Case Study: Automated Proposal Generation Tailored for Each Client
Introduction
In today's fast business world, responding quickly and effectively to clients is crucial. Many companies spend a lot of time creating personalized business proposals for each potential client. These proposals must be carefully tailored to address the specific needs, concerns, and requirements discussed during client meetings. This process often requires hours of detailed analysis, writing, and customization.
Hypothesis
We propose that using ChatGPT to automate client proposal creation will decrease time spent and improve customization and relevance. This should result in higher client satisfaction and more conversions, as proposals will better address specific client needs and concerns.
Solution Implementation
Our solution involves a multistep process utilizing ChatGPT, integrated with a no-code automation tool N8N and Notion, to create detailed, client-specific proposals based on transcribed meeting notes.
N8N is a powerful no-code workflow automation tool that allows users to integrate and automate various online services and APIs without writing any code. While it is not specifically designed for AI applications, N8N recently got very powerful update that introduced various AI-nodes that make development of AI applications very easy and available for everyone. The idea of this use case was not just automate the proposal creation but also battletest N8N.
Step 1: Template Preparation
First, we designed a comprehensive proposal template that includes sections such as executive summary, problem statement, proposed solution, project team composition, relevant case studies and FAQ. This template serves as the framework into which specific client data is inputted.
Step 2: Meeting notes creation
After each client meeting, the conversation is transcribed and saved in Notion. Moreover, this transcribed data undergoes a comprehensive analysis and structuring process where we extract information about the client's business, pain points, expectations of our product, and concerns. This document serves as the primary source of data for the proposal generation process, containing key details about the client's needs and expectations
Step 3: Creating Client’s Persona and Company Profiles
To create detailed client personas and company profiles, we gather publicly available data from LinkedIn and other social media, as well as from the client's website. This raw information is stored in a dedicated Notion page. A specialized n8n script is then run to process this information, creating structured Persona and Company profiles that follow best product management practices.
Step 4: Incremental Proposal Creation and Refinement
Having all the necessary components together, such as the meeting notes, client's persona, and company profiles, we can begin to build our proposal, section by section, using our template. We supplement some sections with additional data from our knowledge bases. For instance, the 'Proposed Solution' section is built using information about our products and services, and the 'Case Studies' section is formed using a database of use cases, etc. This information from the databases is analyzed by LLM (Language Model), filtered, and adapted for the client's needs and expectations.
The incremental proposal creation process provides more control over the content generated by the LLM. Each subsequent section is built using the context of the previous sections. For example, the 'Proposed Solution' is based on the 'Problem Statement' section, and the 'Project Team' is based on the 'Proposed Solution', etc. Building the proposal incrementally, section by section, allows the editor to review and edit the information provided by the LLM, which, in turn, improves the quality of all subsequent sections.
A typical n8n pipeline for writing a proposal section looks as follows:
- Subscribing to Notion database update events and filtering them by a certain action
- Downloading all general information required for every section (Meeting notes, persona and company profiles)
- Downloading section-specific information (e.g., list of services we offer)
- Building a prompt combining all information together
- Making a request to LLM to write a section content
- Making a request to LLM to format the answer from Markdown to be uploaded to Notion
- Uploading the content by Notion blocks.
Step 5: Persona Simulation for Quality Assurance and Personalization
In this final step, we employ LLM in a novel way to further improve the quality and personalization of the proposal. LLM is asked to take on the role of the client's persona, based on the details we have gathered and analyzed. Acting as the client, LLM reviews the draft of the proposal and raises any questions, concerns, or areas of potential confusion that were not addressed in the previous sections of the proposal or during the initial client meeting.
Then, in response to these simulated client inquiries, we generate another round of LLM responses. This time, LLM is asked to provide clear, concise answers to these questions, backed with supporting information where necessary. This process allows us to address potential client concerns proactively, further personalizing the proposal and demonstrating our understanding of the client's needs and expectations.
Evaluation of Results
The automated proposal system has cut down the average proposal creation time from three hours to just thirty minutes. Manual intervention is still needed for gathering client and company data, content review and editing, and markdown and format adjustments.
A seasoned sales expert, boasting over 15 years of IT experience, confirmed the high-grade quality of the AI-generated content.
Limitations and Constraints
Originally, it was a single, continuous pipeline to generate proposals from end to end. However, this was not practical due to potential hallucinations that could occur in the middle of the chain, affecting the rest of the outputs and degrading overall quality. Additionally, the N8N tool began to freeze and crash with long chains during editing. As a result, it was decided to divide the chain into steps to generate sections separately.
It was also discovered that Notion has limitations in its API. Notion does not support automatic recognition of markdown formatting and has a block length limitation of 2000 symbols. Therefore, an additional step was required to convert markdown text into an array of blocks and save each block individually. Reading information from Notion also requires additional filtering and concatenation of blocks in the final output.
Conclusion
In conclusion, the use of ChatGPT, integrated with N8N and Notion, has proved to be a successful innovation for generating automated and tailored proposals for clients. The solution has significantly reduced the time required for proposal creation while preserving quality and customization. Despite some limitations with Notion's API and the N8N tool, the overall effectiveness of the solution has been confirmed. This project demonstrates the remarkable potential of AI and automation in enhancing business processes.
Next Steps
Moving forward, we will focus on the following areas for further development:
Integration with UI Services: We aim to integrate our system with services that can provide a visually appealing interface for our proposals. This will allow us to present our proposals in a more professional and aesthetically pleasing format, whether that be HTML or PDF.
Price Calculation Module: Currently, the price and quantity elements are text-based. To enhance our service, we plan to develop a price calculation module. This module will automatically calculate costs based on the client's specific needs and requirements, providing a more accurate and detailed pricing structure in our proposals.
Validation Through External Business Implementation: Currently, we've tested this system on our own processes. However, we don't have a large customer base or a history of sent proposals to evaluate how the implementation of AI affects conversion rates. Hence, our next goal is to find a business where we can implement and validate this system.