Four Principles For Using AI

Four Principles For Using AI

Pick a basic problem -- it is the single biggest determinant of success. The first version of Vymo only solved a basic need. Salespeople around the world found it tedious to report sales data. Without sales data, managers and leaders couldn’t forecast accurately or help their teams achieve targets, which impacted their topline directly. So, we built a mobile-first solution to detect all sales activities and then, based on sales data, our solution did what a manager would do. This is quite contrary to the general perception of artificial intelligence (AI) solving -- or exacerbating, depending on whose views you subscribe to on the matter -- all of the world’s most complex problems. Basic problems exist all around us, regardless of industry. To paraphrase Anand Sanwal’s advice on running a successful newsletter, it's best if you can find the things people in your industry say but nobody actually ever says out loud. That's not to say a basic problem is an easy problem. In fact, it's often quite the opposite.

AI Solves Basic Problems Efficiently

The major edge data-driven software has over software that just implements standard business logic is that it is more contextual and evolves progressively over time. For example, Vymo’s intelligent suggestions have different intervention thresholds for different types of salespeople and it varies over a period of time. The cumulative impact of this is real and tangible. Its impact is even greater if you can work with businesses to pick out the most useful data sources, find out where the bodies are buried in the data and construct good features to feed into your algorithm. The other advantage of solving a basic problem is that it is generally prevalent (and present in usable formats).

Build Based On Observations

The suggestions our AI gave also evolved. Our first version of suggestions was based on the premise that more sales activities led to more revenue. In simple terms -- more calls, meetings and other such interactions with prospects and customers increased your probability of meeting your sales quotas. It seemed like a perfectly rational thing to assume. In reality, though, only 30% of the best reps were in the top quartile with respect to meetings -- they just averaged higher conversions. This led us down a path toward understanding what activities had a higher return on investment (ROI). As an example, lunch meetings offered disproportionately high ROI for sales reps in wealth banking. We also looked at what leads, prospects or customers had to be prioritized for engagement (spoiler alert -- it's not just the leads that are funnelled by marketing).

The first major pivot was based on an important lesson -- build based on observations and not how helpful you think your application can be. Of course, you start with a basic premise, but once you have enough data to prove or disprove your model, your algorithms should run based solely on field data from end users. We prioritize builds based on what user data is telling us rather than cool new machine learning (ML) or AI capabilities we are excited about. Yes, we do try to stay cutting edge, but that never comes at the cost of being relevant to the user.

Tie Your Application To An End Goal

This is all a user really cares about -- how does your application tie to my end goal? We still use notepads in the digital age because they still serve a purpose. The same is reflected in an app’s usability, too. One of our most popular new features is "nearby," which shows the sales rep prospects and customers that are around his location. Compared to some of our other, more complex builds, this requires us to build a layer of intelligence on Google application programming interface (API) and then make it functional across devices and modes, which, while non-trivial, is definitely simpler than figuring out how prospects ought to be prioritized. So, forget your fancy models and algorithms -- what is the value that you are adding to the end user? A sobering test for this is the usability statistics of the app, which reflects, in reality, the most useful aspects of the app.

It Is Better To Be Vaguely Right Than Precisely Wrong

This brings me to the final point -- it is better to be vaguely right than precisely wrong. AI presents a tremendous opportunity to analyze and learn from large data sets. Often, the payoffs are disproportionately large relative to the costs (which are self-correcting, anyway). For instance, maybe the first few suggestions your algorithm makes are way off the mark, but if your experimental setup is right and your data is sufficiently large, it is bound to get progressively better. At Vymo, we run into corner cases, but if we didn’t expose our models to those data points, then who knows what we could be missing? If you had all of the power in the universe, would you prioritize doing great things or not doing bad things?

So, go forth and conquer!


Source: All the above opinions are personal perspective on the basis of information provided by Forbes and contributor Venkat Malladi.

No Comments Yet.

Leave a comment