在印度開發 AI 購物應用程式前,先測試全球助理無法複製的部分
Before building an AI shopping app for India, test the part a global assistant cannot copy
作者提出一套評估印度 AI 購物應用程式的設計,主張先驗證系統能否改善顧客的購物結果,而非假設本地開發自然帶來優勢;這套設計並非來自已建成或測試的產品。評估應核實商品款式是否有貨、到手價、配送資格及退貨政策,並以相同任務比較用戶現有替代方案的完成率、用戶修正、耗時與服務成本。文章亦建議把本地庫存、語言、售後、分發及履約等優勢視為待驗證假設,並計入對話結束後的營運成本。
An Indian AI shopping app does not automatically have an Indian advantage. If it uses the same models and broadly the same online stores as global assistants, its local edge has to come from somewhere else.
I am sceptical of building that thin interface as a standalone consumer business. But the engineering response is not to give up. It is to test where the system actually improves a customer's outcome.
The examples below are a proposed evaluation design, not results from a product we have built or tested.
Start with an outcome, not a better-looking chat
Choose a narrow shopping job. For example: find shoes in the right size, within the delivered budget, with an acceptable return policy and delivery to a particular PIN code.
A recommendation is incomplete if it ignores an unavailable size, a shipping charge or an unsupported delivery area. Do not grade the system merely on whether the product sounds relevant.
Build a small consented evaluation set of real tasks. Keep personal information out of public test fixtures. Include difficult cases: an ambiguous size, an unavailable variant, a changing price and an unclear return policy.
For each run, record:
- Whether the exact variant was available when checked.
- The total delivered price and when it was observed.
- The source for delivery eligibility and the return policy.
- Whether the user accepted the choice.
- Whether checkout or a later service problem broke the promised outcome.
Distinguish a verified value from an estimate. A search snippet is not an inventory check. A store's generic returns page may not apply to a final-sale product.
Compare against the actual alternative
The relevant baseline is whatever the customer already uses: a global assistant, a marketplace, a search engine or a familiar store.
Give the alternatives comparable tasks. Preserve the same budget, delivery area and constraints. Record the product versions and dates because agent capabilities and merchant coverage change.
Avoid announcing a win from a few favourable demonstrations. Compare completion rate, user correction, time to an acceptable purchase and total service cost across the task set. Keep browsing time separate from delivery time. A fluent answer can still produce a worse purchase.
If a baseline cannot complete checkout, say so. Do not treat it as having failed recommendation quality merely because it lacks your integration.
Treat shared commerce infrastructure as a dependency
Shopify's agentic-commerce announcement describes Catalog, Agentic Storefronts and integrations with several major AI surfaces. This supports a practical assumption: access to online product data may not remain unique to your app.
It does not establish identical access, prices, merchant participation or checkout support for every agent and geography. Verify your own agreements and actual supported flow.
In your architecture, separate the product source from the recommendation and action layers. Track provenance with each offer. Recheck changing price, availability and shipping facts before the user commits. Respect merchant access rules. Do not build the product around unapproved scraping or assume a public page grants a right to transact.
An Indian builder can benefit from open infrastructure as much as a global one. Lower integration cost is good for entry. It is not evidence of defensibility.
Put local advantages on an evidence board
Write each proposed advantage next to the evidence needed to believe it.
Local inventory: Does the product reach useful merchants or products the baseline does not? Is access durable or available to anyone who signs the same contract?
Language and voice: Can users complete the task more accurately in their preferred language? Test code-switching and clarification, rather than counting supported language names.
Returns and support: Who handles the unhappy path? Measure successful resolution, not the presence of a help button.
Personalisation: Does remembered size or preference improve the next purchase with the user's consent? Compare it with simply asking the same question again.
Distribution: Do users arrive through a channel you can afford and return without a repeated subsidy? A launch partner is not automatically a durable acquisition channel.
Delivery: Does controlling fulfillment improve reliability enough to justify its cost? If a partner owns fulfillment, represent that responsibility accurately.
These are hypotheses. A feature list does not prove that any of them is a moat.
Measure economics after the conversation ends
A shopping agent's cost includes more than model tokens. Depending on its role, it can include browsing, data access, payment fees, fraud, customer support, refunds and human intervention.
Keep confirmed revenue separate from expected commissions. Record the share of tasks that require a human, and the cost of those tasks. A cheap model does not make an expensive service failure cheap.
India's large AI audience also should not be mistaken for easy subscription revenue. TechCrunch's Sensor Tower reporting describes a gap between downloads and in-app purchase revenue. That is app-store data, not total market revenue, and historical promotions affected it. It is a reason to test willingness to pay, not a reason to declare that nobody pays.
Culture-specific products need a different evaluation set
Shopping is not the whole opportunity. My examples of India-native categories include astrology and learning the Gita. A global generalist may under-prioritise the exact experience that a local builder understands.
Do not evaluate a scripture-learning app only on fluent answers. Check whether the cited verse exists, whether its translation matches the source, which commentary tradition is being used, and whether the explanation is clearly separated from the text. Add review by qualified people and a correction route. Treat the learning workflow, recitation and reading progress as product components, not decoration around a chatbot.
The Bhagavad Gita app describes several of those features, including source-linked Gita GPT. It is a free non-profit product according to its operator, so it is a category example rather than proof of subscription economics.
For astrology, Astrotalk describes AI supporting human astrologers. Evaluate service quality and trust without presenting predictions as scientifically established or as substitutes for professional medical or financial advice. Co-Star's FAQ also describes AI use, so the claim cannot be that global builders are incapable of imagining astrology.
The test is whether cultural depth, a useful habit and trusted content create better outcomes. It is not whether the category sounds uniquely Indian in a pitch.
Keep operations in the product model
ONDC's shopping guide describes shopping through participating buyer apps across a network of sellers. It is a local-commerce route worth evaluating where relevant, not automatic exclusive inventory.
Mint's recent consumer-AI reporting also describes local use cases in learning, wellness and finance. These do not prove that any particular company will win. They show why the question is bigger than whether a product owns delivery.
The implementation question is not whether the chatbot was built in India. It is which customer outcome becomes better because this system exists, and whether that improvement survives competition and the full cost of delivery.
If the evaluation cannot find that improvement, adding another model or another chat screen is unlikely to fix the business thesis.
來源:DEV Community · dev.to