1 of 34

AI Hype to�Product Impact

Lessons from a Real AI Transformation

2 of 34

Part I

The Foundation�& Strategy

3 of 34

Background & Challenges

Platform Context

A 10-year-old Ruby on Rails monolith built during the company's early development phases.

Strategic AI Evolution

Requirement to accelerate AI transformation. AI-native features + leveraging AI to enhance engineering delivery.

4 of 34

The Speedboat Strategy

Legacy Product

Greenfield AI Product

2 Engineers

1.5 PM/Designer

5 of 34

The Speedboat Strategy

Greenfield product

Decoupled from the monolith and bypassing legacy friction to focus purely on rapid AI innovation.

AI-native tech stack

Utilizing a modern, serverless architecture to leverage AI capabilities without technical dependencies.

Autonomous execution

End-to-end ownership of the roadmap, infrastructure, and deployment, ensuring continuous delivery and high-speed iteration.

6 of 34

Part II

From Hype to�Impact

7 of 34

From Hype to Impact

6X

Velocity Gain

Reduced GTM to 1 month and stable& satisfying product within 3 months

AI-first coding & product flow

Fullstack autonomy

Low communication bottlenecks

Product Quality

Rapid iterations drive 30% higher customer satisfaction.

Dogfooding: Internal use from Week 1

Customer Testing: Active from Month 1

Continuous delivery and iterations

Strategic Path

Triggered the API strategy and paved the way for future architecture

Successful POC for integrating greenfield products with core

Paved road out of monolith

30%+

*measured with limited sample size, will continue monitoring

8 of 34

Part III

The AI-Native�Toolkit That Worked

9 of 34

Vibe Coding & Prototyping for Real-Time Validation

Validating at the Speed of Thought.

Focusing on experience and flow over document. Demonstrating ideas with tangible interactive prototypes.

Vibe coded prototype for reliable user testing

Prototype rather than PRD as the main deliverable for engineers

[ SCREENSHOT: VIBE CODING WORKFLOW ]

Prototype as the main way of communication

10 of 34

Product Management Directly from Codebase

Bridging Tech Constraints and Product Vision. Treating the codebase as the primary spec to slash the "meeting tax," ensures realistic planning, and provides transparency.

Chat with code base through Cursor

Product discovery based on actual code complexity

Dependency and estimation

Progress tracking through PRs

11 of 34

Product Strategy and

Business Modeling

Utilize AI to run competitor analysis, dig into customer feedback, brainstorm on product values and model long-term impact of product decisions on core business metrics.

AI for brainstorming product ideas

Research intelligence to compile feedback

Business modeling

Vibe coded simulator to demonstrate impact of the key metrics with time to facilitate business modeling (the data in the screenshot is random)

12 of 34

Automated Documentation from Single Source

By generating documentation from code and a central knowledge base, we create a single source of truth that automates all help guides and GTM assets.

Convert project data and meeting notes into a smart knowledge base to auto-generate custom documents

Turn source code into living documentation: generate product guides and technical explanations automatically.

Auto generate document from code and knowledge base

13 of 34

Part IV

The Process that Worked

14 of 34

Full-Stack Autonomy for Accountability and Speed

AI made it more than ever possible for every engineer to own the full lifecycle—from UI and backend logic to testing and deployment. This ensures deep accountability for code quality and rapid delivery.

Every engineer in Speedboat is fullstack, they code both frontend and backend, write automated test and deploy the change.

Reduced communication overhead

Increased code quality through total accountability

15 of 34

Continuous Delivery for Rapid Iteration

High automation and cross-functional alignment allow us to ship incrementally. This slashes time-to-market to a month and creates a continuous feedback loop to stably increase customer satisfaction.

The target is not a product predefined 6 months in advance blindly, the target is only customer satisfaction and we make the best decision towards the goal in every moment.

Leveraging high automation for seamless deployments.

Collaborative search for iterative solutions.

16 of 34

Blurred Boundary of Product, Design and Engineers

PM does not write PRD/detailed requirements from beginning any more, they focus on explaining the problem and target together with designers with visual assets like prototype. Engineers are trusted to build the MVP solution. PM/ designers then participate in polish the product further with tickets as well as code.

In this way, we reliably delivered against all our goals and increased final product quality continuously.

Reliable planning and estimation

Realistic MVP and continuous delivery

  • PM and designers define the target with prototype/miro board
  • Engineers build a MVP that solves the core problem 60-70% towards the target
  • Designer schedule follow-up tickets to bring it to 100%

17 of 34

Dogfooding and Real Customer Involvement

We had a workable product for internal use from week 1 and customers using the product after 1 month.

Active usage of products led to higher empathy of internal team and concrete feedback from real usage, which in turn drives higher product quality.

Empathy comes from 1st hand experience

Lays foundation for active participation of engineers

18 of 34

Part V

The Hard Truths�& AI Pitfalls

19 of 34

The AI Pitfalls

Learnings from What Didn’t Work

The Expert Gap

AI augmented experts but failed to onboard non-experts.

The experiment we conducted to allow designers to "Vibe code" user interface in production brought more complaints in code review time than real benefits.

  • Figma MCP created flooded tokens
  • Duplicated components due to inconsistent naming convention
  • Too many changes in one go against engineering best practices

‘It takes more time to review the code than to write it’ - by angry engineers

20 of 34

The AI Pitfalls

Learnings from What Didn’t Work

AI-Generated Specs

Tickets/PRD created purely by AI drifted quickly. They were often unfriendly to human readers and not actively maintained with project evolving.

Our AI-generated PRD is estimated as 67 minutes of reading time :)

21 of 34

The AI Pitfalls

Learnings from What Didn’t Work

AI Generated False Alarm for Code Review

The AI code review attempt by external members proved to be full of hallucination, and due to the lack of contexts, many issues pointed by AI were controlled on API or UI layer.

AI can only serve as a starting point for code investigation, not the conclusion.

22 of 34

The Hard Truth

Transformation is 20% Technology and 80% People

23 of 34

Transformation in Theory

24 of 34

Transformation in Reality

25 of 34

Fragile Team Structure

New tech stack, new skillsets, and a small team led to a lack of backup talent. The team had to be carefully protected against any organizational changes.

1 change in the small team led to another change, quickly led to the collapse of the expertise.

26 of 34

Operational Friction

Using 20+ new tools/services required an efficient operational model to balance security, budget, and innovation, speed without creating new bottlenecks.

An experiment team with new tech stack means a lot of new software and services to implement. An effective operational model to empower the team with fast adoption balancing security and budget is essential to even kick it off.

X20

27 of 34

Information Overload

The "learning by doing" process required high autonomy but created pressure to learn new things fast while delivering immediate results.

The new reality changes every day. To be an experiment team means staying ahead of the trend and challenging the status quo on a daily basis while trying to deliver the business outcome.

28 of 34

Knowledge Sharing Gap

A constant challenge to effectively bridge the experiment team with the core team for knowledge transfer and to cover integration needs.

Knowledge transfer and organisational transformation is an intentional efforts that require dedicated time. Without proper organisational level support and intentional planning, the knowledge stay with the few pioneers and cannot be successfully spread.

29 of 34

Part VI

Reflection &�The Path Forward

30 of 34

To what extent do our experiences align with current industry best practices?

Small Autonomous Team: Small, autonomous teams are the operational standard for leading AI-first organizations. The values include increased agility, autonomy and effective communication, driven by productivity increase by AI. Confirmed by Google, OpenAI, Anthropic, Cursor, Pilot AI, OAK’s LAB. In some cases, 2-people teams are recommended, regular recommendations are 5-9 members per team.

The Full-Stack Expansion: AI tools enable a "new breed of versatile developers" who can operate across backend, frontend, and design silos. At Anthropic, security researchers now capably analyze front-end visualizations, and safety teams build transactional databases.

Fluid roles/Campfire model: cross-functional teams of researchers, engineers, and product specialists gather around a living prototype to sculpt its behavior in real-time, rather than handing off requirements through rigid silos.

MVP first: Start with MVP that solves 80% of the abstract problem without much human control and involve experts to polish the last 20%

AI for code understanding: the most common use case from Antropic engineers and researchers is for fixing code errors and learning about the codebase.

31 of 34

What else could we try or improve? (things we can do immediately without prerequisites)

Shared context: OpenAI encourages the use of unified context systems, such as committing AGENTS.md files to repositories to guide agent behavior via explicit conventions and guardrails. This institutionalizes the "intent" of the team, allowing agents to function with higher autonomy while remaining aligned with organizational standards.

Let AI builds MVP rather than engineers: Engineers use Claude Code for rapid prototyping by enabling “auto-accept mode” (shift+tab) and setting up autonomous loops where Claude writes code, runs tests, and iterates continuously. They give Claude abstract problems they’re unfamiliar with, let it work autonomously, then review the 80% complete solution before taking over for final refinements

Onboard non-PDE people to use codebase for product understanding: legal, support and many other people can use code base as the source of truth too for product questions

Empower fluid roles: proved as a key value in top companies that AI can expand the scope of people and help them get into the domain they couldn’t do before. We need to speed up filling in the gap for designers to work on production code as well as engineers to have enough contexts to fix small UI issues.

32 of 34

What Could Have Been Done Differently

01

Thorough Cost & Business Planning

Conduct a thorough cost planning and business plan simulation to avoid financial and strategic surprises.

02

Think Again if We Really What a Speedboat

Sign clear agreements for budget and measurement. Involve 'investors' correctly to protect team autonomy against fragility.

03

Early Core Team Counterpart

Embed a counterpart from the core team from the very beginning to drive knowledge transfer and manage dependencies proactively.

04

Strategic Project Selection

Pick projects like a startup. Define success clearly (e.g., customer satisfaction, new segment served) and align project selection to those metrics.

33 of 34

How is product operational model involving with AI in industry?

34 of 34

Part VI

Appendix