How to Build Successful AI Products: Insights on AI Development, NLP, and Product Strategies

Guest: Dr. Shao Hao, Vivo AI algorithm expert. Editor: Yu Wei. Platform: DataFunTalk.

Introduction: AI technology has attracted widespread attention from mobile and smart‑device manufacturers. Deep‑learning‑based natural language processing (NLP) drives many fast‑growing applications such as machine translation, recommendation, Q&A, and chatbots. This talk shares concrete thoughts on turning AI research into marketable products.

Current AI development status

NLP and voice‑assistant evolution

How to build a successful AI product

01

Current AI Development Situation

Technical thinking and product thinking complement each other. For example, a gesture‑interaction deep‑learning algorithm that improves accuracy by 5% and reduces inference time to 200 ms may look promising for TV manufacturers, but engineers often make four common mistakes.

1. Over‑ambitious goals: Trying to achieve tasks beyond current technology, such as recognizing animal language or having fully natural conversations.

2. Weak on‑device performance: Algorithms that cannot meet mobile GPU power or battery constraints.

3. Cost‑ineffective implementation: High data acquisition and cleaning costs that outweigh the benefit of replacing a familiar remote control with hand gestures.

4. Misaligned with user needs: Ignoring the core problems users actually care about.

Example: a new smartphone equipped with Snapdragon 888, LPDDR5, UFS 3.0, 120 Hz display, Wi‑Fi 6, and advanced cooling sounds impressive, but most users care about smooth gaming, long battery life, and fast charging rather than raw specifications.

AI Progress and Trends

Gartner’s AI maturity curve shows that machine learning, deep learning, computer vision, FPGA, and chatbots have entered the mature stage and will reach platform stability in 2‑5 years. NLP, autonomous driving, and artificial general intelligence are still far from maturity. The 2017 Transformer breakthrough and the 2020 GPT‑3 release accelerated NLP research.

Deep‑learning pioneers received the 2018 Turing Award, marking the transition from peak to maturity and shifting focus to industry applications.

In the past two years AI investment has cooled; many high‑profile startups (e.g., Anki, Wave Computing, drive.ai) have folded. Investors now prioritize tangible use cases, paying customers, and revenue over pure algorithmic breakthroughs.

02

NLP and Voice‑Assistant Development

Since 2018, AI funding has dropped about 65 %, yet AI features continue to appear on smartphones.

Two illustrative examples:

Computational Photography: Smartphones automatically correct and enhance images using software‑based algorithms rather than dedicated hardware. Originated by Google Pixel 1, it now rivals Apple’s A13/A14 chips.

Voice Assistants: Products such as Siri, Jovi, Xiaoice, Bixby, Xiaoyi, Breeno, and Google Assistant combine speech, NLP, knowledge graphs, and multimodal interaction, yet still fall short of human expectations due to technical ceilings.

GPT‑3 delivers impressive results but requires massive training cost and remains a probabilistic model; its answers are not guaranteed to be correct, which is problematic for deterministic on‑device products.

Voice‑assistant pipelines also suffer from wake‑word detection, noise‑robust speech recognition, and cocktail‑party effect challenges.

03

How to Build a Good AI Product

When technology falls short, product design must compensate. Key steps:

1. Identify the user: Who are they?

2. Clarify the need: What do they want?

3. Define the product: What should it be?

Mobile users have four core needs: connecting with others, using the device for media and productivity, accessing external services (e.g., food delivery, ride‑hailing), and interacting with other devices (e.g., IoT, car keys).

Product types include tool‑oriented, transaction‑oriented, content‑oriented, and game‑oriented. After choosing a type, set a clear core objective.

For tool products, user experience is the primary metric. Common proxies include Net Promoter Score (NPS), daily active users (DAU), retention, usage frequency, conversion rate, and share rate.

The RFM model (Recency, Frequency, Monetary) can be adapted for voice‑assistant products by treating “monetary” as usage duration or feature count.

Target‑user segmentation can be refined by demographic, behavioral, and demand dimensions.

To estimate product ceiling, use DAU forecasting: DAU(n)=A(n)+A(n-1)R(1)+A(n-2)R(2)+...+A(1)R(n-1) , where A(t) is new users on day t and R(t) is retention after t days.

AB testing, adherence to design standards (Google, Apple, Huawei, Alibaba), and ecosystem expansion can sustain growth beyond the expected decline phase, as illustrated by WeChat’s evolution.

Successful AI products require cross‑functional alignment: product, algorithm, measurement, research, and operations teams must each own clear sub‑metrics that roll up to the overall DAU or NPS goal.

In closing, the speaker notes that Apple’s Watch launch succeeded by focusing on everyday user benefits—efficiency, safety, health, convenience, and joy—demonstrating the importance of addressing core user needs.

Thank you for listening.