All Decisions
July 2019 accepted

Avoid Building AI Before the Data and Use Case Are Mature

aidata-readinessstartup-lessonsresponsible-ai
Context

At ILOVEGREENER, I explored AI-powered recipe and recommendation capabilities. The mistake was not using AI; the mistake was underestimating how much data quality, user feedback, evaluation criteria, and product clarity are needed before AI becomes genuinely useful.

Options Considered

AI-First Product Experience

Lead with AI recommendations and intelligent automation as the core product differentiator

Pros
  • Strong innovation narrative
  • Technically interesting
  • Good for demos
Cons
  • Weak if data is immature
  • Difficult to evaluate quality
  • Can create unreliable user experiences

Rules-Based MVP

Use simple rules and structured logic until user behaviour and data patterns are clearer

Pros
  • Predictable
  • Easy to debug
  • Faster to validate
Cons
  • Less sophisticated
  • May feel less innovative
  • Limited personalisation

Hybrid Learning Approach

Start with rules and simple ranking, then introduce AI only where there is enough data and clear success criteria

Pros
  • Balances ambition and reliability
  • Improves trust
  • Reduces premature complexity
Cons
  • Requires patience
  • Less flashy early on
  • Needs disciplined measurement
Decision

I learned to avoid treating AI as the starting point. AI should amplify a proven workflow, not compensate for unclear product value. In my current CTO role, I make sure AI work has clear data readiness, review mechanisms, evaluation criteria, and user value before it becomes part of the product roadmap.

Consequences
Nov 2019

AI features were harder to evaluate because success criteria and feedback loops were not mature enough

negative
Mar 2020

Engineering effort went into intelligence before the core product behaviour had enough real-world usage data

negative
May 2021

The experience made me more realistic about the gap between AI prototypes and dependable production systems

positive
Jan 2024

In later CTO work, I became more careful about responsible AI, human review, data maturity, and measurable outcomes

positive