Using OpenAI Assistants with Fireflies transcripts to Update Company One-Pagers
At Valor we look at more than 20 new startups every week. Because we lead seed rounds for B2B SaaS companies, when we first meet them, a few months can make a huge difference in their suitability for our thesis. To facilitate keeping track of so many companies and their updates with a small team we created Vic, our AI Assistant.
Add to Salesforce and summarize and enrich data
Our CRM is Salesforce, so the first step when we learn about a startup is to add them to Salesforce. These days that process itself is most often automated — typically by sending Vic an email. A company name or (ideally) website URL should be enough to get going. A salesforce trigger will call our research Assistant. This Assistant gathers information by performing a web scrape, conducting a Google search, and adding key findings to Salesforce. The goal at this stage is to fill in the foundational ‘building blocks’ of the company one-pager, which include:
- Description
- Product Story
- Revenue Story
- Funding Story
- Team Story
To achieve this, I run an Open Assistant that I’ve equipped with specific functions: webscrape() to read company website information, targeting pages like ‘team’ and ‘product’; googlesearch() to find LinkedIn profiles and other key pages; and pitchbook() to look up funding information. Once this data is gathered, the addcompany() function uses the Salesforce REST API to integrate the findings and create records automatically. Depending on the available information some of the information might be missing (often funding info and revenue info)
Assessment Assistants
Once the initial record is ready, it’s time for evaluation. Here, multiple Assessment Assistants come into play:
- Thesis Evaluation: One Assistant evaluates the company against Valor’s investment thesis. We’re giving it all the data we have and ask it to check against our criteria, that include location (the south), stage (Seed) and type of company (B2B SaaS).
- AI Rating: Another Assistant generates a Valor AI rating. We use a very simple scale from 0 to 3 that helps create our quarterly AI in the South reports.
- Revenue and Funding Check: A third Assistant helps with more analytics with categorizing revenue and funding information. This makes our reporting easier (example -1: no revenue data, 0 = no revenue 1=<$100k ARR etc)
Every time we update the company basis information, including the building blocks mentioned before, all assessments are re-run as well.
Learning more — meeting and company update Assistants
Internally we use Fireflies.ai to record and transcribe all meetings. Using the Fireflies API, I import all those meetings daily. For each new meeting an OpenAI Assistant is used to identify the account the meeting belongs to and the type of meeting. Then a second Assistant is called to create a detailed summary. For all our meetings with Startups the summarizer prompt is to focus on the 5 ‘buckets’ mentioned earlier (Description, Product Story, Revenue Story, Funding Story, and Team Story). Other meetings will have slightly different summary instructions, like customer reference calls.
When we receive new startup information by email (Investor updates, maybe a new pitch deck) we follow the same procedure: first extract all data from email/decks/attachments using LlamaParse and summarize the information received across the same buckets.
Update the one pager Assistant
Finally, we call an Assistant that will take the bucketized meeting summary and use it to update the original company information. The prompt is very specific to replace older info with newer info. For example:
Current info in Salesforce: Acme.ai — Revenue Story: The company is currently pre revenue but expecting to sign its first paid customer in April.
Meeting summary: …. Revenue Story: Acme.ai signed their first paid customer contracts one with ABC Inc for $30k ARR in May and other one with DEF LLC for $90k ARR in June…
The company update Assistant will update the Revenue info for acme.ai to something like:
Updated info in Salesforce: Revenue Story: The company currently has $120k in ARR from two paid customers.
It will do the same for each bucket, and it will do this every time a new meeting or company update is received.
If any of the buckets are updated, the Assessment Assistants are also called to make sure those reflect the current state of the company as well. In the case above for example, the Revenue Status would be changed from ‘0’ (No revenue) to ‘2’ ($100k-$500k ARR)