By Robin Bordoli, June 7, 2016

Solving Million (not Billion) Dollar Business Problems with AI

Today we announced our recent $10M financing led by Canvas Ventures, Trinity Ventures, and Microsoft to fuel the adoption of our new CrowdFlower AI solution.  This post explores the “why” and the “what” behind this financing.

So first, why did we raise this capital?

It’s a simple answer.  We wanted to bring the economics of applying AI and Machine Learning within the reach of every business.

Why did we think this was the right goal at the right time?

It was based on observing two groups – first our own data science customers, second everyone else trying to deploy Machine Learning.

For the past few years we’ve had a front row seat seeing the emergence of the data scientist role inside companies large and small.  According to LinkedIn there are now over 75,000 self declared data scientists in North America.  We’ve been helping these data science teams convert their unstructured data (text, images, video, audio) into rich structured data on our human-in-the-loop platform.  We’ve done this at scale with over 2 billion judgments applied on our platform.  We’ve seen the evolution of companies go from data engineering to data collection to data science, and now Machine Learning.  Increasingly data science teams are using our platform to create the training data they need to feed their Machine Learning algorithms. Given our front row seat it made sense for us to make Machine Learning a commercially viable proposition for our customers.

In addition to our customers, we observed how Machine Learning was being deployed by everyone else.  We came to two major conclusions.

First, because of the high startup costs of applying Machine Learning (hiring a skilled expensive data science team with expertise in Machine Learning, building your data infrastructure, iterating and iterating and iterating, etc) it has been predominantly targeting billion dollar problems such as driverless cars.  We saw an underserved part of the market.  Namely the millions of million dollar business problems such as support ticket categorization.  These problems aren’t worth a billion dollars to solve because the way in which each company wants to solve these problems is custom to them.  Uber wants to categorize support tickets differently from Tesla which is differently from Fitbit.  One size does not fit all.

Second, even though the major technology companies such as Microsoft, Google, and IBM are now offering Machine Learning as a cloud offering to lower the high startup costs, there still remain two major barriers to adoption.  The first barrier to adoption is lack of training data.  A model without training data is like a car without gas – it doesn’t go anywhere.  The second barrier to adoption is an inability to blend humans and software rather than replace humans 100%. A state of the art Machine Learning model might be 80% accurate but no sane business executive will consider replacing humans with such a model.  Faced with the option of either replacing the humans 0% or 100% the the business decision will always be to not replace the humans.  As a result Machine Learning remains a science experiment for the vast majority of businesses.

So we raised the $10M capital to fund CrowdFlower AI to address the underserved market of millions of million dollar problems such as custom support ticket classification that could be solved by AI.

Now let’s discuss the “what” of CrowdFlower AI.

CrowdFlower AI is a solution that combines training data, Machine Learning models and human-in-the-loop capabilities in a single platform.  So what?  Why does having these three capabilities in a single platform matter?

It matters because we’ve dramatically reduced both the time and cost to deploy commercially viable Machine Learning models that augment humans in some part of a business process.

Let’s break this down into the 5 steps of using CrowdFlower AI

  1. Data.  Real people apply their human intelligence on the CrowdFlower platform to create the specific quality controlled human-labeled training data you need for your business. For example, you could upload your support tickets and create the custom classifications that matter to your business.
  2. Model.  With a few clicks, a custom Machine Learning model with predictive power is created from your human-labeled training data.  For example, the model will show how accurate its predictions are compared to the support ticket classifications in the human-labeled training data set.
  3. Predict.  When you upload new data to CrowdFlower AI, it will make predictions based on the model you’ve created.  For example, you could upload new support tickets and the model will predict the relevant custom classification.
  4. Route.  Since CrowdFlower AI gives a confidence level for each individual prediction, you can route confident predictions to the model and less confident ones back to our human contributors. For example, you could define the required confidence level as 95% so only the individual predictions above 95% are handled by the model.  This human-in-the-loop capability means you no longer have to choose between replacing 0% or 100% of the human activity.
  5. Improve. Once those less-confident predictions are over-ridden by human judgments, they get fed back into your Machine Learning model. That means your models get smarter on the exact use cases they need more training on. This results in the best combination of speed, cost and accuracy for your business.

Despite the hype and scaremongering surrounding AI, we believe it’s the right time for every business to ask themselves the question “What is our AI strategy to improve how we engage with our customers?”.  If you want to cut through the hype and find a practical way to start applying AI to your business, we suggest you start with the repository of unstructured text inside your Salesforce Service Cloud or Zendesk support ticket system.

If you’re ready to learn more about how AI can apply to your business we’d love to talk to you.  You can find out more at