Use Case: Boost Your Productivity with a Custom Email Assistant

Hi there,

Today is a special edition – I’ve teamed up with the exceptional Fabian Werkmeister a seasoned data fellow and master at crafting educational data content, making him the ideal co-author.

Together, we're going to show how you can leverage AI, particularly through a technique known as task chaining, to significantly boost your email productivity.

And to do this, we’ll move beyond the usual "Can you please answer this email for me" prompt (which all too often doesn't work).

Are you ready for this?

Let’s go!

Want to learn how to 10x your data analysis productivity with ChatGPT?

Sign up for my event with O’Reilly on Dec. 12 & 13, 2023! Use this promo code to get free access to the platform and webinar for 30 days. Enjoy!

LLMs suck at complex tasks - until they don’t

Today's LLMs like GPT-4 are notoriously fast thinkers. They immediately spit out the answer based on obvious patterns and data, without any real reasoning behind it. This works well for tasks where surface-level data can provide an accurate answer, such as generating a brief summary from a larger text.

However, when it comes to more complex tasks, like developing a detailed marketing strategy for a new product or creating a comprehensive financial plan for a startup, they often fall short, lacking the 'slow', deliberate thought process that such tasks require.

This limitation is highlighted in ongoing research, as the Paper Thinking Fast and Slow in Large Language Models suggests.

The key to leveraging LLMs effectively for complex tasks lies in task chaining: breaking down larger tasks into smaller, manageable components that can be addressed within a single prompt.

This approach not only aligns with the current capabilities of LLMs but also unlocks a new realm of productivity and efficiency.

Let’s explore how this works on a concrete use case!

Problem: Managing Complex Emails

In business, we often have to break up a complicated task into smaller steps. This can be fun for new challenges, but it can be tiring for monotonous work like organizing meeting notes, handling travel expenses, qualifying leads, or: managing emails.

Each email requires us to go through multiple cognitive steps such as reading, understanding, categorizing, extracting action items, drafting responses, etc.

This process can be time consuming and exhausting, especially when dealing with a high volume of emails.

Let's explore how task chaining can streamline this process and turn the challenge of email management into an efficient, AI-powered task.

Solution Overview

To tackle the complex task of email management, we’ll use task chaining with AI, provisioned through a custom LLM assistant.

The custom LLM assistant helps us to automate the process of reading, categorizing, summarizing, and responding to emails. This not only saves time but also transforms your approach to managing emails, allowing you to focus on more creative and strategic tasks.

To build our use case, we'll use ChatGPT, but will explore other alternatives later (more on this in the "Now Let’s Build This Securely" section below.)

Let’s jump into the individual steps of our workflow!

Step 1: Create the GPT

Go to the GPTs editor in ChatGPT (Plus subscription required), and make sure to select the “Configure” tab (the “Create” mode is still too buggy). Give your GPT a name, a logo, and quick description to get started.

Step 2: Customize the GPT

Next, customize the instructions and available capabilities of the GPT. You can find the full configuration including the prompt here and the final GPT here.

Explanation of the instructions:

As you can see, the instructions contain a series of tasks that the GPT needs to perform:

1. List the sender & all recipients in tabular form

2. List the main topics of the email

3. Assign categories like Update, To-Do, etc.

4. Summarize the content

5. Highlight important notes

6. Read all attachments

7. Extract your todos

8. Draft a response

The cool thing about the custom GPT (i.e. the Assistants API) is that the LLM will automatically detect for which task it needs to bring in which tool.

For example, when the user provides not only the email text, but also some other files (the attachments) the AI will pull in the Knowledge Retrieval tool (a.k.a RAG) to make the contents of the files available to the LLM.

These tools make the whole process extremely powerful. With Function Calling, you could take this capability even further and pull in more data from other services as needed (e.g., look up the right contact for a particular email topic).

There’s literally no limit to what you can do. But be aware that the more “stuff” you cram into this process, the more failures can happen. This is still early beta!

When you’re done, click save and publish the GPT either privately or publicly.

If your GPT is private, you can uncheck the following box to make sure your data is not used for training purposes:

Step 3: Use the GPT

Once your custom assistant is published, you could simply copy/paste emails as shown in this video and you're ready to go! Tons of hours saved!

At least in theory.

Before you go and paste all your emails into ChatGPT, read the next section below.

Don’t skip this!

Now Let’s Build This Securely

Be aware that the approach above comes with some strong caveats:

Privacy and Security Concerns

• Exposing personal data: The approach above could potentially expose sensitive or personal data to OpenAI, which, without explicit consent, breaches data privacy laws. It’s vital to recognize and address this concern!

• Sharing company data with OpenAI: In addition to processing personal data, you would also be sharing critical company information with OpenAI, which - again - would not be in your best interest!

Technical Solution

To overcome the challenges above, we have different options alongside the make-or-buy continuum.

Option 1: Buy - ChatGPT for Enterprise: For organizations not strictly bound by privacy laws, ChatGPT for Enterprise offers a straightforward solution at approximately $60/user/month. However, it requires a significant commitment in terms of minimum seats and contract duration.

Option 2: Buy - Microsoft Copilot or Similar Solutions: Microsoft Copilot integrates directly with Office/Outlook, and while not inexpensive, it's a viable option, especially for those already within the Microsoft ecosystem. If you’re not on the MS stack, consider other integrated options like Google’s Duet AI.

Option 3: Build Yourself – Commercial APIs: Using OpenAI’s Assistants APIs means your data isn’t used by OpenAI for model training. This option allows for better integration with your existing systems, reducing the need for manual data transfer. While these APIs aren’t yet available on Azure, they offer a promising avenue for future development. Additionally, exploring other LLM offerings like Amazon Q or Google Vertex could provide similar benefits.

Option 4: Build Yourself – Open Source: For organizations where data privacy is paramount, such as those governed by GDPR, building a custom solution from an open source stack is a viable option. This approach offers maximum control over data privacy and security.

Conclusion

With the right approach, a custom LLM assistant can be a game-changer in enhancing your organization’s productivity.

However, it's crucial to consider the privacy and technical aspects carefully.

Whether you choose to buy a ready-made solution or build one tailored to your needs, the key is to ensure it aligns with your organization's privacy policies and technical capabilities.

If you're unsure where to start, I'm here to help guide you through the process, ensuring you make the best decision for your organization.

I hope you enjoyed today’s use case!

If so, please leave a testimonial – and don’t forget to try the E-Mail Assistant!

See you next Friday!

Tobias