Having worked in an AI startup, I was exposed to scoping and architecting AI projects much earlier in my career. I’m now a Senior Data Scientist and have worked on numerous AI projects in the past three years.
Even now, I feel every new project I work on has its learning curve. In the past, as a team, we’ve made many costly mistakes. Mistakes that cost us time, money, and energy. We learnt our lessons. These are lessons you can’t find in any degree program or online courses. You need to be there in the industry to face them.
That’s why when I see people like Andrew Ng, who pioneers the adoption of AI across multiple industries, shares his knowledge based on his experiences, I religiously listen. I study them. I adopt them. You should too. They’re priceless gems.
This article was inspired by one such framework shared by Andrew Ng on how to plan AI projects effectively. Let’s get into the mind of Professor Ng, step by step.
Step 1: Identify a business problem (not an AI problem)
This hit me hard.
When a new client approaches us in our early days, we are often required to make a business proposal to help them adopt AI and transform their business operations. We used to approach it from the technical mindset: which AI problem is applicable for this business? Big blunder.
It would be best if you put the business first. You need first to identify a problem that is worth solving. Whether it can be solved or not can be dealt with later, but at this point, your focus needs to solely on the business, and it’s priorities.
The best way to do this would be through involving the domain experts. Every business has experts who have years of experiences in their respective industries. Involve them, let it be surveys or interviews or casual chats, but take their opinion. That’s the best way to identify the problem.
Once you’ve got your business problem at hand, what do you do next?
Step 2: Brainstorm AI solutions
We are all biased, some way or the other. We have our favourite algorithms, familiar solutions. Sometimes we believe that the latest AI research is the way to go. This can work out very well too.
The issue with this approach is you may be missing out on a better solution. This is a classic exploration-exploitation problem. The question is, how much do we explore as opposed to exploit what we already know.
At my current work, the answer we found to this is what we call “problem-solving” sessions. We sit together for a week or so and brainstorm multiple solutions to the problem at hand. There are no right or wrong solutions at this stage, and everybody’s opinion is welcomed.
After hours of discussing back and forth, only a few potential solutions emerge, which gets passed on to the framework's next step.
Step 3: Assess the feasibility and value of potential solutions
The best solution in terms of performance might not be the best solution overall. I’ve learned this through experience, and that was a lesson that will stick throughout my career.
AI solutions and expensive in terms of expertise needed, compute power and cost of implementation. An in-depth analysis needs to be done for the few potential solutions at this stage.
We could do this by consulting with domain experts, analysing what the competitors have done, or implementing proof of concept. A proof of concept has worked well for us, where we test the solution and gain the business’s confidence at the same time.
Based on the feasibility study, you’ve finalized the solution blueprint for the problem at hand. What’s next?
Step 4: Determine milestones
Nobody will accept it when you claim, “But it worked well on the test set!"
We need to identify metrics that the development team should aim for throughout the project. The tricky part is, as data scientists and machine learning engineers, we are comfortable with the machine learning metrics such as accuracy, precision, recall and so on. While it’s essential to meet these milestones, it’s also crucial to define some business metrics.
Some examples of business metrics are revenue generated, customers won back, user engagement, market penetration, etc. We can tie down every project to some metrics that matter to the business, and achieving these would gain the business’s confidence.
The business will realize the value generated through the AI project, and there will be no room for doubts when you set these milestones upfront. We’ve done so well; there’s one last step.
Step 5: Budget for resources
Everyone’s happy till they glance at the budget. But of course, everything comes with a cost.
You might be surprised to hear that most AI projects fail due to the unaffordable cost of resources. It’s worth thinking through the resources required to develop, deploy and maintain the AI project successfully.
Start with the timeline, the data, the development team, compute power, and supporting systems required to deliver the project to the business successfully. Once you list them, assign a cost to every resource and create a budget for the resources.
Bonus: Make the stakeholders commit to the plan
Please get the signoff on what has been decided.
This simple step introduces commitment from all parties and sets expectations regarding the project for the long run. The commitment is vital for the successful execution of what has been planned so far.
You can always iterate on the scope in case something has to change and communicate it.
Transform 2019 of VentureBeat predicted that 87% of AI projects fail and never make it to production. Not having a practical scope for AI projects is one of the major causes of this.
What makes this worse is that we can’t learn such frameworks through online courses or master's degrees. We can only learn these through experiences in the industry. Andrew Ng is one of the pioneers in the AI industry and presents his 5 step framework to plan AI projects effectively. We discussed these crucial steps in detail:
- Identify a business problem.
- Brainstorm AI solutions
- Assess the feasibility and value of potential solutions
- Determine milestones
- Budget for resources
Like you, I’m learning too. I was a beginner once and have made my fair share of mistakes. What’s more important is to treat these as learning experiences and advance in our careers. Keep learning, and I’m sure you’ll be successful too.
For more helpful insights on breaking into data science, exciting collaborations and mentorships, consider joining my private list of email friends.