5 success factors for an AI project — No, it’s not Machine Learning!

After having done many AI implementations, here are some thoughts on what works in a typical AI/ML endeavor.

#1 Define the AI Goal

Artificial Intelligence (AI) continues to drive the technology discussion of numerous organizations and everyone wants to adopt AI solutions. But, first and foremost, you need to know exactly what business problem you want to solve with AI, how it shall integrate into your existing products and technology stack, and what AI enabled product features would demonstrate the most value to your business.

Below are some of the potential AI goals from few industries:

· Healthcare — Assist doctors by enabling your existing PACS system to identify potential abnormalities in CT scan, x-rays and other diagnostics images.

· Banking — Decrease suspicious ATM activities by analyzing ATM video footage and report any such incident when it occurs.

· Insurance — Automate the process of computing claims cost and estimate based on images of the accident provided by the policy holder.

· Ecommerce — Recommend personalized products based on buying pattern, demographics and preferences, to increases overall customer experience and sales.

#2 Back it with Data

Now that you have identified your AI goal, it needs to be backed by data.

A very common misconception: All you need to do is to build the right algorithm, plug in your data and start reaping the benefits of AI.

Data is a huge part of your AI project, and you need to start identifying datasets from various sources. Do note: the data that has to be collected need to be thought about not just in terms of number of records (length), but in terms of its width as well. For example, consider an online retailer’s database of customers spread out in an excel sheet. Each customer gets a row, and if there are lots of customers, then the dataset will be long. However, every variable in the data gets its own column, too, and you can now collect as much data on every customer — purchase history, browser history, text from reviews, customer demographics, date of birth — that the data usually tends to get wide as well.

#3 Involve your in-house Subject Matter Experts (SMEs).

Organization often hire external consultants to deploy new technologies. It will not be that simple with AI, in part because of the dependence on internal data, its structure and business context. For example, consider e-commerce companies https://www.amazon.com (World’s Biggest e-commerce store) andhttps://www.glossier.com/ (NYC based e-commerce beauty brand store). Both are in e-commerce space and recommend personalized products for its customers, but they both may have wide differences in their data model and context. So, data knowledge a product content analyst at Glossier.com has may not be the same as what an amazon.com analyst has.

Involve SMEs in not just the data collection stage, but in every stage of the AI life-cycle. For example, involve them in data labelling which helps find out outliers / biases in the datasets, evaluating the AI models, in testing the models before pushing to production, and overall measurement of the project outcome.

Look out for people internally in various departments like sales, support, customer service, accounting — who best understand the data and actionable insight around your business.

#4 Measure outcome, not output

Measuring the business outcome / impact of AI, is an important aspect of understanding how successful your AI initiative is. For some companies, these solutions are labor and cost savings measures, and should be analyzed accordingly. Other implementations are tied to revenue generation or increase in customer experience or satisfaction. While output of an AI project shall be the machine learning trained models, which has attributes like accuracy, execution time, recalls etc., you may not want to consider a model with the highest accuracy and low execution time as best for your business. This model could be biased and may work differently on unseen data or your changing business environment.

#5 Iterate Fast

Iteration is a central concept of machine learning, and it’s vital on many levels. Knowing exactly where this simple concept appears in the ML workflow has many practical benefits:

· You’ll better understand the algorithms you work with.

· You’ll anticipate more realistic timelines for your projects.

· You’ll spot low hanging fruits for model improvement.

· You’ll find it easier to stay motivated even after poor initial results.

· You’ll be able to solve bigger problems with machine learning.

From our experience, seeing the ML workflow from the perspective of iteration can really help beginners see the big picture and understand all the concepts behind machine learning.

To summarize, having the clear business goals/outcomes in mind, backing it with the right quality of data, involving your in-house experts who have deep business knowledge, measuring the outcome against the goals, and continuously iterating in smaller cycles would make you make fewer mistakes as well as take you closer to your AI/ML dreams!

Do let me know your thoughts or share your experience around it.