Introduction

Are scale-ups perfect machine learning breeding grounds?

The silly answer is off course, it depends on what you do. AI is not the solution for everything and not (or not yet) the right investment for every scale-up. But generally speaking, there are a number of factors that scale-ups have in common that make them machine learning breeding grounds. Meaning that ML projects have a decent chance of succes. Where I like to think of “succes” as a case that is lasting and has positive ROI. Based on my experiences within the scale-up field (mainly NL + Berlin) I picked the strongest reasons. If you are a scale-up, you check all of the five points below and you are not investing in AI yet, you should probably start considering.

1. Digital business model

As you probably know; for machine learning use cases you need data, and preferably a lot. Since scale-ups typically have a digital business model (i.e. an online marketplace, an app-based Fintech, a SaaS provider) the majority of the interactions with the customer is digital and therefore often stored as data from the early start of the scale-up. Preferably, the internal processes are also decently logged from the beginning (see bulletpoint 5). This own created pile of data (rather than some external source) is often the best start for machine learning use cases. Only once the value from this is revealed, you should start to add external sources.

2. Strong development team

Though not every scale-up already has an AI team, high-skilled development teams (front-end and back-end) are usually inhouse. This factor should not be underestimated when moving towards true implementation of machine learning. If all good, those teams created a scalable product infrastructure with decent setup of underlying databases and data-model.  Once your ML use case is finished, those type of infrastructures (with some specific extensions for ML purposes) generally allow for fast integration with your use case. A big, big difference with the legacy systems which are for instance more the case at corporates.

3. Believe in AI from investors

Investors generally support and like new techniques like AI (maybe even a bit too much in some cases?). To convince them of a specific AI investment, create a decent business cases. Though this sounds as an obvious point, the number of companies (of all sizes) that start a data science case without having a decent estimate of the potential business value, is still shocking. See the next point for relevant cases.

4. Relevant AI (business) cases

As stated before, AI is not the solution to everything. Finding the right use case to start with, with a decent balance between impact (business value) and ease (feasibility) greatly increases the opportunity of building lasting AI. Think a long the lines that you scale to find the right use case. Some examples / lines of thinking:

• Growing number of FTE: look for repetitive, mundane processes with a lot of volume (i.e. claim-handling, creating sales offers, customer service), often bringing strong business cases.
• Growing number of customers (& customer types): now your customer base is growing it starts to make sense to get better data driven insights on who your customers are, what their acquisitions costs are and their lifetime value. Use this for segmentation and personalisation of your services.
• Growing product: first you created the easy features to make your platform run and provide the service that you want to provide. Now that the number of interactions with your product is growing, it might be time to make it smarter. Think off solutions like in-app recommendations, ranking solutions, dynamic pricing.

5. Pay attention to logging your data for AI purposes

We started are list with data, and we end with it as well. Though we mentioned already in the first bulletpoint that the business models of scale-ups by default provide a lot of data, you have to be very aware of correct logging of your data, specifically for AI purposes. In the early days of your platform, probably no one cares about future (5-10 years from now) machine learning use cases and what specific data is needed for this. The platform (and subsequently the data that is stored) is designed and developed from an MVP (let’s get it working) perspective. Common mistakes are overrules (rather than timestamping every event) or not logging at all. So for every customer interaction, every (digital) usage of your product & services and every decision made by an expert after a manual process (creating a proposal or handling a claim): Log, log, log!