6 steps to build your AI's project: get our new framework
July 24, 2017, 2 min to read
Just follow the steps, and read the cheat sheet for additional instructions and examples!
And if you want to print the canvas to fill it, you can use this version.
Problem to solve
- What is the precise business case you envision?
- Better said: what is the precise business case you envision, for which you think AI is the only solution?
- E.g. rejected transaction rate for a fraud detection problem
- By how much would you do like to reduce costs? To generate new revenues? To empower your team to spend time on more meaningful issues?
Next (IT) steps
- What do you plan to do with the results of your AI model?
- E.g. what happens after you measure your customers’ mood on social networks? Transfer the results to a customer satisfaction dashboard? Tie them to CRM actions?
Impact on the workforce
- Whose workflows does it modify? To what extent? Does the project replace existing tasks or create new ones?
Types of tasks
It is generally considered that there are 4 of them:
- classification (putting data points into the relevant categories),
- regression (predicting the value of coming data points given previous occurrences),
- anomaly detection (pinpointing which data points are out of a normal range of values),
- and clustering (segmenting data points in a few groups)
Types of AI approaches
- Take a look at our article "AI for Dummies" for categorization (e.g. reinforcement learning; computer vision)
Types of data
- E.g. video, text
- An important distinction is structured (e.g. well-formatted database with rows and columns) vs. unstructured (e.g. videos) data
- E.g. what is your storage capacity? Does the prediction have to be fast?
Necessary 4Vs of data
What do you need in terms of:
Resources (internal / external)
- E.g. leveraging your DMP or API-driven open datasets from government
Tools / solutions
- From the exploitation of off-the-shelf tools (e.g. Google Cloud Vision API) to the integration of tailor-made solutions
Frameworks / languages
- E.g. on your own servers and/or on some AWS instances
- Which roles do you need to define, and how many persons per type that you define? Do you hire inside or outside human resources?
Description of the learning process
- How / when / where does training of your model take place?
Description of the inference process
- How / when / where do you use the results derived from your model?
Description of the improvement process
- How do you plan to improve your model after it has been deployed - in particular through usage data?