How did we join the GenAI race with Smart Email Assist?

David Jambor
Outreach Prague
Published in
13 min readApr 16, 2024

How did we join the GenAI race? I had the privilege to be part of the team to design the first GenAi Outreach.io feature and I’d like to share a story of how we created Smart Email Assist in the revolutionary year 2023. But first, let me show you a short video to introduce this feature so you can better imagine what the next paragraphs are about.

The Journey to Sales Efficiency: How It All Began

It all started with a simple journey in November 2022 when I had the opportunity to go to London with our product & design team and shadow Outreach sales folks for two days. We’ve discovered that the initial phase of sales, commonly referred to as prospecting, is incredibly manual work.

Imagine this — At the beginning of the year your sales manager will give you a bunch of companies you should go after and sell your product. Then you have to research key individuals within the target companies. Once you do that it’s time to initiate contact through email or cold call to persuade them to schedule a meeting for product or service demonstrations.

Writing emails may seem easy, but writing effective sales emails is quite challenging. If you want to increase your chances that potential buyers will reply, then you have to write a concise and engaging email and it takes a lot of time and effort. In fact, this process consumes a huge portion of time. According to research, salespeople spend approximately 28%, on non-selling activities such as reading and responding to emails. So we asked ourselves: how might we make it more efficient?

When we got back from London, I immediately created a visionary concept based on our shadowing insights. We presented this concept to our leadership, demonstrating how prospecting could be revolutionized in the coming years. However, at that time, it was just a dream, it wasn’t feasible to build it. We thought it was at least 3 or 4 years away to materialize.

Fortunately, in December 2022, just a month after our research, the game changed with the arrival of ChatGPT 3. We immediately recognized the need to act quickly and wanted to integrate ChatGPT into the sales workflow to help salespeople write their emails in seconds, not minutes.

Defining focus

We received buy-in from our C-level executives quickly, but there was a catch. They wanted to announce this feature at the upcoming sales conference. That gave us a tight timeline of just 6 weeks to turn the idea into a final design and prepare a prototype for presentation.

While some features may require months of fine-tuning every design aspect, we operate in a fast-paced environment where rapid delivery of desirable products is key so this time constraint was a good motivation for us. In such situations, it’s crucial to prioritize and focus on what matters. We decided to focus solely on email replies, believing it would have the most accurate outcome. Our approach involved leveraging previously received emails from potential buyers to provide context, enabling ChatGPT to offer precise responses.

Our objective was clear: seamlessly integrate ChatGPT functionality into the sales workflow, as we believed it would offer a compelling value proposition.

Product lesson: Know the user workflows even those that are happening outside of your app, keep up with the technology trends, and be ready with the visionary concepts — these factors will drive your momentum and help you work on innovative features

Design explorations

The next phase involved designing the user interface. Initially, I thought it would be a straightforward task, perhaps something I could put together in a few hours — it’s just a single button, after all, right? However, to my surprise, it proved to be quite challenging.

During our shadowing sessions, we observed that salespeople frequently rely on company or personal templates for their emails. However, they often needed to customize these templates extensively to make them more personalized and engaging and they wanted to use their writing style.

Recognizing this as an opportunity, we aimed to enhance the functionality of Smart Email Assist by incorporating features such as the ability to choose the writing tone and set reading time (message length), in addition to the prompt. I understand that this functionality may seem evident now, but at the time, there was almost no design precedent, no app you could try and take inspiration from, so it wasn’t that clear.

Visual identity

Initially, I was considering a different visual identity for Assist. I envisioned a visual representation that would feel a little bit more alive and intelligent compared to a static icon. I aimed to create something that would give live feedback through animation, something that would provide positive emotions and evoke a sense of trust. Thus I’ve created this animated sphere aka our little Outreach Siri.

On the other hand, we needed a quick, simple solution that would be universally recognized. So we ended up using the sparkle icon too. It was a good move because almost every product uses the same visual representation or variation.

But the question is how long this visual representation will last. I think GenAi will be part of every digital product soon. It will work automatically in the background for you and it will be simply expected by users. So it will become the new normal and this special button will be like any other button in your product. So will we need to use any special visual treatment?

Designing user interface

In the first week, I designed approximately fifteen variations, and this particular one was the first promising solution.

We introduced a split button positioned in the right corner of the compose window to avoid overflowing the email text. Users could generate an email instantly by clicking on the sprinkle icon and the expand option would reveal additional options such as user prompt, writing tone, and reading time. Initially, this approach seemed solid, and many products like Copilot from Microsoft or GrammarlyGO have almost the same approach and I think it works for them, but we soon realized that it wasn’t the right fit for us.

  1. The issue was the interaction — Initially, I assumed the primary action would be users clicking the generate button. However, it became evident that users had a different mental model — they desired more control over the process. They wanted to access prompts and adjust voice, tone, or length before generating the message almost always. However, in our proposed design, users would always need to navigate two levels deep, resulting in a cumbersome back-and-forth interaction. Additionally, the design overflowed the email content, which might not be ideal in certain situations.
  2. The second issue revolved around the placement — Our design lacked coherence with other Outreach functionalities in the compose experience. While users had three options for writing an email — selecting an email template or snippet or using GenAI and the placement of Smart Email Assist didn’t align seamlessly with these options.

Although it may not seem problematic in the picture, it didn’t translate well to other surfaces such as Gmail, Outlook, and Chrome extensions, where consistency was crucial. As the conference deadline rapidly approached and we still lacked a solid solution, we decided to take a step back — not focus on the details, but rather think holistically. We’ve defined three design principles to guide design decisions.

Design principles

Consistency and alignment — We aimed to create a solution that felt predictable and cohesive, regardless of the platform. To achieve this, we introduced a new toolbar at the top of the compose window, consolidating content creation actions for easy accessibility when drafting an email. This approach proved effective across the Outreach web app, Gmail, and our Chrome extension.

Efficiency over distraction — Our second principle focused on prioritizing efficiency and minimizing user distractions. The entire process had to be easier than composing the email manually, otherwise, it would be useless. We designed a simple popover that appeared above the email body, offering options to write prompts and adjust length or writing tone.

Transparency & Guidance — At that time, many people were unfamiliar with the concept of prompts, as it was a new technology. It was crucial to educate users on how to formulate effective prompts because it’s the best way to improve accuracy. That’s why we introduced a dynamic placeholder in the user interface with a few examples to offer guidance and inspiration. As people get more comfortable with this technology, this design will change, but it was important to provide some guidance back then.

Also, AI is like a black box. Even if you provide great prompt and context the output can be misleading as GenAI can hallucinate. You can’t do much about that, but you can be transparent and acknowledge this fact. That’s why we’ve added a brief sentence in the UI advising users to review the generated text. Some people questioned the necessity of this guidance. Some said that it’s just additional noise in the UI. However, from our point of view, it’s better to be transparent and provide guidance than say sorry. Moreover, existing guidelines from Microsoft Research lab or other guidelines suggest the same and we wanted to follow them and ship an ethical product.

Product lesson: Be cautious when something looks too simple, because there may be complexities hidden beneath the surface. If the feature is important, many people will be looking behind your shoulder, judging your work, and trying to give you direction. Design principles can be your compass allowing you to confidently push back and justify your decisions.

1. Challenge — Feedback from data science

The goal of Smart Email Assist was to make the salespeople more efficient, by providing a single, accurate email suggestion. However, after consulting with the data science team, we realized that achieving this goal would not be feasible because of the accuracy. As a result, we needed to adjust the design, opting to generate three email suggestions instead.

Initially, I had concerns regarding this approach, known as “Human in the Loop,” fearing that it would slow users down and increase cognitive load. Surprisingly, after user research, we learned that users preferred this approach, mainly because it provided a sense of control.

2. Challenge — Costs of interaction

You might be wondering why users had to click on the Assist and then on the Generate button, instead of receiving suggestions right after clicking the reply button. There are two main reasons:

Contextual usage — Sometimes users write simple emails, like saying thank you, or confirming the meeting. In such cases, the benefit of using ChatGPT is minimal, unless there’s a need to write the email in a different language.

Cost saving — The primary factor here is cost. You have to pay for every input you provide to GenAI (how much) and also for every output generated by GenAI. So it could be expensive if we were to generate three options for every email for every user. Thus, we had to strike a balance between efficiency and product viability.

3. Challenge —Competitive landscape

The first version of Smart Email Assist debuted at the sales conference, and we received enthusiastic feedback from sales leaders. I was convinced that this would mark our little revolution and that we would lead the market. However, as often happens, the market had different plans.

Our initial plan was to launch the Alpha version by the end of March 2023. But do you remember what was announced in mid-March? Microsoft announced Copilot, followed swiftly by Google’s announcement of Bard, and soon after, a flood of other SaaS companies revealed integrations with ChatGPT. Within a matter of days, the competitive landscape changed dramatically.

4. Challenge — Accuracy

Another issue we faced was the accuracy of Smart Email Assist. The email suggestions didn’t consistently meet expectations. Sometimes, it would mix up the seller’s name with the buyer’s, produce inaccurate information, and fail to address prospect questions.

Imagine a scenario where a potential buyer asks a specific question like, “Hey, what is the benefit of your product compared to X?” This is a very specific question and for ChatGPT it’s challenging to fetch correct answers from the sources available on the public internet. So, the question was — how could we address these critical issues?

Sales leaders often create numerous email templates, snippets or Kaia content cards inside the Outreach platform, tailored specifically to sales organizations. Inside these templates or snippets are answers to specific questions that prospects might ask. This was exactly the kind of information we needed to feed to ChatGPT.

We leveraged the strength of our platform and we’ve built new functionality for content recommendation. So now the system automatically suggests using existing content as the reference to generate a more accurate email draft. With this context, Smart Email Assist can accurately answer questions like, “What are the benefits compared to…” or “What is your pricing?”. This is something neither ChatGPT nor Bard can accomplish on their own, as this information isn’t available on the public internet. Thus, this became our new stronger benefit.

Also, we’ve done a prompt tuning. We’ve selected a dozen of internal sales emails and we’ve experimented with slightly different prompts, models, and parameters to identify the prompt that consistently generated the best outcome.

From the user’s perspective, it works like this: When you click on the reply button, Smart Email Assist automatically suggests references based on the buyer’s previous emails. You can then review these suggested references, deselect any you don’t want to use, and simply generate your email.

Outlook version

When it comes to Outlook, it’s a whole different story. Outlook comes with its set of constraints. The only place you can embed your app is the side panel, and the interaction with email content is limited. This forced us to adopt a different UI approach while keeping our goal unchanged. While the user interface couldn’t be identical, the interaction design had to feel familiar to ensure a seamless transition for users already familiar with this feature in Webapp, or Gmail.

Product lesson: Before jumping into high-fidelity design, it’s crucial to focus on building the prompt, fine-tuning it, and discovering the combination that yields valuable output. Even with a well-designed interface, if the generated output falls short, the user experience suffers. Additionally, consider our value proposition, which should be unique and not easily replicable by competitors.

5. Challenge — The biggest one

What was the biggest challenge? As you probably expect Privacy and security were the issue. This is a new technology, and companies are naturally concerned about potential data leaks, which is understandable. Our legal team strongly recommended using the Opt-in approach, meaning we couldn’t enable this feature by default which creates issues with adoption.

Therefore, we had to develop an admin experience where administrators could toggle Assist on or off and manage access through user profiles. To encourage adoption we’ve created a few UI nudges to inform users about this feature and to communicate that Outreach doesn’t share any personal or customer data with ChatGPT. Also, we promoted the product through guided tours, videos, and promotional materials, and this approach was successful.

Product lesson: Privacy, and security need to come first, especially if your focus is on the commercial or enterprise segment. Also building a good product on top of AI models isn’t as simple as flipping a switch.

Final thought

Generative AI is here, creating unique opportunities that come around only once in a generation. So If you don’t work on GenAI yet, just send the message to your PM and start your product revolution today. But building a good product on top of AI models is still hard and if you’re skeptical, I encourage you to watch this brief video with Mira Murati, the CTO of OpenAI (see below). She explains why LLM-based features and products are hard to build.

As someone wise once said, “If you want to go fast, go alone. But if you want to go far, go with your peers.” So, kudos to my partners in crime in Product, Design, Engineering, Data Science, and Marketing. Without them, this would be a lonely journey, and the designs would still live in a Figma file.

--

--

David Jambor
Outreach Prague

Senior product designer, designing GenAI features @outreach.io, Ex-cofounder @Mentedy, father, davidjambor.cz