One of my clients in the outdoor industry approached me with a simple request: “Hey Jeff, can you make the AI draft writer in gmail work better for our business?”. When I dug into the pain points they were having with the off the shelf Google Gemini drafts writer that is native to Google’s G-Suite, it was obvious Google shipped this feature without a lot of thought into the actual usefulness of the product. The AI being leveraged knew very little of the underlying business, the tone was off, and the output just flat sucked. Most of the Gemini outputted drafts needed more work than if you just wrote the email from scratch, but this was an area we should be able to apply AI successfully. So I went to work…
After a few days I had a working prototype (now in production) that combines a simple google doc with information about the business, a good underlying LLM (gpt 4o) and some automation that outputs a drafted email anytime a customer service email comes into the inbox. For fun, you can see an example of it in action here (I had permission from the company to post this).
Under the hood, here are the basics as to how I built this:
- The system looks for new customer service emails in the inbox (cron job or webhook)
- If an email is found that fits our criteria, it is parsed into just what we need to create an answer. (subject, name, body of email).
- The Google Doc (“business rules”) is pulled via API call into the system.
- The parsed email, business rules and uniform system prompt, is sent to gpt 4o
- The response is received and a draft is populated in gmail.
- The label of the email changes to reflect the AI has created a draft. This prevents the AI from rerunning over drafts that have already been written.
- The steps are repeated until all emails that fit the rules of the program are appropriately run through the system.
- The CX team reviews the draft, makes required edits and sends it to the customer who had the feedback.
Those wondering, the tool is set to run on a cron job every hour or, if needed, there is a button that an employee can press to run the tool (say, its super busy and they want to respond to a high volume of emails that came in).
The best part? The company is beyond thrilled with the output. The AI is often outstripping what the (human) cx employee would have written and the number of times the human needs to even add something is shockingly low. I know, its early days with this, but the business knowledge base gives the system enough data to meaningfully answer more technical questions and the process (putting it into drafts, not auto sending) prevents unwanted AI derived answers from going to customers.
Looking to the future there are a handful of things I’ll tackle to improve upon this tool, namely – cost. I’m passing too many tokens back and forth at the moment, which would matter for large CX teams and thousands of emails per day. There are a number of ways for me to solve this, ranging from simple (compress the business intelligence document using gpt), to some form of selective or vector-store retrieval. Obviously, complexity of a RAG/vector store retrieval approach goes way up, but is a solved problem and worth exploring for certain clients. This would lower the AI costs to fractions of a penny per query, which is tenable for even the largest of teams so long as the AI driven content remains useful.
Obviously Google is a powerhouse, shipping some of the best AI tools in the world. They have your entire gmail account at their whim, so I assume in the near future these agents will begin to leverage this data, and easily digest “external” data to work more precisely for whatever output you are aiming. To add, I am beyond impressed with what they announced last week at their Google I/O conference. As they kids say, “they are cooking”.
Ironically, after I shipped this, Y-Combinator’s Pete Koomen detailed his own gripes with Gmail’s draft writing AI assistant and his own solution, which is very similar to what I built. I’d argue mine is built more for teams using one email box and his more built for a single person, but it was cool to see some validation around this idea days after I shipped it. Pete also went into something I want to discuss in a future post under the umbrella of “how software development needs to change”. I won’t go into it here, but software teams are broadly stuck in yesteryear, and its leading to embarrassing results. More on this to follow…
Those readers wondering if they can use this tool, the answer is yes, but only in a business setting. For now, I’ll deploy it within organizations via a “white glove” approach (you contact me, I’ll set it up and make sure you like it, then I charge you a monthly fee). If its interesting, reach out!
Cheers
Jeff
* Google is incredible. My headline was a bit clickbaity – I’ll admit it!