In the last blog, I talked about the definition of Sales Enablement and Sales Operations. In this blog, I will go deeper into some of the examples of applying machine learning in Sales Enablement.
Machine Learning in Sales Enablement
Let’s start with sales enablement. Every sale starts with a list of leads. Some of them become qualified leads, of which, a fraction converts to customers in the end. Often companies buy leads from list vendors, which in return may buy leads from other providers and compile them together. There are quite a lot of issues with this approach:
- Anyone can buy the same lead lists as anyone else.
- Quality is not always consistent. For example, one of our clients spent between $120K-$150K yearly buying leads from a vendor but was getting a high bounce rate on their email campaigns. Plus, more than 30% of the leads did not have phone numbers or titles on file.
- You may be in breach of privacy laws such as CAN-SPAM, PIEDPA, and GDPR. Imagine, if your prospects request that you reveal the exact location and time where you acquired their contact information or whether the origin of your data or of your data vendor’s data acquisition was legal. Could you answer or confirm?
- Do you know more about these leads than just email, phone, (partial) names? Without detailed data, it’s hard for your marketing and sales team to focus their energy on prospects that could otherwise be converted easily.
- Cost, cost, cost. Most of these lead list providers sell on an annual subscription basis, and some of them even lock you in for a prolonged period of time. Even worse, you may not legally be allowed to keep the data on your servers after your subscription ends.
The first solution we devised for our clients is using trained Named Entity Recognition (NER) programs to gather sales leads from prospects’ websites. In simpler terms, we have bots that are trained to recognize names, titles, contact information and other subjects from prospects’ web pages. In our most recent delivery to a client, our bots extracted 700,000 phone numbers, 400,000 titles and 1.2 million leads from more than 541,000 websites. The bots also extracted information such as the developer of the websites, the plug-ins of the websites, etc. Our bots then used search engines, industry association websites and other channels to validate the contact information. Next, we use developed custom identity resolution models combined with Customer Data Platform (CDP) to link these leads with external data such as Twitter and LinkedIn accounts. Our bots then went ahead and analyzed the content to extract key insights into what these leads have been talking about on social media.
So what did we deliver? We crafted a solution to meet the client’s unique requirements and produce qualified leads no other organizations can purchase. It also can be run over and over again on the client’s premises and maintained by the clients’ engineering team with detailed documentation and source code. We delivered a solution that ensures the client is in compliance with current privacy regulations, with the option to filter data so that it can easily comply with future regulations. We also greatly enhanced the quality of the leads by reducing bounced calls by more than 80%.
I briefly mentioned that we leverage data from multiple sources (such as search engines, geocoding, company websites, review websites) to verify contact information. That’s only one method we use. We also use clustering to assign numeric values to signal the validity of contact information. Our machine learning models divide the leads into multiple clusters and then calculate the “distance” between the contact info of a given lead to the cluster that contains the most valid contact information. These numeric values help the sales teams focus on the right contact and reduce the bounce rate of campaigns.
Delivering leads with verified contact information is only one facet of sales enablement. It is highly inefficient for sales teams to target hundreds of thousands of leads: they need to prioritize leads to achieve the maximal conversion rate. Historically, sales teams would use preset formulas that take into account firm size, deal size, etc to rank prospects. This grossly overlooked the unique characteristics of individual sales teams. No two sales teams are identical, and even the same sales team evolves over time. The clients are also drastically different from team to team. Just because one deal is worth more doesn’t mean it is more likely to convert. Let’s use the basic stats concept of Expected Value to calculate the return of two deals:
- Deal A is worth $500K with 5% probability to convert
- Deal B is worth $50K with 80% probability to convert
So the expected return of Deal A is $500K x 5%=$25K while Deal B has an expected return of $40K.
Ok, so why don’t you just use expected values to sort deals?
Because assigning probability is very subjective. Unless you have a large set of transactional records, it is statistically insignificant to calculate probabilities with small sample sizes.
Our expertise is deep learning (a type of machine learning), which excels at taking into account a large set of features (imagine an Excel table with hundreds or hundreds of thousands of columns) before making accurate predictions. In our client’s case, our deep learning model analyzed data such as:
- the sentiment of the communication between the sales reps and the prospects
- the current fiscal status of the targeted firm
- the social media activities of the prospects
- the number and types of job openings at the targeted firms
Evidently, this is way too complex for Excel to handle. We leverage the power of tools such as Pandas, Tensorflow/Keras to devise solutions that can easily make the prediction. We tried several methods to make predictions. First, we used a binary classifier to classify a deal to either “convert” or “not convert” and see the probability that’s close to the cut-off line. The probabilities then become our basis to rank deals. We also tried using clustering algorithms to predict a given deal into one of the several clusters. We found that different methods achieve different results when the data changed. For example, when we are trying to prioritize upsell/renewal leads, a classifier works the best; while for new leads, clustering has worked the best in our client’s system.
This is another reason why we are keen on building solutions rather than products. As you can see, when the data changes, the tools/models need to change as well. As a result, there is no one-size-fits-all solution. Your data (structure, quantity, and integrity) is unique to your organization and, as a result, your prediction models should also be adjusted to reflect the unique nature.
Great, now you know which leads to you ought to spend more energy on.
When is the last time you called to buy insurance or open a bank account? What questions did they ask you? Did the agents spill a string of questions like a robot? Now let’s translate the same experience to bigger sized deals. How would your buyers feel when your sales team opens up the conversation about topics related to your buyers’ interests like fishing or cycling? The key is the personalized experience, which highlights that your sales team has done the homework to understand your buyers; it signals to the buyers that the sales professionals have gone the extra mile to understand their needs.
At the same time, your sales team may not have the time and resources to perform such extensive research of individual customers while making hundreds of calls per day.
Given the objective and constraints in mind, we first developed a custom User Profile Enrichment (UPE) solution. UPE, as its name suggests, provides additional data enrichment of user profiles that are otherwise absent from the existing data set. For example, you may only know a name or an email of an individual. UPE’s objective is to add more information such as “what are this person’s interests?”, “which college did this individual go to?” etc. The process is relatively straight forward:
- We first acquire additional links to public social profiles such as Twitter by using email/phone number reverse lookup.
- We analyze the text on these public profiles to extract key insights that are useful to build a personalized sales playbook for sales reps.
Part (1) of our solution for that particular client leverages our own bots as well as a well-known CDP to acquire additional insights. CDPs typically buy data from a variety of vendors and stitch the data together to form larger pictures of individual profiles. Some well-known CDPs include FullContact, Clearbit, LiveRamp etc. They all provide different levels of details. For example, LiveRamp only provides batch-level insights, such as demographics while FullContact provides links to social media profile pages. It’s vital for you to plan your needs and costs before making a decision on buying one or several of the services.
As we briefly touched upon in (2), we leverage Natural Language Processing (NLP) (a type of machine learning that focuses on analyzing human languages) technologies to extract key insights of individuals:
NLP by itself is a very complex topic to write about. Sometimes, I write technical blogs on it. You can also find an abundance of references on the internet to learn more. The end result of using NLP is to compress large bodies of texts (corpora) into small snippets that sales teams can easily digest so that they don’t have to spend a large amount of time doing it themselves.
Here is an example of a prospect profile before enrichment:
Here is an example of a profile of the same prospect after the enrichment
- Profile Picture
- Schools graduated
- Topics of latest tweets
- Topics often talked about on social media
- Estimated revenue of the Firm
- Recent rounds of funding raising
- Recent news of the Firm
- Number of employees at the firm
Armed with rich insight like the above, sales reps are much better positioned to have a personalized conversation with the prospects. The profiles are also shared across multiple teams to ensure the experience is always consistent across the company.
Stay tuned for Part 3