Problem

The sales team was spending a significant portion of their time on business development and outbound sales. An external SDR team cost significant money with virtually no results to show for it.

Solution

Sales is a domain where process automation has been used for a long time with varying results. We’ve seen the rise of literally thousands of software companies attempting to sell their “AI sales automation” tool. However, when you scratch the surface you realize most of these tools don’t really make much of an impact at all.

There is a number of reasons why many existing AI-sales approaches produce unsatisfactory results

• Lack of deep integration with your CRM and actual ways of working
• The people who configure the system don’t really understand your company’s value proposition and customers buying behavior
• The AI-companies are skimping on the LLM model power to keep costs down
• Over reliance on AI to do everything diminishes internal knowledge and critical thinking

The result is that most AI-sales tools create generic, boring and poorly calibrated results. As more and more companies adopt these approaches its also an increasing arms race where recipients of prospecting messages will become less receptive.

AI is a fantastic tool, but you must know how to wield this power.

To navigate around these issues and to help the sales team focus more on actual selling we created a bespoke virtual assistant that was assisting the sales in prospecting. We decided to build something more similar to a team mate – a bot that you can give tasks by describing what you want in plain language.

The virtual assistant had custom built capabilities for:

• Autonomous planning: The most important capability we gave the assistant is to decide on its own what it needs to do based on the task description. Tasks were written in plan language, which then get broken down to a sequence of activities.
• Research: The assistant has the ability to execute web searches, visit company websites and discover LinkedIn profile data. It summarizes all its findings to a temporary memory that can be used to steer other parts of the assistants work. The research agent was trained to look for information that was specific to the value proposition of the company.
• Writing: Because the assistant can read all the data from CRM as well as conduct research from various sources, it was capable of writing extremely natural and personalized messaging.
• Update CRM fields: If the bot found new information that was not aligned with the CRM it had the ability to update some chosen fields, like web address or number of employees, etc.
• Find profiles: The bot was able to discover decision makers based on job profiles. We found the best results by focusing on job titles and profile keywords.
• Quality score leads: Thanks to its research capabilities, we could easily build a module for evaluating the quality of leads. The bot was trained on the complex rules between country, company size and vertical specific dependancies. With these approaches the bot was able to guess a leads quality score with basically perfect certainty.

From the get go we felt it was important to keep sales people in charge. Every message created by the system was approved by an AE. The idea was that machines don’t replace sales people – they help them work faster and with higher quality.

What was the outcome of this approach? The sales reps were able to do prospecting consistently and with high quality at a rate exceeding 10 times the regular pace. This meant more time could be spent actually selling. After about 6 months in operation the system could be seen as having contributed to 40% of the enterprise pipeline.

Automation pays off as the system not only resulted in pipeline growth but also cost 80% less to operate than the earlier approach. On a day-to-day basis this system is also virtually maintenance free as we put in rigorous error handling and made sure the system continues to operate even if some parts of the automation fail.

Customer

Enterprise SaaS

Systems Used

• Custom APIs
• Web scraping
• Process automation
• Pipedrive
• GPT-4 API
• GPT-4o API