Project Overview
ListenBravo was building an AI-powered tool to help their clients' revenue and growth teams in discovering new accounts, enriching existing ones, and finding expansion opportunities. With us, they requested the addition of agentic capabilities to power their MVP, which already had a few clients.
This was a relatively straightforward application of AI agents to the problem of optimization of RevOps, but nevertheless goes to show how much automation can help businesses optimize their workflows. You can see a demo of their product here.
Problem Statement
Given a database of target customers, enrich it on-demand based on the client's prompt and answer schema.
For example, given a list of target company data like name, industry, yearly revenue, number of employees, description of operations etc, the client would add a column called "Has Chatbot" and provide a prompt saying "Visit this company's website to find out if they have a chatbot operating."
The initial database itself can be either:
- Provided by client
- Asked to be provided by scraping data from web sources based on a signal of the client's choosing (for example: scrape from LinkedIn, Glassdoor etc. for Hiring Signal data)
Solution: Agents with proper tools
We set up a simple Langchain agent based on gemini-flash-1.5 and provided it with prebuilt tools like google-search and PlayWrightBrowserToolkit and custom tools like PyPDFLoader and SeleniumURLLoader. This would enable the LLM to do things like navigating to a URL, browsing and loading the customer website, searching for news, loading annual reports etc. This minimal agent setup is shown below:
Then we set up a system prompt like this:
Consider the details of a target customer given below.
{<insert row data from customer database>}
Now answer the following question requested by the client.
Question:
{<client prompt>}
The answer should be a JSON object as shown below:
{{
"answer": <value>,
"reason": <reason-text>
}}
containing "answer" and "reason" keys.
<value> needs to be in a {<client schema>} format.
<reason-text> needs to be a single sentence explaining why you chose the answer.
Do not make up an answer if you are not sure! Do not use Markdown!
You have numerous web search, browsing and pdf tools at your disposal, so use them as necessary!This minimal setup worked surprisingly well for the MVP, with the agent routinely finding correct information on the web, finding and parsing PDFs correctly and reporting it back in the requested format. There were, of course, the usual failure modes of the agent getting stuck in a loop, context-window limits and failure to spot "needles in the haystack". In a more rigorous setting, these can be handled with appropriate guardrails, timeouts, context-engineering and narrower scoping of tasks.
Conclusion
The entire end-to-end solution delivery from problem assessment to demo took less than 10 days, owing to the straightforward applicability of AI agents for ListenBravo's use case. They reported great success with their MVP, claiming remarkable outcomes for their clients in a very short span of time:
- Reduced time for MQL to SQL qualification by 90%
- On an average increased sales pipeline by $10,000/ quarter
- Reduced average SDR required per campaign from 10 to 3
- Helped increase number of campaigns per quarter by 20%