I've been experimenting with a workflow over the last quarter that has significantly improved the quality and consistency of our initial lead follow-ups, and I’m really curious to hear if others have tried similar patterns or have ideas for optimization. The core idea is using ChatGPT to generate a first draft of a personalized follow-up message, which is then reviewed and edited by a human before being imported into our CRM for distribution.
Here’s the step-by-step process we’ve settled on:
* **Data Export:** Our marketing automation platform triggers an event when a lead reaches a certain score. We export a batch of these leads daily into a simple CSV file. The key fields are: lead name, company, industry, and the specific whitepaper or webinar they engaged with.
* **Prompt Engineering:** We have a dedicated, fairly detailed prompt for ChatGPT (using the API). The prompt instructs the model to generate a concise, professional follow-up email that:
* References the specific content asset the lead downloaded.
* Asks one relevant, open-ended question about their current challenges related to that topic.
* Offers a single, clear next step (e.g., a 15-minute discovery call or a link to a related case study).
* Maintains a helpful, consultative tone—no hard sell.
* **Batch Generation:** We feed the CSV data into a simple Python script that calls the ChatGPT API for each lead, generating a unique email body. This runs as a scheduled job on a serverless function.
* **Human Review & Edit:** This is the critical gate. The output is compiled into a shared document where a sales development rep reviews each generated message. They check for:
* Appropriateness of the question.
* Any awkward phrasing.
* Company or industry nuances ChatGPT might have missed.
* Overall flow and tone.
* The rep makes minor tweaks directly in the doc—this usually takes 30-60 seconds per lead.
* **CRM Import:** The finalized messages, linked to the correct lead IDs, are imported back into our CRM (Salesforce) via a custom object. They populate a dedicated "Next Follow-Up" field on the lead record, from which the SDR can send them with one click.
The results have been promising. The human-in-the-loop edit is non-negotiable; it catches subtle context errors and adds genuine empathy. However, the draft from ChatGPT provides an 80% solution, ensuring a consistent structure and saving our team from staring at a blank screen dozens of times a day. It also enforces a baseline quality and includes that strategic open-ended question we used to sometimes forget in a hurry.
I'm particularly interested in the intersection of this with cost management and observability. We're tracking token usage per lead and have alerts on unusual spikes, which feels like a tiny FinOps practice. My main open questions right now are:
* Has anyone built a more integrated review interface, perhaps directly inside their CRM, to streamline the edit step further?
* Are there effective patterns for dynamically adjusting the prompt based on lead source or industry without creating a maintenance nightmare?
* How do you measure the impact beyond just time saved? We're A/B testing response rates against our old template-driven approach.
The workflow isn't fully automated, and that's by design. It feels like a pragmatic middle ground between generic automation and fully manual, bespoke outreach.
~jason
~jason
This is a smart use of the API. I'm curious about the prompt structure, especially how you handle the single open-ended question. Do you feed it a list of pre-approved questions tied to each content asset, or does it generate something unique each time based on the industry and asset name?
That's a solid process. The human review step is the most critical part there, I think. It's easy for that kind of automation to drift into sounding generic if someone's just rubber-stamping the drafts.
My main question would be about scaling the editing workload. Does having a detailed prompt actually reduce the editing time, or does it just shift the effort from writing from scratch to reviewing and tweaking? If the team gets a hundred of these drafts a day, that review queue could become a real bottleneck.
Keep it civil, keep it real.
I've been thinking about trying something like this in Freshdesk. Your point about referencing the specific content asset is interesting. How do you handle it when a lead downloads more than one asset? Does your prompt prioritize the most recent one, or does it try to combine them?