Turn Right in 500 Feet: Adopting a Machine Learning Mindset

Just get me to where I want to go.

Remember paper maps? Me neither. Our maps apps are amazing at getting us from Point A to Point B as efficiently as possible. 

Thanks to those apps, drivers everywhere have unknowingly absorbed the basics and the benefits of Machine Learning (ML). Machine Learning is a machine that continuously learns to suggest optimal, actionable decisions at any specific moment in time. Most of us don’t have the time or energy for route analysis and optimization; we’re more than are happy to hand off that cognitive load of deciding a route to a machine learning tool. 

Unlike the ubiquitous maps apps on our phones, Machine Learning is hardly mainstream in the business world, which is a problem. Only 10% of businesses in the United States have established ongoing machine learning capabilities and those skew to the bigger tech conglomerates. Product teams and developers everywhere else are using a standard technology toolkit to make their products and processes more efficient: A/B testing, decision trees, conditional logic.

It would seem as though ML is the obvious go-to tool to automate informed decision-making for known everyday business events (things like personalizing content on a website, determining a delivery logistics path, upselling a gamer on a power pack). So why isn’t it being used in more places?

Part of the lack of ML utilization in the workplace has to do with our discomfort with data. We know that oceans (lakes, sets, mountains, pick a term) of data can be useful. But entrenched biases in historical data sets skew results unfairly. Also, it takes time to analyze the vast data sets that encompass all user behaviors. Cool data analytics software (Tableau, Power BI or Qlik) only helps us visualize our data well after the point where any optimization opportunity could be acted upon. 

A data-driven moment in time should be optimized at that moment in time. 

Most everyday decision-making events in a product or process don’t require data scientist-like analysis, or even a data scientist to define. 

To be competitive machine learning needs to be mainstreamed. Not just building the model but the whole clunky process from start (data collection), to building models, using models to make decisions, and managing it all (ML Ops). Right now, almost all ML solutions require in-depth expertise, complicated piecemeal set-ups for different parts of a decision, and expensive ongoing management.

Select your destination.

Successes or failures in meeting our business goals are now determined in microseconds or minutes, not days or months. Two-week sprints are becoming the tech industry’s version of paper maps. Successful companies are the ones using ML to power the optimal paths for their customers and clients. 

It’s time to shift our mindset in the business world, similarly to when we evolved beyond static paper maps to dynamic digital directions. Let’s evolve the value of data from analytic naval gazing (which may or more not have predictive use) to using Machine Learning to process data that reflects real-time information to drive the best decision. Data is not the destination, the decision is the destination.

Decisions, Decisions, Decisions.

We read a lot about the great potential of Artificial Intelligence and Machine Learning (AI/ML)*. But there’s not a lot written about the pitfalls of dumb software.

Businesses use software to operationalize and scale their decision-making processes. This can range from internal operations or customer management platforms to external product recommendation engines for shoppers. Ideally, the software perfectly optimizes each task and each user to drive the most value for the business. Sounds pretty simple.

One problem for many businesses is that their decision-making software is relying on basic decision trees, A/B testing, case statements, and conditional logic tricks. The decision-making isn’t coded to learn and react to changing conditions. It operates under the assumption that if a change is needed, a human process will take over to analyze the data, evaluate that need, design a plan for a change, test out that plan for change and then put that change into the software. But in today’s world change doesn’t wait for a two-week sprint to complete (it’s too slow). The software we think of tried and true just can’t keep up. It’s tired. It’s true.

The software we think of tried and true just can’t keep up. It’s tired. It’s true.

Rewire Tired Software

Enter AI/ML with its power to continuously sift through data, find patterns, and constantly optimize decisions in changing conditions for individual customers. Amazing, right? It’s no wonder that 83% of executives cite AI as a strategic priority. So why is not every business using AI/ML to make better decisions, faster? Well, some are – certainly, we know about the massive investments in AI by Google, IBM, Amazon, Tesla. These investments have created market-moving value and growth in their businesses. But few businesses have the time, agile cycles, and resources to throw data scientists and developers at every potential use case. They read articles about AI/ML focusing on the complex, the massive, or the bizarre and think, “No, thank you.” AI has been thoroughly defined as costly, complicated, and time-consuming. But that’s about to change.

The Time is Now

The best product teams and leaders are masters at doing more with less, regardless of the size of their teams or business. Same with the next great advancement in AI/ML resources. We have reached an inflection point, where we can do a lot more with less. The costs of crunching data are rapidly decreasing, thanks in part to marketplaces that assist in data aggregation, and decisioning models become increasingly standardized. Making better decisions for every single use case isn’t a promise for tomorrow; it’s a reality today.

Everyday AI

An e-commerce shop or fintech firm doesn’t need the AI technology used to detect interstate money laundering activities to determine the best promo offer or payment plan to present to each user. A content creator doesn’t need to go deep on Tensor Flow to present personalized and relevant videos to their audience. What everyday business leaders need is the ability to optimize each pivotal moment in a path to ensure the ideal outcome. In other words, to make better decisions faster, every time.

All executives need for the next iteration of AI/ML technology is to know their ultimate business goals. They don’t need data scientists, data pools, large budgets, large teams, or MLOps. Soon businesses will be able to add AI in an afternoon. By the next day, the software will have started to learn. (If “soon” isn’t soon enough, how about now instead?). Software that can learn turbocharges capabilities by empowering businesses to offer increasingly targeted, personalized, streamlined options; ultimately driving growth and increasing value. Who doesn’t want that?

You don’t need a sledgehammer to drive in a single nail. You don’t buy training time on a TPU to decide the best time to present a ‘buy more lives’ offer to a mobile gamer. You just need a little AI.

*I use AI and ML interchangeably, and I am sure some of you may rightfully quibble with that. Technically AI is the umbrella under which ML sits. Meaning AI refers to machines acting “smart” and ML is a technique or application of AI where the machine learns and adapts to experience. There are even more definitional nuances, if you ever want to go down a rabbit hole. 

It’s The Little Things: Big AI for big problems and Little AI for everything else

Courtesy of TechCrunch. March, 2021

It’s the little things. Like when you wake up to find your phone fully charged, for once. Or when you head downstairs in a rush, in need of your car keys and they were exactly where they were supposed to be. Those are all wins; little problems solved that made your everyday life a little smoother, a little bit more efficient, a little bit better.

The little things. 

But when most of us talk and think about AI, we only think BIG. We watch dancing robots on YouTube, read about protein folding in science blogs, switch on our car’s auto-pilot mode, download “Not Hotdog” app, or discuss how a blue box schooled Ken Jennings on Jeopardy! We think about all those insanely hard problems that only AI with its extraordinary data-crunching capabilities can address. This is BIG AI: Using terabytes of data to solve (or create, depending on who you ask) all the biggest problems in the world.  But in our rush to radically change everything, we forget about the tons of little problems we face and solve every day. We might want to rebuild our homes with the latest technology, but first…we have to find where we left our damn car keys. 

That is where Little AI comes in. We’re surrounded by basic everyday problems, the ones it wouldn’t even occur to us to use AI or even machine learning to solve them. BIG AI has taught us that we need big datasets, big models (that demand big teams, and big consulting fees…), not to mention an endless appetite for compute time, to solve the enormous problems.

Our world is filled with Little AI problems. That doesn’t mean Little AI solutions are trivial  – it means that designing and implementing them simply isn’t hard.  Little AI works on common, less complicated problems: the ideal call to action on a home page; the best offer to show a customer to lift conversion; helping determine the fulfillment center to process an order or even the best time of day and the best channel to communicate with a customer.  Right now, we solve those everyday pain points using decision trees, conditional logic, manually-driven personalization, or even A/B testing. 

All those little pain points can lead to big problems or big efficiencies if solved with a Little AI.  Little AI uses machine learning to discover and implement the best decision faster and better than any traditional Agile process. So the next time you read about some remarkable new discovery made possible by AI, look around and consider the millions of other everyday problems that AI could make better. It’s the little things. And they’re everywhere. 

Oh, there are your car keys, exactly where you left them…in the car.

Agile Development is Too Slow for the World of AI

It’s time to yank down our software from deterministic decision trees and expand our impact. “Good enough” design and functionality based on Agile Development methodology isn’t good enough anymore. We have software capable of learning, adapting, and coming up with the best design and experience for each user. So why aren’t we letting it do so in our everyday lives?

Partly because Artificial Intelligence and Machine Learning (AI/ML) is viewed as hard and inaccessible. It’s time to move past the 1999-era Agile development mentality. The fact is that self-evolving software is faster and better at achieving design, content, or functional decision goals than the most crack developer team in a traditional Agile framework. Product teams should be thinking about strategic solutions, while our machines are busy figuring out, learning, and implementing tactical design and UX decisions. We’re the generals, our software the tacticians. If the generals are down on the battlefield trying to address every single interaction, we aren’t spending time thinking about winning the war.

What’s the cost of staying agile, rather than adaptive? Moving too slowly and missing the boat. Missing an edge case that could drive a next-level experience for a new user.  Missing a micro-trend that could earn us more money.  Missing out on forming a relationship with a new consumer whose online profile hasn’t caught up with her new six-figure job.

Often we’re not even aware of these misses because our technology is working, grinding away at offering preset programmatic solutions. Or we’re aware that we’re missing out on these edge cases but we have more pressing issues. But because leveraging even a single, slight advantage can cascade and culminate into millions of discrete moments that ultimately up-level our entire business or industry we’re actually missing out on incredible potential. The good news is that we don’t have to.

What feels the most foreign to the rigor of technology? Ambiguity. It’s time not only to get comfortable with ambiguity but to enable it – while still setting boundaries – in our software that designs user experiences. Often AI/ML solutions are complicated, cumbersome, and costly. It’s why the majority of risk, compliance, and design teams don’t allow their software to create, test, and implement user experience decisions.

Which is too bad. Because a machine that learns will make a better design, content, or functional decision. So what do we need to do to make that leap of faith, climb down from our decision trees, and allow our software to evolve? We embrace probabilistic over deterministic and allow our machines to learn more to do more.

The technology industry has always been comfortable with evolution. Our processes have evolved from client servers, to cloud, from waterfall to agile, from object-oriented to event-driven – in all cases – using processes that are planned, mapped out, specific: deterministic. Adapting to a new type of development model, a probabilistic one, will allow our software to learn and address all the possibilities inherent in our product and user experiences. In the meantime, we get to dream up ever more impactful use cases for our software to test.

It’s time to climb down from our trees. There are worlds to conquer in screens all around us. It’s time for us, and our machines, to adapt.