How can we use AI to bring new value to businesses? An investigation.Back Blog
What you can do with AI depends a lot on the industry you work in, but more importantly it depends on your approach - how you evaluate your possibilities and then how do you develop and build production ready AI solutions.
I am here to talk about this approach, the bigger picture, the process, as well as share our own journey how we learned all this.
AI - new technology
So, let’s take a broader look at AI - what are the two important things with any new technology, including AI, that we as companies and clients must deal with?
- What to expect from it
- How to work with it
We had new technologies before and biggest challenge with new things on the market is always expectation management and a nice neat process that leads to success.
Expectation example - SEO marketing
SEO marketing was new at a certain point in time. Back then a vast amount of time was spent to convince non-believers it will work or to prepare early adopters for potential failure.
Back then when SEO was the new kid on the block do you remember what question each SEO expert hated the most? Bring me to the top of the Google search page. SEO experts knew that this is not possible but they needed to spend a lot of time explaining how they can try, how it can work for some search queries or some regions but not for others, how people need to wait or how search engine changed algorithm at the exact time when the proposed changes went live.
A similar story is happening with AI nowadays. You need to be open and patient upfront, most discovery calls end up with you educating them what can and what cannot be done with data they collect, and most of the times you end up with a completely different idea than they had before the call.
Or you end up first building the data collection pipeline for them since they wanted to do data analysis without data but more on that later.
How to work with it
You heard of Andrew Ng, right? He created an AI course for everyone to address this issue and explain what are good use cases where to apply AI, right now. Check it out, it’s pretty informative even if you’ve never touched AI in your life.
The other thing with new technologies is the fact that they are, well, new. So we do not know how to work with them, we do not know which steps to take to make a successful project.
When somebody wants to build a new website it is sort of easy. We have been doing this for 20 years now. You have an idea, create some website content, designers create mockups, a moodboard, logo and all those shiny things designers call brand identity, and then the website developer needs to code it. You can easily verify if it is the same as in your design and you are done. Now you just need to call your SEO expert to bring it to the first place on Google right? So this is a clear process to follow.
So how confident are you in your expectations and your process? What results is it bringing to you? There’s a clear way to measure it :) Check how many projects are in production.
Not happy with the number of projects in production? Well, if you want to change this and have more production ready solutions, then you will have to change your approach. Here’s how we did it.
We learned this the hard way - a dozen projects over 4 years
We at SmartCat worked on dozens of AI projects of various sizes in the past 4 years, and more than 10 ended up in production. We have honed our process the hard way. Over time we began to discover what factors strongly influence whether the project will be a success or a failure - how we and our clients should approach our project.
Or, to put it another way, both our clients and ourselves made lots of mistakes, and failed a lot. Thankfully, we’ve learned enough that we now have a very clearly defined process that either takes you to success, or tells you early enough that you shouldn’t waste more money on this or any AI projects.
The result? Lots of lessons, a clear process and a realistic perspective on AI
And this is a brief overview. You want to start with understanding the business and for this we organize a workshop to sit down with you, speak about your day at work, ideas, pain points. Next up is to analyze the data you collect, what can we see in it, how we can use it. When we understand the business and understand its data, we can move on to the next phase: we can present you with a couple of projects you can do to improve your productivity with the data you are collecting. We choose the best one and proceed with implementation. First we do a web demo so you can verify we are on the right track and we then move to integration and deployment phase to bring this AI solution in production.
Pretty easy, right? NOPE. There’s a lot of potential pitfalls along the way
Along the way there are a lot of dragons sleeping. We have learned this the hard way so I am here now to share some of our experiences and hints on how to avoid the problems on this exciting journey.
Before we begin - One important thing on how to approach your first AI project… and in case you’ve already done them, please pretend your next one is the first one :)
ALL THE WAYS YOUR AI PROJECT CAN GO HORRIBLY WRONG
- Use case is not clearly defined (do not do analytics for the sake of analytics)
- Success criteria is not clearly defined - you want to prove certain hypothesis, AI is perfect tool and this one is good for expectation management, it gives common goal to both stakeholder and AI expert. Start with why, why you do this? What is important, how much?
So, we’ve done the business analysis, the project is good to go, now we just take the client’s data and voila! Right? OF COURSE the client has data, right? I mean they’re doing an AI project, data is like the only thing you need? WRONG!
- Lack of data - one of our first questions on discovery workshop, do you have the data or did you just read an interesting article. There’s a high chance we’ll part ways until you collect enough data because for data science you need DATA
- Non-labeled data - machine learning is letting machine learn from patterns, you need to teach it. When you want to teach your child to see the difference between cat and dog you show them cats and dogs, you need to do the same in ML
- Non-representative data - we do gap analysis and we will let you know if you have gaps in your data, this is part of our due diligence process.
- Stale data - data is not frozen in time, it is alive and it is catching patterns through time. We solve this using constant retraining and keeping a human in the loop. So, someone wanted to make an AI judge in US, and what do you think happened when it started processing cases? Was it fair, unbiased? Hint: they took the data between 1950-1990s.
So, once you’ve defined the data, you might think that your work is done… however. It just means you’re ready to transition from data to continuous improvement.
- Do not ignore the tuning phase - AI needs regular updates to grow and keep bringing value, just like your phone or computer
- Running costs - use it when you analyze your investment. Your solutions will run on infrastructure with huge amounts of data, infra and maintenance is not cheap
- Human in the loop - a good old fashioned human being is essential to check if everything is running smoothly and accurately
- Post delivery phase - you need maintenance and updates - like your phone or computer - because you need constant improvement, because you are building software for humans and humans change their behavior so you need to adapt. Since we do not have Artificial General Intelligence yet you need constant improvement through maintenance.
AND THAT'S BASICALLY WHAT YOU NEED TO KNOW - BUT THERE'S MORE
Data is everywhere, it is so easy to dive into it and end up being buried with technical solution but remember that we are here building data solutions for people.
We need to think about business goals, user experience, expectation management. We need to talk a lot and understand the real need behind solution we are creating. We need a good process which will put stakeholders in every step of the process so we can tune, steer or give up in a timely manner. Remember this is an experiment, hypothesis, basically RnD. This means there is a chance it might not work out. Pay attention to those mentioned pitfalls along the way. Share yours if you have them, send them to us, write a blog post, educate our future AI audience.
If you approach AI this way, you will have the highest success, you will improve the fastest and thus position yourself best in the future/reap the most rewards from AI. Or you will conclude you are not ready yet, save your investment and put it somewhere where it’s more needed.