In this article, I’m using several AI-related terms interchangeably. The text relates to Seam - a chat-based SaaS unlocking the power of data insights and operations for business users. Seam is built on top of GPT-4, a Large Language Model. Large language models fall under generative AI (models capable of generating different types of content based on their training). Generative AI falls under a much broader category of Machine Learning, which refers to a subset of artificial intelligence algorithms capable of analyzing patterns in data and performing tasks without explicit instructions. All terms are encompassed by an umbrella of “Artificial intelligence,” referring to any computer or machine aiming to simulate a human-like thought process. While this language makes sense in the context of Seam, it’s worth noting that the terms are not interchangeable in general, and each refers to a specific, distinct software category.

Introduction

In the past five years, most major tech companies have shifted their focus towards AI. Google, Microsoft, and Meta are betting heavily on artificial intelligence, particularly generative AI. Even though the technology is evolving rapidly, the potential (and demand) is astronomical. 

Working with AI startup founders daily and seeing their ideas dawn, pivot, evolve, and succeed, we've also started noticing our approach to creating products evolve. When designing human interactions with machines simulating human-like behavior, a whole set of factors, from practical through ethical to philosophical, has to be considered. Some might call it "science fiction" - we call it a Tuesday afternoon.

This article will explain the paradigm shift in product design we've observed over the past several years and how it affects our approach to design at Semiflat.

Fundamental differences between AI and traditional computing

The concept of artificial intelligence has captured society’s imagination ever since tools like ChatGPT and Midjourney were released to the public. Their ease of use allowed even casual users to interact with machines that could help them make decisions or even be creative. For the first time, in the eyes of an average consumer, everyday technology was human-like. People across all industries have found ways of improving their workflow with tools like ChatGPT and “AI,” as the umbrella term has leaked into the mainstream. 

Our work as product designers centers around problem-solving. With the emergence of systems that operate (to an extent) independently, designing experiences that help users adapt to that change is critical for driving their adoption. Since AI systems operate based on previous training rather than just explicit user input, we need to evaluate how humans use technology in their everyday lives. The design process behind products leveraging AI needs to consider people's evolving relationship with machines, which is a direct consequence of the significant difference between artificial intelligence and classical computing.  

So… What is the difference?

Conventional computing operates deterministically, following predefined rules to execute tasks predictably. Take sorting, for instance; selecting 'sort from A to Z' will always arrange records alphabetically. In other words, each user action yields the same outcome every time, allowing for total predictability of software interactions. 

The emergence of large language models like GPT-3 has fundamentally changed the way humans interact with machines. Instead of returning an output based on a predefined algorithm, these models can analyze the context and return a nuanced response based on the patterns learned on large data sets in the training phase. The way it generates responses is probabilistic. While both a calculator and an LLM can correctly answer the question “What’s 4+4?” - how they arrive at that answer is vastly different. While a conventional computer would return an answer based on permanent instructions, an LLM will compute the likelihood of various responses being correct based on the patterns it learned during its training.

What does it mean for product design?

Our work as product designers and user experience designers enhances human lives. We use a combination of research, proven patterns, and problem-solving to create products that help humans perform tasks faster or enable them to do something previously impossible. 

‍

Creating aesthetic wireframes is not enough in a world where countless digital products compete for user attention. What any founder should want is to design a relationship between the user and their software. According to The Media Equation, humans tend to assign human-like characteristics to computer systems. Building a relationship with the user based on trust and reliability is critical to keep them continuously using the product. This is even more true for products utilizing any form of artificial intelligence. With interactions happening primarily through a natural language interface, the line between a human and a machine can become blurry. That’s why many people instinctively use words like “please” and “thank you” when interacting with LLMs (as seen on multiple LinkedIn memes). 

When designing a digital experience based partially or entirely on artificial intelligence, it’s essential to ensure that the end user understands the capabilities and limitations of this interaction. This way, users will know how the product can help them and what they can expect, drastically reducing their frustration and improving the overall experience. 

Consequences of neglecting design in AI products

While user experience is rarely mentioned in the general narrative surrounding AI, failing to consider how users will interact with it has profound consequences on business metrics like user retention and churn. No matter how amazing the technology is, from a business standpoint, it is ultimately worthless if it doesn’t reach the right users. Neglecting product design in an LLM-based experience can result in the following:
-list-normal-

• Users won’t trust the product.
• Users won’t know its capabilities.
• Users will have unreasonable expectations.
  

-list-normal-

These factors will inevitably lead to lowered confidence in the product and users turning away before they can fully experience its potential. Good user experience practices can salvage users' relationships with your product, even if they have a bad interaction (for example, if the model returns inadequate or incorrect results). 

How we approach designing AI products at Semiflat

Being involved in designing multiple new digital products utilizing Machine Learning algorithms and Large Language Models, we’ve developed a set of principles that underlie most of our design decisions.

Core principles we stick by

Every product is different, but there are core principles that we always consider when designing products that utilize any form of artificial intelligence.

-list-normal-

• Transparency - to establish trust from the first use of the product, we call out the features utilizing any decision-making algorithms. This way, users can immediately distinguish between “regular” and “smart” features, which sets their expectations for the interaction.
• Continuous feedback - users need to be able to easily provide feedback to the responses generated by the model so it can continuously improve. This ensures that users understand the algorithm is supposed to learn and can improve over time, encouraging them to interact with it further.
• Explainability - every decision made independently by the algorithm needs to be easily explained. Users need to verify the logic or steps the algorithm took to arrive at the response. This reinforces trust and helps identify potential errors in the algorithm’s reasoning. 
• Full control - it’s essential to communicate to the user that they have control over the system. They need to confirm any important decision before the AI can execute. This way, we build an understanding that the algorithm serves the users and will not make any decisions that might have undesirable consequences.

-list-normal-

The controversy of the magic symbolism

In principle, I fundamentally reject the notion that AI or its subsets are supernatural in any way. I find this belief more harmful than beneficial. This attitude is reflected by AI experts worldwide, like IBM’s Rob Thomas:

“Today, too many people view artificial intelligence (AI) as another magical technology that’s being put to work with little understanding of how it works. They view AI as special and relegated to experts who have mastered and dazzled us with it. In this environment, AI has taken on an air of mysticism with promises of grandeur, and out of the reach of mere mortals. The truth, of course, is there is no magic to AI.”

Rob Thomas

Senior VP, Software and CCO at IBM

At the same time, many people have developed a universal correlation between magic and AI. The magic wand, or “glitter” symbolism, has become synonymous with the ubiquitousness of this new technology. I accept that this is the case, especially among non-technical users. Because these correlations have developed - I believe that in some cases, using this symbolism is the fastest way to signal to the end user that they’re dealing with a Machine Learning or LLM Algorithm. 

When designing Seam, we sporadically introduced magic wand iconography to label LLM features for that exact reason. Seam targets business users without a technical background. We’ve decided to call out these features to increase transparency between the software and the end user by relying on a familiar metaphor (the magic wand). This is not universally applicable, though, and while faced with representing AI in a different product for more technical users, we’ve opted to refrain from referencing the supernatural in our custom iconography. (Future - Link to the article about designing the custom icon set for Composabl)

Meet Seam - the AI platform for customer information

The mission to democratize data access with AI

Data rules the modern software landscape. It helps companies make better product decisions and gives insights into reaching more users and generating more revenue. As software companies scale, so does their data. The sheer number of sources makes it hard to keep up. Over time, information becomes siloed across dozens of sales, marketing, customer success, and finance applications. Access becomes even more difficult for business users who don’t know SQL. They are effectively cut off from access to customer data they need and rely on technical users as their gatekeepers. 

The founders of Seam, Nick, Mikhiel, and José, noticed this problem while working at Okta. As the company scaled from $100M to $1B in revenue, accessing, understanding, and activating the data they needed became increasingly complex. That’s when the idea for Seam was born.

One interface to rule them all

Seam aims to solve all frustrations around data insights experienced by business users at fast-growing software companies. It’s a simple chat interface that ties together any data source their company might use. Once the sources are connected, a simple query in natural language will unlock complex data operations without writing a single line of SQL.

‍

Two user personas living in harmony

While created primarily for sales, marketing, and revenue operations teams, Seam’s interface affords technical users full control. Engineers and data analysts can create and tweak verified queries and definitions to ensure the entire team can access the most reliable insights contextualized for their company's specific use case. The interface we designed considers different ways business and technical users interact with customer data and provides each group with a Seam-less (pun intended) experience specific to their needs.

Link to the article about how we’ve designed an interface for both personas

Live demo

Interested in trying out Seam for yourself? The live demo is waiting.

Try live demo

What we’ve considered in the design

Minor enhancements with a big impact

Seam has approached us with an early version of their MVP, which they had built internally. Our immediate goal was to identify areas for improvement that would yield the most significant impact with a relatively small effort.

One of the first things we noticed was the terminology around conversations users would have with Seam. The team was referring to them as “analyses.” While the term is correct, it felt less human and more machine-like. The term “analyses” brings to mind the actions performed by the model “under the hood.” We’ve proposed renaming them to “threads” to help users think of interacting with Seam as simply conversing with AI. On top of that, we’ve simplified the navigation structure and enabled users to access recent threads from every place in the application, as opposed to just the “Analyses” tab.

Two steps ahead of the users

One of many things AI systems are particularly good at is anticipating user behavior based on historical data. Seam’s users would likely need to perform the same queries periodically, so an intelligent suggestion mechanism was essential.

Seam’s initial MVP included a ‘saved models’ functionality and a ‘prompt library’ accessible directly from the start page. These features weren’t clearly highlighted and, therefore, were easy to miss, creating confusion around their purpose.

Striving for maximum clarity, we’ve included a prominently displayed suggestions box as the central piece of the empty thread page. This gives users who aren’t yet familiar with the app or interacting with LLMs, in general, a clear call to action they can take while interacting with the product for the first time. Seeing the value with just one click reinforces user’s confidence in the product and encourages them to experiment with their own queries. The suggestions are also helpful to returning users who run the same query regularly. We’ve designed suggestions to empower users in two ways: as a shortcut for frequently run queries or to help them understand their data better by suggesting contextual follow-up questions under Seam’s original response. 

Always bring receipts

Building on the core concept of transparency in product design for AI products, we’ve walked a fine line between providing access to complex information (data sources, operations, definitions, and steps followed) and maintaining the simplicity of a chat interface. While giving information about generated answers, the existing interface was unstructured and cluttered with context that might not have been relevant to the user. This increased the cognitive load and defeated the purpose of using a familiar and straightforward chat-like UI. 

We’ve simplified the thread interface and displayed only the most relevant information (a direct answer to the user’s query), making it easy to access more granular information with a single click. This decision ensured a more contextual experience. Business users can quickly get the answers they’re looking for with minimal cognitive load. Technical users can launch a ‘details’ tab that augments the view by providing information about sources, definitions, and table operations performed to produce the answer.

What’s next for Seam?

Seam has just launched its product and updated its visual identity to the public. The team aims to help more high-growth teams democratize access to data and become the trusted data platform for customer information.

Hundreds of hours saved

The team has been working with fast-growing teams at companies like Redpanda (a Series C streaming data platform from San Francisco) and Apploi (a Series B workforce management platform from New York) to unlock their data for business users and save hundreds of hours per year. 

In Redpanda’s case, customer data was siloed across multiple systems. Access and reporting were a nightmare for business users and required manual processes that were time-consuming and non-scalable. Seam provided a flexible and scalable solution to track attributions and identify new pipeline opportunities. The growth team also incorporated more buying signals in their account-based marketing strategy based on product usage data, all with a minimal, two-day onboarding process. Thanks to Seam, Redpanda saved hundreds of thousands of dollars across multiple tools and more than 720 hours annually on manual data work (Read the case study on Seam’s website). 

• 60 hours a month savedRedpanda reduced the team’s time spent on manual data work by 60 hours per month (or more than 700 hours annually).
• $100,000 saved on toolsBecause Seam replaces multiple data-related use cases, Redpanda was able to save six figures per year in tooling alone by building solutions in Seam.

“We have so much data in systems, and we want everyone –sellers, marketers, CSMs, SDRs, executives– to have that data at their fingertips. We want them to wake up and know what’s happening in their territory. Seam is going to be a huge part of getting there.”

Yuliya Maystruk

Sr. Manager, Marketing Operations at Redpanda

Apploi - another one of Seam’s customers, was also able to achieve great results within months of implementing the app in their organization. Before using Seam, their go-to-market data was scattered across multiple systems. The lack of a single source of truth cost them missed revenue, and reporting was almost impossible. To make matters worse, different teams were working on different metrics (with every team favoring the tool they were most familiar with). That led to substantial alignment challenges and decreased confidence in their data, ultimately hindering their ability to grow. 

“Even if you say system A is the source of truth over system B, teams will orient towards the system they are most comfortable with. Different metrics derail the conversation, you’re less focused on solving the problem and you’re more focused on if you even have the right inputs.”

Elliot Farris

VP of Operations and Analytics at Apploi

Apploi implemented Seam as its analytics engine on top of the entire GTM stack. The company unlocked customer intelligence that was previously outside of its reach, automated manual processes, and made sure every team was on the same page. This resulted in substantial savings - 328 and 590 hours annually on manual analytics and operations work, respectively. Additionally, their team identified over $31M of new addressable market for sales to target. (Read the case study on Seam’s website). 

• 328 hours of analysts' time saved per yearWith Seam, Apploi was able to connect all of its data and start generating customer intelligence automatically, saving hundreds of hours on data analysis annually.
• 590 hours across operations teams saved per yearApploi used Seam to build automation that helped the finance, sales, and customer success teams save 590 hours annually.
• $31M of new market opportunities identified with Seam
  

The team is very excited about the future they’re now able to build with Seam:

“I think this is the first tool I've actually come across where the options are almost limitless. If you can think it up, there is probably a way to do it with Seam. And it won’t be hard.”

Elliot Farris

VP of Operations and Analytics at Apploi