Domain-Coloured Glasses for AI

Artificial intelligence (AI) has been a hot, “new” thing for a while now. Its promise grew with the exponential growth of our computing and telecommunications capabilities, and the ease of data collection, storage, processing, and transmission. And now, we hear about AI being used in virtually every domain. Of course, there can be a debate about what can pass off as AI, and what cannot. However, in this article, I use the term somewhat loosely to refer to the slew of new, intelligent, data-powered products that use statistics, machine learning, NLP, or other forms of “AI”.

In particular, one domain that has garnered a lot of attention is the industrial domain. This is due to the large amounts of data that are already produced here, and the domain’s timeless and global nature. The promise of AI in industry is all about leveraging data to produce more value at a lower cost.

Over the last few years, I’ve worked on creating products that use AI for industries such as mining, oil & gas, energy, petrochemicals, cement, steel, glass, automotive, etc. The journey so far has highlighted one lesson over and over again — the need for deep domain expertise alongside AI to create any meaningful value. In other words, a team developing AI-based products in well established domains must necessarily combine experts in AI with experts in the field of operation.

Here’s why.

Problems come before solutions. And they aren’t readily known.

AI becomes the proverbial hammer if one only chases the “AI-powered” tag for their business or product. Of course, this is not sustainable. Therefore, the identification of the right problems to tackle with AI is a key task before developing any solution.

So far, in my team, we’ve seen that not every customer knows the problems to solve with AI. In fact, their only ask is “increase production, lower costs”. How you do it? That’s up to you. Including finding the problems.

In fact, every single customer engagement of ours has involved a phase of domain-related exploratory work to identify problems worth solving.

Therefore, in this environment, having enough domain expertise to find relevant problems at a scale that is both tractable and profitable is indispensible. More specifically, this is because:

1. Existing systems have many existing processes around them to address eternal pain-points. The way to developing an accepted, sufficient, and sustainable AI-based solution requires understanding the nuances of the pain-points and processes. This is possible through deep domain knowledge and a close engagement with the customer throughout the problem-finding process.
2. The higher value problem you find and solve, the more you get paid. This is self-explanatory.

Data is often incomplete, unclean, or too raw to consume.

Along with the identification of the problem to solve, it is also incumbent upon the provider of the AI solution to ensure that enough data is available to actually solve the problem. This includes enough data volume, enough variety, and enough quality. This means:

1. You must be able to identify key data-streams that are missing, or are unreliable. In other words, knowledge of the domain is required to ascertain whether all the factors that can influence a system are being reflected in the data. For example, if the temperature distribution inside a furnace is not known, the glass that is produced from the smelt of the furnace may seem to have unpredictable defects. Further, the data scientists’ ability to trust or distrust a data-stream is also increased if they know how that data-stream is supposed to behave relative to other data-streams and the system as a whole. This, of course, requires domain know-how and experience. On the flip-side, to improve the performance of the AI models, superfluous data-streams must be eliminated.
2. Data-streams often must be transformed or combined together to make an acceptable model. This is feature selection or feature engineering. In fact, we’ve also seen instances where a rule-based model (no AI, just physics equations) has provided sufficient value and has been preferred due to its simplicity and easy “tweak-ability”.

The buyer speaks and listens to domain language.

The people you are selling to, often (almost always), are not AI experts. They understand their business and domain like no one else in the world, and they LOVE talking about it. They also may not always appreciate AI in the way that you do. Therefore, talking domain is the way to pique interest. However, apart from building rapport with your potential customer, talking domain is crucial in marketing the value of your AI product in a way that a domain expert would appreciate. This, again, is only possible with deep domain knowledge and experience.

So, get yourself some domain folks to help you position, develop, market, and deploy your AI product. Simple.

Also, read about other lessons we’ve learned as a team in this article written by my colleague.