RICHARD DALLAWAY A roadmap for find tricky solutions to business problems PROBLEM SOLVING WITH AI 

PROBLEM SOLVING VS “NOODLING ABOUT WITH DATA” 

STEPS 1. The problem 2. Your theory 3. Search 4. Testing 5. Trialing 

This is the problem (or perhaps “the dream”)? - “We want to automatically [ what? ] so that [ what? ]” - Can you define success? - Is there a smaller step? 1. “WE WANT TO AUTOMATICALLY…” 

- Hunch - Bet - Guess - …this is your theory. - (your experience plays a big role in finding a good bet) “I wonder if we could [ solve the problem ] by using [ data in reach ]?” 2. “I RECKON….” 

This is where you need some AI / machine learning experience. - Literature search (has anyone already done this?) - What would make a good model for this problem? - What data do we have? (are we allowed to use it?) - Would this model disadvantage any group? - What data do would we need? …and so, does this sound like something worth trying? 3. “WHAT IF WE TRY….” 

- Building a stand-alone experiment (running code) - Getting hands on: can we really do this? what gaps do we have? - Does the model work, perform well? - What confidence do we have in this? - How much data do we need to get to production? - What are the consequences of the model being wrong? 4. “LET’S SEE IF OUR HUNCH IS RIGHT…” 

- Run and monitor without using the results - There’s usually lots of plumbing work to do - Find out what is working and what is not - Feedback behaviour to improve the model - …and be sure you can rectify errors 5. “LET’S GET IT OUT AND SEE…” 

IT’S MORE INCREMENTAL THAN THAT 

THANK YOU @D6Y