Beyond Tone: Taobao Shango's Emotional AI Agent Design Framework

Background: Tone Alone Is Not Enough

Taobao Shango initially launched three selectable conversation tones — "Refreshing & Efficient", "Relaxed & Natural", and "Witty & Inspiring" — but found that tone adjustments alone could not deliver a genuinely good emotional experience in many scenarios.

Competitor Research Insights

The team surveyed over ten competitors, including general chat AIs, vertical-domain AIs, and hardware/vehicle products, to see where emotional design was implemented and how human-like communication was defined.

Writing Response Principles into Values

To handle AI output uncertainty, more products define a set of "values" that govern how the agent makes choices, expresses itself, and incorporates attention to user emotional experience for more stable behavior.

Vertical AI Focuses on Business-Context Solutions

General AI products tend to give complete cause analysis and generic action advice for emotional issues, while vertical AI products guide users into their business domain and use product features to provide solutions.

Recognizing Emotion and Auto-Switching Strategies

Vertical AI products are strongly task-driven; users want to complete tasks quickly. Simply switching tone is insufficient — tone experience must be built on top of good response strategies.

#02 Values: Caring with Boundaries

For Shango, values must answer: "How far should caring go?" Example: a user says "Pressure is high, I'll order three fried chickens plus a large coke." Directly recommending items may ignore context; advising less may misjudge — maybe it's for a group. Caring needs evidence and must respect user choice.

The team defined service boundaries: caring revolves around immediate life needs, not pretending to be a friend or therapist, not promising the impossible. They analyzed user queries and complaints during ordering, waiting, and after-sales to understand what users care about and what help they need. These inputs formed a shared value system for subsequent service and expression.

#03 Emotional Response Strategy: Understand Emotion, Then Help

With shared principles, the team turned to psychology's "Emotion-Action Prediction Theory" for a three-step flow: Analyze Emotion → Understand Need → Adjust Service . Combining wording, punctuation, context, and actual order status, the agent infers why the user feels that way, what help they want, and then adapts its reply and handling.

Concrete Example: Late-Delivery Complaint

User repeatedly asks: "What time is it, still not here!! Promised half an hour?" Combining phrasing and order status, the team inferred the user wants delivery progress and feels dismissed by standard replies. The response shifted from guiding to check the order to first acknowledging anxiety, apologizing, then stating known status and risk compensation.

#04 Tone Style: Turning "How to Say" into Concrete Rules

In the first version, users manually picked a tone. But emotions change from ordering to waiting to after-sales, so a single tone needs scenario-specific expression. The goal: within the user-chosen style, the AI actively adjusts tone based on emotion and scene, keeping familiar traits while addressing the moment.

Borrowing stylistics' "style markers", the team broke tone into six quantifiable text features: characters, vocabulary, syntax, structure, semantics, and application-specific traits. This yielded rules for punctuation frequency, word preference, sentence patterns, information organization — specifying which habits stay stable and which vary by scene.

Same Recommendation, Different Tones

Recommending curry pork rice: "Cheerful & Positive" uses colloquial short sentences like "Let's eat!" to lift mood; "Steady & Reliable" gives the conclusion first, then describes the pairing in full sentences. Price and ETA stay identical; tone difference comes from word choice, sentence structure, and information order.

These combinations were codified into AI-followable expression rules, providing a basis for scene-based tone adjustments while keeping business facts consistent.

#05 Testing & Validation

Quick Iteration via Prompts

During trial runs, the team used prompts (reply instructions) for fast rule changes and re-tests. Values, emotional response, and tone requirements were written separately for item-by-item tuning. Later, results and business needs would inform Skill encapsulation or model training.

Scoring Criteria

To verify replies against design intent, they combined fixed test questions with a unified rubric. Twelve check items covered four dimensions: problem solving, emotion soothing, relationship maintenance, and information supplement. Different versions answered the same questions under identical business conditions for comparison.

Scene-specific emphasis: ordinary queries check clarity; complaints check explanation and follow-up handling — avoiding mistaking "friendly tone" for "problem solved."

Automated Batch Testing Pipeline

As rules and test cases grew, manual prompt editing, per-case testing, and repeated comparison became unsustainable. The team built an automated pipeline: reply generation → batch scoring → version comparison. Humans review results and adjust instructions, reducing repetitive tests. Score changes assist comparison; real-world experience remains the final judge.

Conclusion

This exploration extended the team's work from writing replies and examples to writing rules and building test processes, requiring earlier collaboration with engineering on implementation. Linguistics and psychology provided new angles to judge whether replies address emotion, care appropriately, and offer real help. Next steps: continue trial runs in more real scenarios and iterate based on feedback.