How Startups Can Leverage Artificial Intelligence and Machine Learning

Startups today are in an exciting phase. Amazon, Facebook, Google, Netflix, Tesla, and many of the world’s biggest household names in technology were small fledgling startups not so long ago. 

Even now, every few days, we hear of another startup like Spotify, Airbnb, or something or the other becoming a unicorn, going public, and before we know it, becoming an integral part of our everyday life – replacing companies and brands that had been a part of our lives for as long as we can remember. If not all, most of them can do so because of the tremendous advances that we have seen in technology in the past few years. 

AI & ML services today offer solutions to many problems that we considered insurmountable in the past. They improve existing solutions in cheaper, quicker, and more efficient ways. While it can be challenging to manage expectations from Artificial Intelligence and Machine Learning efforts even today, the results can be worthwhile if done right.

Here are some ways in which startups are uniquely positioned to leverage their flexibility to make AI and ML efforts successful.

Access a lot of data

It is the era of Big Data, and you can’t do AI or ML today without data. Depending on your product’s scale, you need to ensure that you have access to enough data on the problems you are trying to solve. Startups, by being small, tech-savvy, and operating at scale, use technologies to digitize whatever material they can, automate whatever tasks they can, and often capture large amounts of data for training models and building AI products.

According to an MIT Sloan Management Review study, 59% of the responding companies have an AI Strategy now as compared to less than 39% in 2017.

Machine Learning projects often run into a lot of cold-start problems, not having data resources and whatnot. Manual effort and one time solutions can get you out of quick pickles. Still, to eventually come up with a scalable and long-term solution, you will need access to large amounts of data on not only what works but also what doesn’t by tracking large datasets over long periods; and startups that realize this consider data as an asset that they actively try to build.

Results of a recent survey of ML Practitioners on their current data requirements (Source)

Find the Right Fit

According to VentureBeat, 87% of AI projects don’t make it into production.

Companies often need time to figure out their product/service concretely, what features it can realistically offer, how it distinguishes itself in the market, and many things that can only happen through trial and error. Large corporations come under a lot of scrutiny by both the public and investors. 

At the same time, startups generally can spend a lot more time exploring multiple options and finding the right fit for themselves. One of the best examples that come to mind is that of Slack Technologies. 

Currently one of the world’s most popular chat and productivity tools – used by 77% of the Fortune 100, Slack was initially a gaming company called TinySpeck. After shutting down their MMORPG Game “Glitch” in 2012, they went on to port the game’s “familiar” pet rock into Slackbot, a friendly bot that helps you find things around, gives you notifications, and keeps you informed about the workspace. Slack was launched in 2014, got listed at NYSE in 2019, and was acquired by Salesforce for $27.7 billion in cash and stock.

Minimal Bureaucracy

Compared to large enterprises, startups are more flexible and dynamic in setting up business and cultural procedures, ensuring that they set processes right with a long-term vision right from the get-go. People wear many hats at startups, but there is also a lot more active direct communication and understanding between the business teams, product teams, data science teams, and domain experts.

Communication in startups is more comfortable, and analyzing the problem from multiple angles and understanding a lot of different perspectives is simpler. Because of the closer working relationship among team members, it is easier for them to get access to any datapoint that they may need, get resources to test out different technologies and operation processes until the right fit for the team(s) has been discovered. This system is also present in more prominent companies, but in startups, the system is cheaper, quicker, and easier to manage.

Building a Knowledge Base

While large corporations often handle an entire end-to-end solution, startups generally spend a lot more time carving out a niche and pursuing excellence first in that small niche. Startups have smaller, more connected teams that are all focused on solving that particular niche problem quickly. Storing information about the difficulties faced, sharing lessons learned is more straightforward. Startups can afford to be flexible, make mistakes, fail fast, and work out solutions to what works or will eventually. This builds a domain-specific knowledge base about the problem that serves a critical role in building AI-based products.

Domain expertise not only tells you about the problem you’re trying to solve, but also about what others have already tried and what works or not and why, so that your data scientists not only know what behavior is expected from the models and why, but can also identify if a model passes sanity checks, is using irrelevant features to overfit or “doing too well” or underfitting and “playing too safe.” To build AI models, you need to know what both success and failure look like. Understanding the domain helps set your expectations right on what value the effort can produce, what it is currently making, and what is realistic.

Active Learning (and continuous labeling)

While Unsupervised Learning has been showing impressive results academically, its large-scale adoption is still not very widespread in industry. Most ML models in production environments across the globe are supervised models. Therefore, to get those super impressive outcomes, the data used to build the model has to be huge in size and diversity and labeled/verified manually by dedicated human teams. 

In 99% of cases, this labeling does not stop even after your models are deployed. It would be best if you continued getting data labeled and tracking the model performance even after a model you have built has gone into production to ensure that the model is up to date and has not “drifted” with a change in the data ground reality. Remember, your model can only be as good as your (labeled) data.

What is Active Learning = Image Credit: Settles, 2009

Learn to Manage Failure

It is a truth universally acknowledged that things don’t always work out right on the first try. In real life, you seldom get useful data, conditions, or scenarios. But you want your AI products to be able to handle those messy real-life corner-case situations with low battery and lousy Internet connections, too, on conditions that may have been overlooked while testing. 

There have been numerous instances, even by tech giants like Google and Facebook, where technical issues like a lack of representative dataset used for training or not thinking of exceptional corner cases resulted in their efforts being jeopardized multiple times, sometimes even during product launches. Even if you make sure that your data is not skewed, has all the features it can get in real life, is not unknowingly biased, has tested the upload/download part is working tens of times, it may be wise to keep some additional buffer for failing a few times. 

According to research by Gartner, only 53% of all successful AI Prototypes ever make it into production.

Startups are better positioned to value data assets and successfully utilize them by taking the right amount of risk. Still, they should also not forget the value of doing a worst-case analysis, testing out models in different situations, and starting with small tests to gain confidence before launching big. Techniques like dogfooding have gained prevalence in startups to ensure their quality is on the par with the expectations.

Embrace Constraints

While the amount of interest generated by AI and the number of ML models created is increasing rapidly, the number of people successfully turning them into finely tuned AI products that can successfully give fair business value is not keeping up. This effort requires business teams, product teams, data scientists, analysts, data engineers, and software developers to work hard, understand each other’s pain points, and build a smart, competitive solution that satisfies all customers’ and stakeholders’ constraints. 

Startups can rapidly incorporate feedback from the real world, be flexible, and improve their existing technology or strategy to ensure their products/services add more value and create innovations like never before. 

However, giving more time, money, and resources to mature may not necessarily lead to better outcomes. Recent research suggests that moderate input constraints and small output constraints on teams produced the most innovative solutions to problems, as long as the team’s morale was high.

Startups have a responsibility to their investors, their customers, and authorities, and they should take delivering on their promises as a challenge to be solved. The most successful AI startups are those where stakeholders agree on what is good enough, how long is long enough, and enforce reasonable consequences at every step.

LEAVE A COMMENT

We look forward to starting a meaningful conversation!