A chatbot recruiter to humanize job alerts

Torre.co is a job matching network startup focused primarily on remote jobs, with over 1.5M users. It has a powerful matching algorithm that uses 121 different factors to determine if a person’s genome is a good fit for a job and vice versa.

If we can suggest full-time and internship opportunities to job seekers in a way that feels human and personal, a significant amount will be more likely to engage with the posts.

After many conversations with users about the product, we were bullish on building a secondary job alerts system that could leverage our algorithm in a way that felt a lot more personal and premium.

In essence, we wanted to build a “Senior Recruiter” email bot that would send job posts to users and automate around 90% of our one-on-one conversations with job seekers. The interactions had to be carefully crafted to simulate our real recruiters.

For context, the following screenshot shows what job alerts looked like at the time. Sufficiently straightforward for some, a bit cold for others.

In order to build a bot that could pass for a human, there were a few things to take into consideration:

From the start, I was aware that this project would require little to no UI design. This experience would happen in conversations between our email bot, job seekers, and real recruiters from Torre. However, in order to send most of the emails, we would have to set up triggers linked to several of our core interactions on the platform. Plus, we would have to create endpoints and integrations with Hubspot for recruiters to chime in when it was necessary.

The deliverable for the engineering team would be a BPM detailing all possible interactions and a document that included every email with 30+ variations for each one to better emulate a human writer.

Building the BPM took me about three weeks of full-time work and many back-and-forth conversations with the CEO. We started out with a very general idea of what we needed but making the bot believable required much more attention to detail than we first estimated.

For example: As I reviewed one of the first versions of the full journey, I realized in order to prevent spamming, the bot would have to ask the user about their frequency preferences (in a non-robotic way). Also, our algorithms manager requested a way to get some feedback about the matches so they could be improved.

In order to achieve this, I came up with a system of auto-replies to the recently sent email. Depending on many factors, after sending a job post email, the recruiter bot would send a reply to that same email with a note.

Some of those reply emails were:

Another very important addition was the “Book me for a call” line at the end of each email. Before we decided to add it, I had a strong feeling that something was missing. Even with all the versions of emails, the interaction still felt robotic. The [real] possibility of speaking face to face with María would add the final touch to humanize the interaction.

Of course, this addition wasn’t as simple as a line of text. We had to coordinate with our team of recruiters for them to keep a Calendly and take as many calls as possible. Luckily, the link was rarely used because communication by email was very fast.

In summary, since I can’t publish the BPM, here’s what the bot would do according to the flow:

This project could not be tested with a regular prototype. It required a guided one-on-one experience in which testers would go through the entire flow of a new Torre user.

For it, I set up a very basic presentation with Figma that would allow me to show different emails to the user through the video call depending on their response to the previous question.

I personally tested the prototype with 10 users through video calls, walking them through the journey and asking their thoughts after every hypothetical interaction.

I would purposely avoid mentioning the recruiter would be a bot during the test so I could ask about it at the end.

After each test was finished I would ask general questions about their opinions of the service, the help being offered, and the tone of the emails. Finally I would ask them two questions that they should answer numerically (1= not at all 10 = completely). The first one was about their overall rating of the experience and the second was to what degree they suspected they were talking to a bot.

The results were overwhelmingly positive:

The development took a couple of months. The BPM was extensive, the system touched on many different areas of the platform, and there was little room for error. As the project manager and designer, I was there whenever necessary to answer questions and guide the developer through the details.

A few weeks after launch, once we had solved the major bugs and the flow seemed to be running smoothly, we conducted a usability test with the real product.

We randomly selected 8 users who had received emails from the recruiter bot and asked them about their impressions.

Excerpts from the results:

One month after the product was fully launched, we analyzed the quantitative impact it had on the platform. Results were very promising:

A year after the product was launched it has become the main channel for applications on the platform. The bot has beat regular job alerts in open rate by a 13% difference and in applications by 16%.

Hundreds of people have been hired thanks to the recruiter bot and it has continued to grow with new functionalities. María (the bot) constantly receives praise for the attention she has given to our users. Here are some of those comments: