Chapter 2: Creating Product Ideas

Coming up with great product ideas isn't easy. It's both an art and a science, and those with the ability to come up with great ideas are remembered in history, from people such as Steve Jobs to Mark Zuckerberg.

In this chapter, we'll explore the five pillars of AI, which are driving new and innovative ways to create product ideas: language understanding, visual understanding, information extraction, information organization, and creative AI. Understanding the pillars of AI will help you become more strategic about how you plan, manage, and invest in your AI projects.

For instance, some product teams might like to use AI to generate new product designs, which would focus on the pillar of visual understanding, while others might like to use AI to understand desired customer features that will lay the framework for new products, which would involve pillars such as language understanding and information extraction.

After covering that, we'll cover building, selecting, and iterating product ideas. Finally, you'll learn how to use Commerce.AI to improve the product ideation process and take advantage of billions of data points to gain a competitive advantage.

In this chapter, we will cover the following topics:

• Understanding the pillars of AI
• Why is product ideation so hard?
• Using Commerce.AI for creative AI
• Building product ideas
• Selecting product ideas
• Iterating product ideas

Product ideation is crucial to business success. In the past, companies have failed to create products customers wanted because they didn't know what the customer wanted until after the product was built. But that approach is now increasingly obsolete. The internet has given companies unprecedented access to customer data and insights about their customers at any given time, allowing them to build better products faster than ever before, aided by the power of AI.

Understanding the pillars of AI

These are the five pillars of AI, which lay the groundwork for using AI for product innovation:

1. Language understanding
2. Visual understanding
3. Information extraction
4. Information organization
5. Creative AI

When you combine the first four pillars with creativity, you get what's called creative AI. In other words, the first four pillars are needed to create the data structure that fuels a gamut of creative use cases.

Creative AI is an advanced form of artificial intelligence that can solve problems previously thought impossible for machines, whether that's designing wholly new products or coming up with truly innovative ideas, such as how Google used AI to design rounded and organic computer chips much faster than human engineers, or how designers at Autodesk use AI to design skeletal scaffolds that are lighter, stronger, and more efficient than regular designs. In this section, we'll explore the five pillars of AI, and how they tie into product creation, in greater detail.

Language understanding

Language understanding is the ability to read users' minds.

One of the most important pillars of AI innovation is language understanding, which allows machines to interpret human text and reasoning, and then return a response that the user can understand. The ability to do this is often referred to as natural language processing (NLP). NLP has been around for decades, but just recently it has witnessed significant advances through deep learning algorithms.

In contrast to traditional machine learning methods, deep learning can identify patterns in data using neural networks. This approach makes use of large amounts of datasets and produces accurate results at higher speeds than previous methods. Deep learning can not only predict future outcomes but also perceive the user's intent or state of mind from their voice or writing.

For example, if you ask a Google Home device about the weather in San Francisco tomorrow, it might respond with It will be sunny with a high of 82 degrees Fahrenheit. This is due to deep learning technology interpreting your spoken words as requests for information, and providing you with what it believes you want based on its vast knowledge base.

The point here is that today's modern NLP technology enables machines to understand our intentions better than ever before, which means we can effectively build conversational
agents easier for a variety of tasks (for example, when scheduling a meeting). The good news is that because modern NLP technology isn't very complex, compared to other AI technologies such as image recognition, we don't have any shortage of examples where it could benefit our lives.

One way language understanding helps product teams is by generating new product ideas from existing ones. For example, say you have an existing product idea for a new hard hat for construction workers. You could use language understanding to automatically extend that idea into a smart hard hat that monitors the worker's location and warns them about their proximity to dangerous objects. At a high level, this is like autocomplete on steroids. We'll explore how this works in the Transformers section.

Another application of language understanding is helping companies scan online user reviews for specific features that users want to be added to their products. For example, if you go to Amazon right now and look at the reviews section for any popular product, you will see dozens of customer comments asking about the product features, either explicitly or implicitly. A popular snack includes comments such as flavor is mediocre and the snacks are clumped together.

With language understanding, these comments can be extracted and turned into a customer wishlist. Product teams can use these wishlists as feedback signals, telling them what their customers want from their products so that they can improve them over time.

Given that 95% of the data on the internet is unstructured and largely textual data, language understanding is a crucial pillar of AI. This technology applies to voice data as well, which enables new ways to collect and analyze customer feedback for market research, such as voice surveys.

In fact, at Commerce.AI, we've found that 95% of research participants prefer voice surveys over traditional survey forms.

Visual understanding

Visual understanding is about recognizing objects using images. An intelligent computer program can understand the elements of a picture and use that understanding to generate new ideas, and AI programs can also be trained on user feedback to develop new products based on what customers want. These programs can gather customer feedback, such as through our product data engine, voice surveys, or focus groups, and then turn it into product ideas using AI. This process produces an enormous amount of data about users' interests.

Let's explore how visual understanding is used in three different types of product innovation:

1. Generating better insights on products and product feedback

   First off, visual understanding addresses the task of finding data about people's needs, where an AI system can be trained to recognize objects within images. This enables better insights on products and product feedback.

   For instance, there are millions upon millions of products listed on Amazon. Factoring in Amazon Marketplace; the number is estimated to be around 350 million
products. Many of these listings have not diligently listed all the visual information found within the photos. In other words, Amazon product photos are another crucial source of product data. Visual understanding can be used to understand these photos and their details, which is particularly useful when the textual descriptions fall short.

   Apple uses facial recognition, a form of visual understanding, to enable users to effortlessly unlock their iPhones. They also use augmented reality, another type of visual understanding, to create personalized emojis.

2. Developing new products by mining user feedback

   Beyond generating better insights, visual understanding can also help you develop new products and features. For example, many online reviews include images of the product while focusing on standout features, but also defects. Visual understanding can be used to analyze these images, in addition to regular text data, to help inform product teams of what products and features to change, and to avoid similar defects in the future.

3. Developing solutions for users who are visually impaired or blind

   Let's look at one more benefit of visual understanding related to user experience innovation: Visual understanding is also used to improve software that allows blind computer users to access websites more easily. For example, many images on product
listings lack alt-text, or the invisible description of images that are read aloud to blind users. With visual understanding, this alt-text can be generated automatically.

   This can be done by using AI libraries such as OpenCV, which is a popular framework for computer vision and machine learning. OpenCV uses techniques such as Convolutional Neural Networks (CNN) to perform image classification tasks by extracting patterns and features to classify images based on what has been learned previously.

These neural networks might extract features such as color histograms, edges, and shapes in the image, or any other feature that makes it easier to identify and distinguish between different objects that, once identified, can be added as alt-text.

Information extraction

Information extraction is the process of extracting information from unstructured textual sources to enable finding entities, as well as classifying and storing them in a database. This is a big part of data science and AI.

At the time of writing, we can use a variety of tools to extract information from unstructured textual sources, such as NLP and deep learning.

By using information extraction on customer data, you can start to turn data into insights.

Using customer data

Information extraction is, of course, a huge part of extracting product data, but it's also used in the context of customer support services and can help you with questions such as the following:

• What do customers say about our products?
• How do customers react to different claims in our product descriptions?
• Which words are most commonly mentioned in our product reviews?

This customer data can come from a huge range of sources, including forums and blogs, surveys, videos, customer support tickets, CRMs, and more. The following screenshot highlights the very tip of the iceberg of customer data sources:

Figure 2.1 – A sample of customer data sources

The holy grail is to answer customer questions preemptively or make suggestions based on what users have said so far. This kind of service might sound futuristic, but there are some companies already offering it. For example, Expedia has just released an AI bot that uses machine learning and NLP to help people book hotels. More specifically, this uses what's called intent extraction, which means finding out what kind of information a user wants. For example, if a user types my card isn't working in the chatbot, the system needs to process that as a request for payment information.

The intent extractor looks at all the possible intents and tries to match them with what the user has typed. At a technical level, this works by using a model trained on past conversations that were labeled with user intent. The model will understand patterns in the user language to match the user's messages with intents, even if the keywords aren't identical.

However, the biggest challenge here is figuring out how to scale up NLP for commercial purposes without getting trapped into using canned responses or preselected question sets. Chatbots should be able to have natural conversations with humans, and canned responses can stunt the human-like qualities of communication, which means that high-quality NLP tools are needed.

Using AI

This is where AI comes in handy: to generate new ideas that can be tested out before spending money on building a new chatbot or hiring new employees for customer service jobs.

While extraction is not the solution in itself, it will enable us to ask better questions about problems and identify opportunities for improvement. Data extraction provides a means to quickly learn about product reviews, sentiment, and market data. By providing insights into customer behavior (for example, what features users use most frequently and how often they encounter particular problems), we can ask better questions about why users face the problems they do, and how to fix them.

After extracting data, we can move on to organizing it.

Information organization

Information organization is the process of organizing extracted data to come up with greater insights.

Information organization creates structure and establishes the relationships between products, brands, and attributes, enabling the creation of new ideas, descriptions, and even ad copy. For example, the following screenshot highlights information about watch brands organized by Commerce.AI into a Feed Summary. From just a glance at this organized information, we can see that customers appreciate Casio's prices, but complain that they're made out of plastic and have bad batteries.

Figure 2.2 – A sample of organized information about watch brands

AI can organize information faster, cheaper, and more effectively than human beings. Let's look at five specific ways in which AI can help with information organization.

Reducing information gathering time

While collecting data is crucial for product innovation, it is also time-consuming. This is one of the biggest drawbacks of collecting data from users and customers.

More specifically, you need to collect massive amounts of user feedback or customer preferences over a certain period (which takes up a lot of the budget). But with the advent of powerful AI tools such as machine learning and NLP, you don't have to spend as many resources on data analysis anymore.

Instead, you can save tons of money by outsourcing data analysis to AI that knows how to read user feedback. This approach allows you not only to save resources but also to make more informed decisions about products and improve your business results faster.

Information organization depends on gathering the right product information across brands, product categories, and attributes. By using AI to speed up information gathering, we can make the right product data available for information organization, faster.

Eliminating duplicative data gathering

Another major drawback in running surveys or interviews is getting redundant responses from customers or users.

For example, if I want to find out about my customer preferences on various products available in the market today, I would ask various questions, such as, Which smartphone do you currently own? What brand do you prefer? Are there any other features that come standard with these phones?

With the help of artificial intelligence tools such as machine learning and NLP, you can get rid of repetitive feedback by automating tasks such as sending out surveys or conducting interviews automatically using chatbots instead.

Making better suggestions based on big data

There are times when we tend to rely on our judgment rather than analyzing what people want or need. Some examples of this are when a new product idea comes to mind, when we imagine a feature we'd like our products to have, and when we think about improving an aspect identified as an issue.

Since users often explicitly mention desired features and product attributes, whether it's in Amazon reviews or YouTube unboxing videos, it's a better idea to tap into this big data rather than relying on gut feeling. High-quality organized information requires big data, especially when it comes to understanding customers at scale. By using AI for big data analysis, companies can mine their data to find trends and patterns that they might not have been able to see before, enabling higher-quality information.

Gaining value from existing data

The most common method of feedback gathering that's used by companies today is asking users questions.

However, this feedback often already exists in the form of millions of online product reviews, both by the firm in question and its competitors. The problem with this data is that it's unstructured and unorganized, which means that there's untapped value. With AI, companies can finally structure and gain value from this existing data. As AI becomes more and more accessible, product teams are starting to take advantage of this unstructured data, but it's mostly still trapped under the technical burden of analyzing large amounts of raw data. As a result, product teams can still gain a competitive advantage by using AI.

Creating heatmaps

Machine learning helps generate heatmaps visualizing information gathered from online forms, filled up across multiple platforms such as web pages – both mobile websites and desktop sites – provided by services such as Google Analytics. This organized data helps companies before and during creative ideation. By organizing what users care about into a visual hierarchy, product teams can focus their creative ideation process on what truly matters. For example, suppose you're an automotive firm, and users are far more interested in your self-driving features than fast charging — this would inform your creative ideation process to focus on improving your self-driving features even further.

Creative AI

By using AI to generate ideas, we can build more products that delight customers and turn them into lifelong users. In a nutshell, we leverage the latest AI technologies, particularly large language models, to guide product ideation.

This process is called concept generation with language models (CGLM). CGLM is an effective strategy for generating a range of novel concepts that meet user needs and desires.

Beyond generating natural language, creative AI can be used to design new products using a technology called Generative Adversarial Networks (GANs). The idea of GANs is to train two competitive neural networks, where the first network generates fake images and the second network discriminates whether the images are real or fake. The generative model iteratively tries to fool the discriminative model until real images are indistinguishable from the fake ones. GANs were infamously applied to create deepfake photos and videos of celebrities and politicians, but they can also be used to generate product concepts, whether it's a new apartment layout or a sneaker design.

These types of creative AI support the product ideation process, which is typically a long, hard endeavor.

Why is product ideation so hard?

One of the most popular methods for generating new product ideas is through the use of brainstorming exercises, such as ideation exercises or mind mapping techniques, which are used by designers, architects, and engineers who are stuck or have reached dead ends in their creative processes. These methods are very useful because they enable people from different fields to come together and share different perspectives.

However, this method has limitations in terms of generating truly novel commercial products, given its focus on imagination rather than on actual customer needs and desires.

Another technique that's used by many companies today involves including customers in ideation processes, via surveys or focus groups, where they share their opinions about what they want. While this approach can be better for incorporating customer feedback, it's also costly and time-consuming. Domain experts have to manually take notes, collate and analyze feedback, and merge this with external data to move forward. They then need to find a way to collaborate with the actual product development teams to ensure that feedback is understood and used.

This lengthy, expensive process is also highly limited in terms of the data that's collected. A focus group can only include so many users. With AI, all available product data can be analyzed, including textual product reviews, product descriptions, video reviews, voice surveys, and more. AI is also much faster than humans, which means this data can be extracted, organized, and analyzed in real time, providing insights directly to product teams. This organized data acts as the fuel for creative AI as well, including product ideation. At the end of the day, product ideation is vital for success, lest your business end up as a graveyard firm that failed to innovate.

And it doesn't have to be burdensome, as Commerce.AI shows. Commerce.AI is a platform
that allows you to easily build and manage your product ideas, from ideation through development and launch.

Using Commerce.AI for creative AI

As we've established, creative AI is powered by language understanding, visual understanding, information extraction, and information organization. With Commerce.AI, these components come together to let you generate new product ideas at will.

Commerce.AI's standard analytics include a product leaderboard, graphs of the top products and brands by market share over time, a review breakdown of the top products and brands (by stars), a market landscape graph, and more. The creative AI component uses this data, in addition to selected customer wishlist information, to generate new product ideas. Similarly, you can generate ad copy as well, which is typically a tedious, manual task, that can now be done effortlessly.

As we can see, AI can be creative, and a powerful tool for product teams to expedite their innovation process. Now that we have a background in AI and product ideation, let's use this knowledge to learn how to build, select, and iterate upon product ideas. Naturally, building product ideas is the first step, but not every idea will be a hit, which is why it's important to diligently select and iterate upon the best ideas.

Building product ideas

Innovating new product ideas is a serious challenge. With AI and, in particular, large language models (LLMs), product ideation becomes effortless.

LLMs can take massive amounts of textual information (whether it's text reviews, YouTube videos, or voice surveys) and generate novel text. The more data that has been annotated for the model, the better it can generate original ideas. For example, with Commerce.AI's product idea generator, machine learning is used to scan large amounts of product reviews from a given category, such as men's wristwatches, and then extract customer wishlist information.

The next step is taking this customer wishlist information and feeding it through an LLM to generate a new product idea. In the following example, given the wishlist points of I just wish the hands glowed and I wish it had a metal caseback, the LLM generates the following text:

A carbon steel watch with a glowing face. Under each hour, you have a number of dots that light up in sequence from left to right, then back to the left again when it reaches 12.

Figure 2.3 – Generating product ideas based on a customer wishlist

If you haven't used a language model yet, you can try out AI models such as OpenAI's GPT-3. GPT-3 was trained on billions of content pieces on the internet to generate new text, like your phone's autocomplete on steroids.

For now, these systems are pretty basic: they just generate words one after another without real intelligence or understanding what they mean. That said, GPT-3 can be used to write like a human for a variety of real-world use cases. For instance, it could be trained to produce coherent arguments in favor of a political candidate. Or it could be trained to write more evocatively about a particular memory or experience. It could also be trained to generate new product ideas.

So, while it may sound whimsical at first, there are quite real applications for this technology, which is based on an AI architecture called the transformer.

Transformers

GPT-3, for instance, uses the transformer. Transformers are the secret weapon behind some of the world's most popular natural language models, from the likes of Facebook, Google, and Microsoft.

At a technical level, transformers use what's called a sequence-to-sequence architecture, or Seq2Seq. As the name suggests, Seq2Seq is a neural network that transforms a sequence
of elements, such as a string of text, into another sequence.

Seq2Seq models are made up of an encoder and a decoder. An encoder takes the sequence and turns it into a set of latent variables, which are then passed to the decoder. The decoder
is responsible for turning the elements back into a sequence.

To give an example, if you have a document that says The quick brown fox jumps over the lazy dog, the model might like to know what animal a dog is referring to. This task involves encoding the word dog into a numerical token that can be calculated regarding other tokens.

For example, the token for the word dog might be located between the tokens for the words wolf and pet. Of course, this is hugely simplifying the matter, as every character and every word gets a token, creating a huge network that computers can use to understand language.

This allows computers to extract semantic features from documents in a way that had previously been thought impossible. As we mentioned previously, this model uses sentence structure as an input feature – or you could call it an input vector – so we can see how this works at a conceptual level here.

Sentences are parameters in supervised sequence models such as Seq2Seq; they are fed through an encoder, which turns them into vectors (latent variables). These are then passed through one or more networks before being turned back into sequences for output processing (decoding).

A quick note on terminology

These models aren't just called transformers because they transform sentences; they also transform characters into characters (or words). A character transformer doesn't just process words but also letters. It processes each character individually instead of taking groups of characters at once.

There are differences between character transformer models and word processors such as Microsoft Word or Apple Pages. However, they all do something similar under the hood: they process raw data with a computer algorithm until a result pops up on the screen that we humans can read easily.

Transformers are a crucial AI development as most people agree that robots will need excellent reading and writing skills before becoming truly useful. Whether we want robots to vacuum our floors or serve us dinner depends on how good their language comprehension capabilities are. We need robots who can understand natural language sentences in context before we teach them specific tasks.

While transformers can be trained on text, which gets encoded as numerical token values, they can also be trained on pixels, which get encoded as numbers (that is, a matrix of pixel RGB values). As a result, transformers can be used to make images in much the same way that they can write text, as seen in this example of a chair, which has been designed with the help of AI image generation:

Figure 2.4 – A generative design by Emmanuel Touraine, CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0)

From product designs to logo assets or other design collateral, AI-powered image generators will play an increasingly large role in the product innovation process.

Selecting product ideas

Innovation is the lifeblood of product teams. However, innovating new product ideas is just one step. Let's explore how to select the right product idea for business success.

Many business leaders understate the importance of thinking about a product as an extended process. They think that once they have a good idea, they can immediately build it and get it to market. But this line of thinking is fundamentally flawed. The most successful products in history were not developed overnight but rather evolved through systematic planning and execution processes. This basic path is how many companies sustainably grow their market share year after year.

Important questions here include the following:

• Where should product managers focus their energy?
• Should they focus on the original idea (or ideas) or should they be developing products based on other opportunities that could provide an even stronger return on investment?
• How do you know which ideas are worth pursuing?

Every company faces different external competitive pressures. That means every company has unique opportunities and threats associated with its products. For example, when Facebook was founded in 2004, there were no smartphones available to consumers; today, there are thousands of smartphone apps that compete with Facebook for users' attention. So, which ideas make sense for Facebook to pursue?

To help answer these questions, you can consider six key factors:

1. Your company's need.
2. What problems the product solves for your customer.
3. Whether those problems are being neglected by competitor.
4. How big the market is for solving that problem.
5. Where you are in the life cycle of a product.
6. Whether there is a new version of your product coming out soon, or whether you are currently in the market with an existing version.

The answers to these questions will help you prioritize your product ideas. For example, if your company is far along with its current product and it's not suffering from any major competitive threats, then perhaps it can afford to take some time off to work on something completely different. However, if you're still working on development and planning for a new release of your current product, then pivoting into something new may make less sense.

If you have highly skilled employees who already know how to make products like yours, they may be able to start building right away without much need for training or retraining. Alternatively, if it takes many people with complementary skill sets to build products successfully (as it does at companies such as Facebook), then you may want to expand your team first before starting a project that requires more than one person (such as hiring designers and writers). You might also consider partnering with other organizations that have relevant skills (for example, Twitter has relied on outside contractors as part of its engineering teams).

These partnerships can also help mitigate risk by spreading out development costs across multiple organizations.

This factor changes significantly, depending on where your company is in terms of fundraising activity and cash reserves. If you're raising money or looking for investors now but don't have significant revenue streams generating profits, then spending money on risky projects that aren't likely to be profitable immediately isn't so crucial at first. However, once things get going again (and assuming investors continue believing in your business model), launching products quickly will allow them to generate revenue faster.

It's important to note that having strong answers to all questions does not guarantee success; it just helps teams make better choices about which opportunities are promising enough to pursue further down the road.

Iterating product ideas

However, selecting the right product idea is not enough. Once a potentially fruitful idea has been selected, idea iteration takes over. Idea iteration is a highly strategic process.

Many companies try to execute their original product ideas without formal planning or iteration. These companies often fail to sustain momentum and ultimately kill promising but unproven products.

In contrast, many successful products have gone through explicit iterations that are guided by principles of organizational learning. While the idea itself is not changed, teams add new features, new capabilities, and new functionality in response to what users want. An example is Yahoo's acquisition of Flickr in 2005. At the time, Flickr was already a popular destination for people looking to share photos and videos with other users online. But by buying Flickr, the Yahoo team could create an even more compelling user experience by providing photo sharing.

Another example is the evolution of Facebook from the hot or not rating site to a global communications and media platform. Facebook started in college dorm rooms, and many early users wanted a casual way to stay in touch with their friends. But as Facebook became more popular, it attracted more professional users who wanted greater benefits from the site. In response, the team created new features such as Events and Groups that helped people find groups of friends they didn't know well but could use as sounding boards for career advice or to help them organize their workday.

Of course, not every product idea will reach such a level of success. But if your organization is going through an innovation process similar to this one, you'll want to make sure you have a strategy for iteration at each step of the way. The lesson here is that an innovation process should be guided by user wants and needs.

Summary

AI is a very broad concept. The way we use the term varies, depending on the context. In most cases, it refers to machines that can carry out difficult tasks on their own by learning from data.

Commerce.AI, as an AI platform, stands between you and your data so that you can build products that solve customers' problems more effectively than before, including by using creative AI. It captures data from every part of the product ecosystem and allows you to ask questions about products in ways a human never could, and it turns raw data into valuable information that you can use to improve your products and make better decisions about how to grow your business.

Through the pillars of language understanding, visual understanding, information extraction, and information organization, you can use AI to empower product teams with more efficient creative ideation. Creative ideation is vital for product success, but it's not the only skill you need. Industry trends change, and it takes a keen eye to see those changes coming and then pivot your product strategy accordingly. While this chapter has explored how to generate product ideas, you'll want to make sure that you're coming up with product ideas within a trending, viable market.

In the next chapter, we'll explore how to predict industry-wide trends using big data.