarantee

Unlike chatbots, business opportunities fleeting; ad decisions often millisecond-level — agent however smart, tens of milliseconds late, bid window gone. Ad agent must do LLM inference, vector retrieval, and massive business logic. Performance not from single-point hardware stacking but algorithm–compiler–OS–chip full-stack co-optimization. Collaborated with domestic/international teams from bottom chip/communication up, to OS customization, AI compiler customization, even retrieval algorithms tuned at 0.5ms granularity. Ultimately top-layer ad agent decision speed reaches tens of milliseconds.

06 Conclusion: Making Ad Placement Smarter

Agentic placement is not simply "LLM + advertising" but a new paradigm jointly supported by B-end advertiser efficiency, C-end consumer understanding, and a full-stack compute capability behind them. JD expects to work with merchants to make ad placement smarter through Agentic.