Back to articles

The AI-by-Design 6-Step Framework

•3 min read•By Ioannis Zempekakis
AI by DesignFrameworkCRISP-DMDouble DiamondPart 3

The AI-by-Design 6-Step Framework

Part 3 of the AI-by-Design series. Start from Part 1.

The AI-by-Design framework merges the designer's approach to problem-solving (Design Thinking) with the data scientist's expertise in AI innovation. It takes inspiration from the Double Diamond design approach and overlays it with the CRISP-DM data management methodology.

Design Thinking as the Backbone

Design Thinking is the structured and iterative approach which designers use to:

  • Empathise with their customers
  • Define and frame problems
  • Formulate and challenge assumptions through experimentation
  • Iteratively develop and improve solutions

"Design Thinking is human-centred, and places focus both on solving the right problem as well as solving the problem right."

The Gap Between Data Science and Design

By comparing the two approaches, we can see where gaps exist:

  • Data scientists have a natural tendency to jump fast into solving the problem, sometimes losing perspective of end-users
  • Designers may have unrealistic expectations of what is possible with AI and aren't always able to assess feasibility
  • Designers are often excluded from the final delivery step, yet AI's dynamic nature requires designed feedback loops

"Data scientists and designers need each other to create desirable, feasible and viable AI systems."

The 6 Steps

1. Discover

Build understanding of the project goal, customer needs and problems, and business opportunities. Usually involves customer research.

2. Define

Define the challenge scope — select a problem to solve or an opportunity to pursue. This includes researching the context and AI possibilities.

3. AI-by-Design Decision

Ask: Is this a problem that can and should be solved with AI?

  • If yes: assess which data is needed and research potential unethical consequences
  • If no: that's also a great outcome. AI is expensive and time-consuming. If alternatives can solve the problem, that should be the first approach

4. Develop

Explore different solutions and look into the data and modelling needed. Data scientists perform exploratory data analysis (EDA). Design for trust, transparency, explainability, and unbiasedness.

5. Test

Before committing to building and deploying, identify the riskiest assumptions and validate them. Build quick prototypes to test desirability. The first model we ship is often going to be the worst model.

6. Deliver & Evaluate

Iterate, refine, pitch, and deliver. But the process doesn't stop here — continuously iterate over the solution with a monitoring system and feedback loops to catch biases and data drifts.

The Sweet Spot

Successful innovation lives at the intersection of:

  • Desirability — We believe our customers want this
  • Feasibility — We believe we can do this
  • Viability — We believe this is worth our time and effort

All three should be taken into account. Find a problem that has big customer impact, is in line with strategy, and is technically feasible.

Next: Part 4 — The Framework in Practice: Real Cases from OLX

Previous: Part 2 — Why Do We Need AI-by-Design?