In July 2017, our team of 4 stepped into the world of applied AI solutions with some interesting ideas and a lot of trepidation. We have covered a lot of ground since then, made some mistakes, broke a few things and built some truly awesome stuff. This is a good time to share some of those.


Customer support is broken. It can be fixed.

We had set out to fix the universally broken customer support experience.

We have all been there – explaining your issue for 10 minutes without anyone understanding it, being asked to repeat the same thing multiple times over and, of course, getting put on hold. Once we began to unravel why this was the case, we saw opportunities to fix the experience.

Enterprises work off a bewildering ensemble of software built anywhere between the 1980s to last month. The workflows often date back further to account for discontinued products and, quite often, the 80s software. Things get slow and confusing. Agents, who have weeks to learn both the systems and processes of their client companies, often bear the brunt of these inadequacies.

To fix the agent’s experience, we inserted ourselves between them and the software jungle powering their computers. We gave them a super-fast, clean interface where they see only what is needed in that moment and nothing else. And then we release the power of machine learning.

  • We pre-complete all routine tasks (think data recording, searching for information, prefetching customer’s account information) for them so they can focus entirely on the conversation with the customer.
  • We provide helpful indications throughout the case handling process to ensure a smooth experience. For instance, we show indicators that tell the agent that this case may need health & safety issues. Agents are able to provide the right context to their customers instantly.
  • We provide extensive directions to the agents about how to respond to the customer. This takes the form of appropriate content from knowledge articles as well as suggested full-form responses. This allows new agents as well as experienced agents facing a new type of query to resolve it faster and more accurately.

We have just begun this journey. Every day we envision new ways in which we can make the customer support process a seamless, positive process for people everywhere.

Machine learning ≠ Product. Machine learning = Team.

Being a machine learning-centric company, we often get lost in cool things that technology can build. This is even more so when the technology itself is evolving rapidly every day. However, in the world of enterprise software, you do not sell technology. You sell solutions. And that means focusing on what tangible benefits can machine learning provide (even when it is not cool). It means building stronger bridges between product, engineering and data science. It means ever so often aligning the team to achieve the business objectives.

At the same time, its the ability to get lost in cool things that ensure you have a team that can build great solutions. Machine learning practitioners today are faced with so many interesting choices, the only way to make sure they remain enthused and focused is to create a space to play. This means carving out part of the time for research, provision to participate in work-adjacent projects, contributing to open-source communities and involving themselves in research outside the organization.

A lot of this is easier said than done but the right mix of the two ensures you have happy people making awesome products.

AI includes human-defined rules.

We tend to think of AI as a self-contained, self-sufficient, all-knowing unit that learns from datasets and then operates on its own. As far as I know, the number of production systems that adhere to that image – ZERO!

Our AI platform attempts to mimic the actions of highly experienced, highly efficient customer support agents. The workflows it attempts to mimic have been created over the years by people. The funny part about human-defined processes is they come with weird exceptions, quirks, anomalies, and distortions. Most of these are not documented. Yet, not only does every agent on the floor know them, but they also know exactly when to apply them.

For example, workflows may define what to do when a customer complains about a product but agents take matters into their own hands all the time to resolve a complaint before it becomes a problem. Deep learning-based systems can learn this but take a lot of time and data. Instead, production systems will often come with many manually defined rules. The AI system must work in harmony with them to ensure there are no surprises.

Exceptions get noticed. Exceptions get measured.

For all the hype surrounding AI, the wow factor dies down pretty quickly. This is NOT an anomaly. In fact, this is the state that we aspire to every day – complicated processes becoming so routine that no one thinks about them anymore.

The technology, however, is not perfect. It WILL make mistakes. And within a few months, it is only those mistakes that will get noticed, not the thousands of others that you get correct. Once again, this is NOT an anomaly. This is the sign of acceptance of your product as well as the fact that your product’s success has now been assumed. Clients will expect you to get it right every time. Embrace it. Get better.

No machines for the bad news.

Once we were in production, we were getting great results for almost every kind of customer query. Some of these queries, however, were more sensitive than the others. Most of the customers writing into our clients were expectant or new mothers and, every once in a while, we would encounter someone who had recently encountered personal tragedy.

For machines, the margin of error in these situations is zero. Even so, we were able to generate great results. However, these situations required so much more than accurate results. They required a gentle, personal touch of an empathetic person. We built a filter to avoid any decisions to be taken by the machine and ensure a human agent goes through the case right from the start.

As a species, we are taking really quick steps towards letting AI playing a big role in our lives. However, when it comes to crunch situations, we do not trust anyone but ourselves. If you want to get people to trust your product, make sure you build in the right controls so someone can take over when needed. Over time, these controls will be needed less and less but today, they are absolutely critical.

Users are human!

AI has come to be associated so deeply with technology, we often forget that end of the day, it is a person who is using it. We also forget that we started out building the product for a person and not for operating in a vacuum. We have had more than a few situations where the product has been rejected. All of these complaints fell into two main categories:

  1. The AI workflow was different than their current workflow. Workflows are designed to improve efficiencies but process changes are hard and well-meaning products can bear the brunt of this inertia.
  2. The user expected a particular predicted action from the platform but received a different one. A specific level of inaccuracy is in-built into machine learning-based solutions but users do not care about this. They expect the product to work correctly 100% of the time.

To handle the above issues, one needs to put extra effort into understanding their users and their workflows. One needs to spend an inordinate amount of time observing their users make sure there are no surprises (no one likes surprises at work). One needs to ensure there is little new learning involved to simply use the product. This means putting special emphasis on how exceptions are handled, what errors tell the users and how elegantly machine-human hand-offs are designed.

Building machine learning-based products will get easier with time. Getting people to use them will always remain a challenge.


The last year has been interesting, to say the least. We have not only developed a solution to a real problem, we are seeing it deliver benefits to thousands of people every day. We have already set our sights for the next big challenge and will be ready with our solution before the year is out.

Onwards and upwards!