It started with a simple observation. I noticed that some of our most brilliant engineers at JvG Technology—the ones who can design a complex solar module production line in their sleep—were spending a surprising amount of their time on work that felt more administrative than innovative: writing proposals.
These aren’t simple sales documents. They’re dense, highly technical blueprints that can run over 50 pages, detailing everything from machine specifications to throughput calculations. Each one is a testament to our expertise, but building them from scratch is a significant time investment.
Research from Bain & Company confirms this isn’t just our challenge; it’s a universal one. They found that top-performing sales reps spend 52% more time on actual selling activities compared to their peers who get bogged down in administrative tasks. Seeing my engineers wrestling with document formatting, I knew we had a bottleneck. This became the starting point for a new experiment: Could we build a digital apprentice to help them?
Our goal wasn’t to replace our engineers—it was to augment them. We decided to test if we could train a custom Generative Pre-trained Transformer (GPT) on our own project archives to handle the heavy lifting of drafting these proposals.
The Core Problem: The High Cost of Manual Proposal Generation
Every hour an engineer spends searching through old documents for the right technical spec or formatting a table is an hour they aren’t on a call with a client, solving a unique engineering challenge, or refining a system design. This opportunity cost is immense.
The complexity lies in the bespoke nature of our work. A client in Arizona has different needs than one in Northern Europe. A proposal for a high-efficiency PERC line looks very different from one for a specialized glass-glass module line. While the core components are often similar, the configuration, context, and commercial details are always unique.
This is precisely the kind of challenge where AI is poised to make a significant impact. McKinsey estimates that generative AI could automate business activities that absorb 60 to 70 percent of employees’ time. For us, proposal drafting was the obvious first candidate.
Our Hypothesis: The ‚Centaur‘ Model for Sales Engineering
I’ve always been fascinated by the concept of the ‚centaur‘ model, a term I first encountered in a Harvard Business Review article. It describes a collaboration where human intelligence is augmented by AI, not replaced by it. The AI handles the massive data processing and initial drafting, while the human provides strategic oversight, emotional intelligence, and final judgment.
This became our guiding principle. We didn’t want an AI to write and send proposals on its own. We wanted an AI that could produce a solid, 80%-complete first draft. This would free our engineers to focus on the most critical 20%—customizing the solution, adding strategic insights, and perfecting the final details that win a project. This is a practical application of my core interest in building systems that scale by combining human expertise with machine efficiency.
The Experiment: A Step-by-Step Breakdown
This project is part of our ongoing process of running experiments to improve operations. Here’s a breakdown of our approach.
Step 1: Curating the Knowledge Base
The first and most critical step was data preparation. The ‚brain‘ of our custom GPT would be our own archive of past proposals, technical datasheets, and project plans from JvG Technology. This repository contains years of accumulated knowledge.
However, it’s also filled with sensitive client information. Our first task was to meticulously anonymize this data. We systematically scrubbed all client names, specific pricing, confidential project details, and any other proprietary information. This was non-negotiable. We needed a clean, secure dataset to serve as the AI’s single source of truth.
Step 2: Building and Training the Custom GPT
With our anonymized knowledge base ready, we used OpenAI’s platform to build a custom GPT. We uploaded our curated library of documents and gave it a clear directive. I wrote its core instructions to define its persona: ‚You are a junior sales engineer at JvG Technology. Your task is to assist senior engineers by drafting technical proposals based on the provided knowledge base. You must adhere to our company’s tone and formatting standards.‘
This process transformed a generic language model into a specialist with deep knowledge of solar module production technology.
A flowchart illustrates the complete system we designed—a workflow where technology and human expertise are integrated at key stages.
Step 3: The Prompting Process
With the system built, an engineer can now start a new proposal not with a blank page, but with a simple, structured prompt. The prompt provides the AI with the core parameters of the client’s request.
This prompt contains just enough information for the AI to search its knowledge base and begin assembling the relevant components for a first draft.
Step 4: The AI’s First Draft
Within minutes, the GPT generates a structured document. It pulls the correct machine specifications for a 100MW PERC line, includes boilerplate text on our company history, and even incorporates the ‚Desert-Proof‘ framing option from a relevant past project. The output correctly structures the sections: introduction, technical scope, commercial terms, and appendices.
This aligns with a trend identified by Gartner, which predicts that by 2025, 30% of outbound messages from large organizations will be synthetically generated. We are essentially applying that same principle to a much more complex internal workflow.
Step 5: The Human Engineer’s Critical Role
The AI-generated draft is impressive, but it’s not the final product. This is where the centaur model comes to life. The engineer’s role shifts from writer to strategic editor. They review the draft, validating every technical detail, and add the nuance the AI can’t—referencing a specific conversation with the client, adjusting a technical parameter based on a new insight, or rewriting a section to better address the client’s primary concern.
This human review stage is where the proposal transforms from a competent document into a winning one. The engineer’s expertise is the final, indispensable layer.
Initial Findings and Next Steps
The early results are promising. We’re seeing a significant reduction in the time it takes to create a first draft—in some cases, cutting it down from hours to minutes. The quality of the drafts is consistently around a 70-80% completion level, providing a fantastic starting point for our engineers.
More importantly, it’s changing the nature of their work. They are spending less time on tedious assembly and more time on high-value strategic thinking.
Our next steps are to continue refining the knowledge base with new project data and to develop more sophisticated prompt templates. We’re also exploring how to integrate this tool more deeply into our CRM and overall digital infrastructure for a more seamless workflow.
Frequently Asked Questions (FAQ)
Is the AI creative or just copying old proposals?
That’s a common question. The AI doesn’t just copy and paste; it synthesizes information. It identifies relevant patterns and components from dozens of anonymized documents and reassembles them to fit the new prompt. The result is a new document that is structurally familiar but contextually unique.
What about data security and client confidentiality?
This was our primary concern from day one. The system is only as trustworthy as its data, which is why the anonymization step is so rigorous. No confidential client data is ever uploaded or used in training. We operate within a secure, private instance to ensure our proprietary process information remains protected.
Could this system make a mistake and send it to a client?
No. The system is designed with a non-negotiable ‚human-in-the-loop‘ safeguard. The AI is an internal drafting tool only. Nothing it generates ever reaches a client without a thorough review, edit, and explicit approval from one of our senior engineers.
How much technical skill do you need to build this?
The tools to create custom GPTs are becoming increasingly accessible, but the real expertise lies in the data curation and system design. The most difficult part wasn’t the AI setup; it was the strategic work of cleaning and structuring our knowledge base and designing a workflow that empowers our team instead of getting in their way.
This experiment is a small but clear example of how we can build systems that amplify our best people. The goal of technology shouldn’t be to replace human expertise, but to free it up to focus on the problems that matter most. Our digital apprentice isn’t going to close a deal, but it’s proving to be an invaluable tool for the brilliant engineers who do.
