AliExpress AI Native Homepage Promotions: Automating 30+ Country Campaigns with Human Governance
AliExpress details its AI Native system for multi-country homepage promotions, replacing manual workflows with a knowledge-base-driven agent orchestration that automates 8 of 10 process nodes while retaining humans only for intent submission and final approval, using automated real-device screenshot validation across 30+ countries.
Background: The Complexity of Multi-Country Homepage Promotions
During an A+ mega-promotion, AliExpress must configure homepage placements for 30+ country groups, two core user segments (non-SMB returning users and SMBs), three platforms (APP, PC, msite), and two phases (preheat and formal). This translates to hundreds of schedules, each bound to datasets, tasks, and style packages, flowing through a state machine from pre-release to live rendering. Creative assets multiply similarly: dozens of country groups × ~10 image schedules × 5 slots = hundreds of manual uploads; at Double 11 scale, thousands of images across multiple languages per slot. A single missing image leaves a country's slot empty for that period.
The difficulty is not translation but four dimensions of per-country divergence:
Rule divergence: Each country group has unique slot definitions, targeting rules, and style-package matching keys; some lack dedicated groups and fall back to GLOBAL.
Data divergence: Slot configuration change points differ per country; schedules must be split at the union of all change points (e.g., slot 2 changes on 9.12, 9.14, 9.17; slots 3–4 on 9.12 → three segments). Missing a segment leaks preheat targeting into formal period.
Permission divergence: Country owners edit only their own floor schedules; homepage product only approves, does not edit.
Role divergence: Five roles (operations, design, dev, country assistants, product) each hold exclusive actions.
The core design principle: extract the common process, sink per-country differences into a knowledge base .
AI Native Definition
AI Native is not adding a chat box to a human-driven flow. It means first encoding business as machine-readable artifacts, then letting machines execute, with humans retaining only judgment and authorization . On the localized homepage: one shared process, country differences in knowledge bases, humans appear at two mandatory points — intent submission and sign-off.
Key Terminology
Country Group — Target country grouping (US, EU8, LATAM, GCC, OTHERS). Analogy: Which district the building sits in.
Long-tail Country — Countries without dedicated groups; rules fall back to GLOBAL. Analogy: No custom blueprint, use generic blueprint.
Activity Package — Promotion container per country with permissions, audiences, platforms. Analogy: Project charter defining who can modify.
Schedule — Placement record for a country, audience, platform, time window. Analogy: Which goods occupy which shelf for how long.
Slot — Homepage display position (e.g., Banner slot 1, Welcome Coupon slot). Analogy: A shelf compartment.
imageKey — Unique identifier for an image in a slot; essentially a Tair cache key. Analogy: Compartment ID.
Image Dataset — Collection of imageKeys produced for a promotion. Analogy: Batch of compartments opened per blueprint.
Placement Rule Knowledge Base — Per-country-group isolated rule collection; Agent's single source of truth. Analogy: Construction manual, revised per country.
End State: Ten Nodes, One Chain
Flattening the end-to-end flow yields ten nodes:
Business Requirement Submission
Intelligent Creative Production
Placement Rule Knowledge Base
Style Package
imageKey
Image Dataset
Activity Package
Placement Schedule
Release Approval
Time-Travel Validation
Before/after ownership comparison:
① Business Requirement: Before — Business → Design, back-and-forth on requirements & creatives. End State — Business submits intent & design asks once; Design Agent handles understanding, generation, size & rule validation.
② Creative Production: Before — Design → Country Assistants. End State — Design Agent outputs, assets flow directly into pipeline.
③ Rule Knowledge Base: Before — Business writes tables → Dev compiles. End State — Business & Agent multi-turn dialogue; rich text auto-converted to structured KBase.
④ Style Package: Before — Homepage assistant creates one by one. End State — Agent creates & publishes in parallel by platform, country, audience, component.
⑤ imageKey: Before — Country assistants upload manually. End State — Agent creates, validation pass = output.
⑥ Image Dataset: Before — Country assistants create slots + rally people to upload. End State — Agent creates slots; assets auto-matched by agreed path name.
⑦ Activity Package: Before — Homepage assistant creates → country assistants configure permissions. End State — Agent auto-identifies key info, creates package & backfills; country permissions shipped with package.
⑧ Placement Schedule: Before — Homepage assistant → country assistants per country. End State — Agent sharded generation, auto-matches creatives & prices, cross-validates.
⑨ Release Approval: Before — Homepage assistant → country assistants → product. End State — Humans retained: BPMS reviews only data & rules.
⑩ Time-Travel Validation: Before — Country assistants / product manually check per country. End State — Automated validation: time-travel screenshots per country, audience, time window.
Eight nodes automated, two human: ① intent origin, ⑨ responsibility endpoint.
Foundation 1: Rules as Conversational Knowledge Base, Not Documents
The starting point is not model integration but making rules machine-solvable. Knowledge base construction in the end state:
Business does not fill a form; they converse with an Agent over an interactive knowledge base — creative style, placement rules, country groups, audiences.
Agent holds historical placement analysis (creative sizes, audience-rule mappings), recent business asks, real-time category content; proposes options; business selects/edits.
Business submits rich text; Agent converts to structured KBase: homepage slice, MyAE slice, slot records, targeting rules, style-package keys each to their place.
Core validation runs before write ; non-compliant content cannot enter the library, no post-hoc inspection fallback.
Long-tail businesses follow standard template; Agent auto-completes, humans fill only deltas. Long-tail countries often outnumber primary ones; without this, earlier savings are eaten.
Four hard rules for any multi-country placement:
Four-dimensional targeting: country, page, audience, floor type. Missing one dimension misroutes US Banner config to EU audience.
Long-tail fallback path mandatory: Countries without dedicated groups use GLOBAL; if missing, fall back to original datasetId, never hard-fail. Fallback is primary design, not error handling.
Separate knowledge bases per business line per channel: Homepage mainline and SMB procurement are two independent KBase docs with own country groups. Same activity ID can exist in both with different content. Routing via promoName keyword at creation; promoName must not contain "采销节" or it routes to wrong library.
Domain-isolated knowledge bases: Homepage and MyAE use separate docs. Mixing causes cross-domain pollution and hard-to-locate issues.
Drafts generated in dialogue are not written to the library until confirmed. Business can iterate; Agent re-parses; drafts never pollute production KBase — this is data safety, not just UX.
Agent capability boundary is defined by the knowledge base: what's in KBase it does fast and accurately; what's missing it infers from context (plausibly but unreliably). KBase completeness directly determines usability.
Summary: The knowledge base is not reference material for the Agent; it is its sole source of truth.
Foundation 2: Operations as Tools, Permissions as Code
After rules become solvable, they must be executable. Two distinct tasks:
Expose Domain Capabilities as Tools
Five domain objects — Activity Package, Schedule, Floor, Style Package, Dataset — exposed via gop-model-s HSF interfaces to the Agent. Each tool does one thing.
Three examples:
Style Package: Previously designers handed files per country/platform/component; ops created one by one. Now Agent requests design platform protocol, fetches fileKey from OSS, splits by platform/country/audience/component, creates in parallel, batch publishes. Days of work → single run.
Image Dataset: Phase 1 solved "slot creation": Agent calls material center twice — first batch-create imageKeys with full path names, second create dataset from those keys. Humans still fetched hundreds of upload links, notified designers, notified devs to refresh. Phase 2 removed that: three parties agree on single path name (venue/country-group/schedule/slot); material update auto-matches unique imageKey, backfills & syncs Tair, sends completion notice. Formerly every hop needed human coordination; now auto-backfill, auto-notify. Rework changes from "re-run full flow" to "push update again". Human retains one action: upload image. Next step: design draft direct connection — design update triggers material pipeline automatically.
Activity Package: Previously per promotion, per country, per audience, per platform, then configure permissions individually. Now Agent reads key info from activity itself (which promo, title, time, platforms, countries), auto-creates package & backfills, ships country permissions with package. Audience info separately validated to avoid promo-side vs audience-side definition drift.
Chained end-state flow: Design draft parsing → create imageKey (tool-side validation pass = output) → Image Dataset → Style Package → Activity Package → Placement Task. Each step via tool; tool output = next input; no human data shuttling.
Placement Task Generation — The Costliest Failure Point
Tasks not created one by one but dimension-sharded: page → audience → country, each layer auto-completes, filters, changes, splits. Each shard step can fail: creative mismatch, price conflict, country config unsupported. Two mandatory actions in end state:
After auto-matching creatives & prices, cross-validation required ; conflicts stop the line.
Any failure in country config, data, capability, multilingual must be retryable per item, not whole-batch abort.
Encode "Allowed" in Code, Not Prompts
gop-model-s enforces a three-state machine on the publishing platform: SAVE → PRE-PUBLISH → PUBLISH. This is not a doc convention; it is forcibly executed in plugin-chain lifecycle hooks like beforeBatchPublish. Skip-step publish requests are intercepted. Agent decides "what next" but not "which step can be skipped".
Two iron laws:
Any Agent lightweight path must not bypass domain service mandatory plugin chains. Lightweight = flow, not validation.
Batch scenarios must consume per-item details; never collapse by dimension.
Idempotency at Run and Skill layers: user re-sends message, model retries tool call → no duplicate schedules. Invisible until a promo night explosion.
Result encapsulation: one replayable structured context per promotion — data sources called, configs read, delivery targets, validation results, business callbacks. Debug by replaying to exact step, not guessing from logs.
Conclusion: Model decides what to do next; code decides what is actually allowed to happen.
Foundation 3: Screenshots for Every Country, Every Time Window
Time-travel validation is the most human-eye-dependent node. Action: placement configured → open country homepage → switch to audience account → wait for effective time → verify. Multiplied by 30+ countries × audiences × time points = impossible manually. Automated in three steps:
Time-travel: System scans schedule effective windows, jumps to effective moment, captures real-device screenshot. One main task per country; one sub-task per schedule × language.
Replay: Results mapped back to "country × audience account × time-travel point" on the original submission card. Screenshots for country owners, not devs; anomalies red-boxed with explanation. "Who owns the schedule reviews the result" — saves more communication than any report.
Feedback loop: Fix & resubmit → second validation auto-triggered, round incremented, result bound to round. Later audits can answer "which version screenshot is this?".
Three must-keep designs:
Validation decoupled from publish action. Triggered by schedule effective window scan, so schedule time changes via approval are naturally compatible. If coupled as final publish step, any time change invalidates validation.
Validation is a sidecar. Inspection platform down → build mainline unaffected; single kill-switch. Positioned as "extra pair of eyes", not "mandatory gate".
No self-reported coverage. Coverage counted from downstream screenshots & render results, not someone saying "all checked" in chat.
Validation criteria become auditable evidence per country, per audience, per time point.
Human Positions in End State: Two Sign-off Points, One Data Port
After three foundations, the trap is "handily removing humans too". Multi-country business demands a pause.
Activity Package Example — Clearest Split
Old flow: Homepage assistant creates package → notifies country assistants to create floor schedules → country assistants send links back → Homepage product approves on BPMS. Country assistant = executor; Homepage product = accountable.
End state: Agent creates package once, ships country permissions with it. Country assistant opens submission page, sees only their country; initial creation triggers no approval; only when they modify their own floor schedule does BPMS fire, Homepage product appears — approves only, does not edit.
Two design judgments:
Initialization exempt from approval; modifications require approval. Not because "approval is tedious" but because nature differs: initialization executes already-confirmed rules, outcome predictable; modification changes a live placement, someone must be accountable.
Country owner edits only their own country. Not a frontend hide; permission model pushed down to activity package layer. If permission is only UI-level, someone will bypass via API.
Why Humans Remain at Specific Nodes
Business Requirement (①): Most primitive input. Business must clarify "what this version expresses, which creatives, how many segments". Agent can help ask questions, pre-generate creatives, pre-validate sizes & rules, but "this version" must come from business.
Release Approval (⑨): Never planned to automate. Approval is not a technical act; it's a responsibility act. End-state approval board shows only two things: data correctness, rule correctness. Execution details guaranteed by tools & validation; no need for human fifth-pass eyeballing.
Hard-coded boundary: Capital-risk parameters (promo start/end times) never given to model. Maintained by humans in config center; tool parameter interfaces lack the field; model cannot even pass it.
Data Port
After chain completes, one human action remains: review feedback data — live publish, CDN effective, customer service & pipeline feedback, product info return, tracking data. This is the only segment not given to machines, and not planned to be. All prior automation executes "known rules"; data feedback is the only signal telling us "are the rules themselves right?".
AI Native boundary here is not "can technology do it" but "should the machine make this decision". Technical feasibility ≠ authorization.
This End State Did Not Grow Naturally
Three structural transformations were required; missing any blocks arrival.
1. Entry Point Shift
End-state entry not in workbench forms but in business dialogue & submission cards. Business completes submission in chat: which promo, platform, country group, audience, creative copy, banner group — each with defaults; business confirms/edits, no zero-start Q&A. Submission lands on "My Promo Submissions" page: card shows promo name, platform, country group, audience, banner group; expand for preheat/formal timeline; post-build, timeline appends automated validation screenshots.
Human still in loop but only two confirmations: mid-tier confirms baseline info (once per promo init), submitter confirms upload (per country). Full-country submission complete → single unified run allowed; audiences split into returning, new, SMB batches; single batch failure isolates; re-run per card.
2. Orchestration: Super-agent as Router + Workers
Workflow form needs human guard at entry, exception, stuck. End state removes these three positions structurally, not just "open UI". Called Super-agent but not a bigger Agent; two roles separated:
Router: Only three jobs — classify ownership (new task or supplement), assign dispatch (to whom, what scope), judge completion (close, re-dispatch, escalate). Three hard constraints: no tool execution, no direct HSF calls, no content generation. Looks like capability limit; actually resource allocation: Router's asset is global view; if it acts, global view lost.
Workers: One task, one session, mutually invisible. Build Worker runs 7-step pipeline; Validation Worker handles screenshot diff. Isolation not obsession: 20+ write ops per chain; three countries' schedule params + two creative versions mixed in one context → impossible to pinpoint fault.
Completion judged by task ledger terminal state, not self-report. Validation dares full coverage because it depends on zero self-reporting.
3. Capability Ladder: Do → Verify → Improve → Discover
Four rungs: Do (machine operates), Verify (machine judges correctness), Improve (machine refines method), Discover (machine finds opportunities). Cannot skip rungs: Verify prerequisite for Improve — without trusted verification, "better than last time" unprovable; Improve prerequisite for Discover — method can't self-iterate, discovered ideas fossilize after one landing.
Interactive knowledge base and automated validation are the last two puzzle pieces: one governs "how rules enter", the other "how results are judged". Missing either, preceding tools & state machines only replace typing, not judgment.
Order dependency: Validation needs baseline first — without certain rules, screenshot replay only detects presence, not correctness. Conversely, conversational KBase enables full validation because comparison has a baseline.
In past runs, all placement countries except KR ran on Agent pipeline. Coverage expanded by pipeline extension, not headcount stacking.
What to Copy, What Not to Copy
This foundation now runs three capabilities — Build, Ops Q&A, Alert Auto-close. Different entries, same five phases: unify entry → pre-filter → judge dispatch → execute → return result to human. Below are transferable principles for any multi-country, multi-domain Agent-ization:
Copy These
Rules first. Don't rush to plug model. Answer: where do your placement rules live, can machines solve them? If not, write them down so business can author directly.
Converge differences into keys. Country, audience, platform, time window — four dimensions are universal coordinates for multi-country business. Fold differences into these four; code writes once.
State machine as guardrail. Which actions allowed — in code, not prompts. Model plans path; cannot skip steps.
Humans keep two seats: intent & signature. Separate "who accountable" from "who executes"; not automating approval lets automation run boldly.
Validation reads evidence. Not model self-report; downstream system terminal data & real-device screenshots, auditable per country, per time window.
Don't Copy These
Don't start with full multi-Agent orchestration. Changing six places at once → can't tell if skeleton or entry is wrong.
Don't expect model to replace approval. Approval value not in correctness judgment but in consequence ownership.
Don't merge knowledge bases into one. Domain-isolate; mixed pollution hard to locate.
Don't trust self-reported "coverage rate". If not counted from downstream, it becomes a nice-sounding lie.
Don't couple validation as a publish step. Once validation depends on publish action, any time change breaks it.
Closing Thoughts
Looking back, AliExpress's AI Native practice for localized homepage did nothing spectacular.
We moved business rules from Excel & heads into a conversational knowledge base, decomposed human ops into discrete tools, encoded "allowed" into state machines & plugin chains, folded per-country differences into four dimensions, turned validation from "someone saw it" into "screenshots exist for audit", then placed the model in the middle to chain these actions.
The real hard parts were never the model, but three mundane things: are rules written clearly, can mistakes be rolled back, who confirms completion .
One-sentence summary of our AI Native understanding: It's not about replacing humans; it's about turning humans from "form fillers" into "decision makers".
Multi-country business is inherently multi-centered: 30+ countries = 30+ judgment sets. Previously sustained by headcount; end state writes rules clearly, lets machines run those 30+ iterations, humans appear only where accountability lives — submit what this version needs, sign whether this version can go live, then watch data feedback tell us how to adjust the next version.
Then country owners' work shifts from "build schedules one by one" to "glance at screenshot, say this is wrong".
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
