Struggling to convert my ChatGPT plugin into a paid SaaS product

I’ve been working on a niche tool for six months that uses the OpenAI API to automate invoice extraction for small accounting firms. It works well, but I’m hitting a wall trying to move beyond a free tier. Right now, I have about 200 active users, most of whom are using it via a basic web interface. The cost of API calls is eating up my margins because I’m using the GPT-4o model for every single request, regardless of complexity. I tried switching to GPT-4o-mini for the initial parsing step, but the accuracy dropped significantly on messy PDFs, and users started complaining. My main question is about pricing strategy and technical optimization. I want to charge $49/month per firm, but I feel like that’s too high for a tool that doesn’t have a massive brand behind it. I also don’t know how to handle the security concerns properly. Right now, I’m just passing the PDF directly to the API without much local redaction, and a couple of users have asked if their data is being used for training. I know the answer is no, but explaining that to non-technical accountants is tricky. Has anyone successfully navigated this transition from a side project to a sustainable business? Specifically, do you recommend building a custom vector database to handle the document context before hitting the LLM, or is that overkill for this use case? I’m leaning toward RAG to lower the token count, but I’m worried about the implementation complexity slowing down my launch timeline. Any advice on balancing cost efficiency with user trust would be appreciated. I’m mostly using Python and FastAPI for the backend, so keep suggestions within that stack if possible.