Despite the recent successes in the use of machine learning (ML) to perform tasks more accurately than humans such as in cancer detection and hand-writing recognition, moving beyond a demo to roll out a live commercial product requires much more than a fancy algorithm. There’s no doubt that the technology will re-define most industries, but it’s worth keeping in mind that we are just at the beginning of a multi-decade cycle and so entrepreneurs should be cognisant of this when implementing ML in a commercial environment with clients. What follows are some lessons learnt from being in the trenches.

It’s worth clarifying that most problems in ML follow a similar pattern, loosely called “predictive analytics”:

Building these models requires huge datasets of labeled historical records. For example: loan applications with tags stating whether a loan repayment event occurred in order to make a prediction whether an applicant will repay a loan.
The dataset requirement may not sound problematic (even assuming the requisite datasets are available), but the reality is that these datasets are often residing in different areas of the business, in different formats with different labels and with different decision-makers for each dataset. Moreover, depending on the industry, there may be regulatory constraints such as data that contains personally identifiable information, or perhaps the data needs to be kept within a certain geographic location. Another critical concern is how often are these datasets updated and how quickly can one gain access to them in order to re-calibrate and improve your model? In fact, 80% of a data scientist’s time is spent on preparing the datasets in order to build the predictive models.
My advice to entrepreneurs when partnering with companies is to focus on the following:

• Be realistic about the length and complexity of the sales process – it’s certainly more of a consultative sell than picking items off a menu. Find out early on who makes the decisions (both technical and commercial), what data is available, and whether the company has a gap in their engineering / implementation roadmap to accommodate you.

• Having said that, if you’re a product company it’s important to know your boundaries and not to develop a bespoke product that won’t scale with other companies.

• Emphasise that, although further down the road you may require the very large datasets, to begin with you only need a sample set. Try narrow the problem and stick to simple models in order to iterate quickly.

• Set expectations early on. Companies will not give you access to valuable datasets if you promise a “black box solution that magically transforms the data into a new multi-billion dollar revenue stream”. Help them understand what type of problems can be solved using ML and what it cannot do. Transparency is key.

• Although ML and AI are buzzwords, they are just a means to an end. ML startup teams tend to be deeply technical and have a tendency to focus too much on the technology and not the business value provided. One of the advantages of providing a ML solution is you have access to valuable datasets. This means that one is often in a position to actually quantify the value-added to the customers’ business – be it a new revenue stream or quantifying the additional profit generated as a result of a more accurate solution.

• Often companies already have a data team. In most cases, it’s not a data science team that’s working on the same set of problems as your solution but it’s important to establish whether your solution is complementary or competitive to this team’s roadmap, and to get their buy-in if it’s the former. Often this team serves as a great source of information on the inner-workings of the business processes and data flows within the company.

• Understand the relationship between the accuracy of your solution and the product risk you pose. If the solution does not yet exist and your product is better than a random guess, then you could already be adding value; however, if you’re providing self-driving car software, the cost of an error is extremely high.

• It’s important to get the company to pay for a proof-of-concept. It serves as an important validation that your solution can provide enough business value that it’s worth risking money on a test with uncertain results.

Naturally this process will evolve quickly as the technology matures and becomes better understood. Right now, though, it’s important to understand that, as with most technology waves at the beginning of their cycle, the technology is poorly understood and it’s your job as the Founder/Sales Lead of a ML company to bridge this gap.

About the author

 Rael Cline,  Co-Founder & CEO, MediaGamma

Holds a Masters in Finance from Cambridge University, & has experience in sales, marketing, commercial & managerial skills within both finance & technology organisations, co-founding AppChat, a mobile virtual network operator in South Africa. Previous to that Rael worked in private equity, at one of South Africa’s largest funds.