Tech enablement done right

A few introductory words

MyJobNow has been around for less than a year. But in that time we have accomplished a lot. I’ve been working with/around startups for 5 years now and in that time, I have often come to make a broad distinction between tech startups and tech enabled businesses. In the former category one would find companies that build up tech IP and in the latter any business that leverage new technologies to offer better or (and) cheaper products for established market needs. Now I know that this is not a very precise set of definitions but it always seemed to me that people that I talked to got it and concurred with this high level model. However, I really came to internalise this notion of tech enablement with MyJobNow. In retrospect, although we have deeply product focused trajectory, I came to the realisation that >70% of our resources have been spent on sales & marketing operations (and rightfully so). MyJobNow is a marketplace and our product/service quality vastly depends on our users. This is something I believe a lot of startups get wrong. Building complicated and technologically challenging features only works if your users actually need them. And understanding what users need is not a trivial task. Therefore we have invested a great amount in making MyJobNow operationally efficient so as to receive, understand and productise user feedback rapidly. We have accomplished that by doing our tech enablement right.

Lead prioritisation — A great example

A bit of context

I would like to share our latest accomplishment in terms of leveraging technologies, as a good example of how we have used state of the art practices and technologies in our stack to optimise sales calls prioritisation for new customer conversions. I am going to go through the technologies/tools and how we have used them to solve a specific problem.

Our leads are business owners or HR personnel that wish to hire people. We provide a trial period for our product where leads can post 1 job opening in our platform, receive applications from candidates and communicate with 1 candidate (they have to pay for more). Our average purchase size at the moment is ~€40 which makes it very challenging to deploy a field sales team. Therefore we have sales reps that call leads (whilst on their trial period) to convert them. At the beginning a blitzkrieg approach was pertinent because our lead generation processes produced just enough leads so that a couple of people were able to handle them. In the past couple of months we started getting into a phase where inbound leads are way too many (and not an insignificant portion of them irrelevant) for our sales team to handle them all. So we started prioritising by using rule-based systems like calling leads that are further down the pipeline in priority to others. The sales team started taking initiatives which seem to work very well by assessing each lead’s usage and profile before making the call. This yielded great results right away increasing our conversion rates +30% on the attempted conversions. However, things slowed down in terms of call capacity since every single lead had to be manually evaluated before attempting to call. So the goal became to have some sort of automatic prioritisation so as to reduce the probability of assessing a bad lead as low as possible.

Our Stack

Here is where we start talking about our tools and how we have used them to create our sales process. So one by one:

1. Pipedrive CRM: Pipedrive is a great and extremely customisable CRM. We have done a lot with it and it would take ages to list all of our customisations. The most important thing to know however is that we have used their API to create a sales funnel using our production data. A deal/person/organisation is automatically created when a new user is on-boarded and his/her stage in the pipeline depends on his/her actions in the product.
2. Intercom: for messaging & ticketing. Intercom is a great tool, and we have invested a lot into using their SDK in our apps to catch all the events users go through, which gives us massive space for optimised and customised messaging.
3. Blendo: My favourite tool of all, blendo makes ETL extremely easy. All of our CRM, Intercom, Adwords, Facebook ads and many more data are seamlessly fed into a database that we use for analytics and automations.
4. Python Pandas & sklearn: Python and sklearn allows us to build simple (and complicated) data driven analysis and machine learning models in a couple lines of code.

Lead prioritisation

The goal of this task was to find a way to prioritise leads by probability of conversion. We tried to see if we can predict whether a lead is likely to convert according to:

1. The Customer profile (Location, Industry, etc..) — Pipedrive & production DB data
2. Effort invested by Lead (quality of job posting) — Production DB & intercom Data
3. The value received during his/her Trial — Production DB Data
4. Timing data — Pipedrive Data

The model used for such a classification is not as important as obtaining the feature-set to train the model. A schematic below outlines how our stack is being used to gather the feature-set and productise the model trained.

Sales reps can view and input data via the CRM and Blendo does a great job in gathering everything in one location. Using the database populated by Blendo we were able in less than a couple of days to play around with the data and find a pretty decent model to use.

A Random Forest classifier was the best choice for ~7,000 won/lost deals used as the training set. I was actually quite surprised with how well this worked. It produced an accuracy of ~76% and f1 score of ~68%. Recall on predicting conversions was quite low but by adjusting the threshold I was able to improve this significantly. What was very impressive is the relationship in terms of predicted probability against conversion rates as shown in the figure below.

As as result we used the predicted probability as a scoring function to prioritise leads.

To productise this model we created a service that rank-orders all open deals every hour according to the trained model’s predicted probability and classifying them:

• A: the top 20% of the population (in terms of scores)
• B: 20–50% of the population
• C: bottom 50% of the population

Given that our conversion rate is ~25% it seemed logical to use these percentages . What we end up with is with a simple A,B,C classification for all open deals. Sales reps are to prioritise accordingly.

Did it work?

F***k yes! Although a bit preliminary here are the conversion rates on closed deals for the past 15 days:

Lead Class/Conversion Rate

A : 42%

B : 21%

C : 7%

Moreover our conversions have increased by ~15% in the last couple of weeks :)

Conclusion

We’ve done quite a lot of cool stuff in the past year and maybe I’ll write about some of them in the future. In any case, I am pretty proud of the work we have done at MyJobNow so far and wanted to share some of the awesomeness, Tech enablement done the right way.